diff --git a/.dockerignore b/.dockerignore index c0d8a84b5..3295cc1b4 100644 --- a/.dockerignore +++ b/.dockerignore @@ -22,6 +22,10 @@ outputs rl media +# Local virtualenvs (the image provides its own) +.venv +venv + # Logging logs diff --git a/.github/workflows/benchmark_tests.yml b/.github/workflows/benchmark_tests.yml index b82c59a8b..3493e5048 100644 --- a/.github/workflows/benchmark_tests.yml +++ b/.github/workflows/benchmark_tests.yml @@ -167,9 +167,9 @@ jobs: # ── LIBERO TRAIN+EVAL SMOKE ────────────────────────────────────────────── # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then - # immediately runs eval inside the training loop (eval_freq=1, 1 episode). + # immediately runs eval inside the training loop (env_eval_freq=1, 1 episode). # Tests the full train→eval-within-training pipeline end-to-end. - - name: Run Libero train+eval smoke (1 step, eval_freq=1) + - name: Run Libero train+eval smoke (1 step, env_eval_freq=1) if: env.HF_USER_TOKEN != '' run: | docker run --name libero-train-smoke --gpus all \ @@ -196,7 +196,7 @@ jobs: --output_dir=/tmp/train-smoke \ --steps=1 \ --batch_size=1 \ - --eval_freq=1 \ + --env_eval_freq=1 \ --eval.n_episodes=1 \ --eval.batch_size=1 \ --eval.use_async_envs=false \ diff --git a/AGENT_GUIDE.md b/AGENT_GUIDE.md index 57a33fdba..03b270dce 100644 --- a/AGENT_GUIDE.md +++ b/AGENT_GUIDE.md @@ -138,7 +138,7 @@ lerobot-replay --robot.type=so101_follower --robot.port= --robot. --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0 ``` -**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration. +**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. No local GPU? Add `--job.target=` (e.g. `a10g-small`, list them with `hf jobs hardware`) to run on Hugging Face Jobs instead. See §6/§7 for policy and duration. ```bash lerobot-train \ diff --git a/Makefile b/Makefile index d3987101f..ea3b6e261 100644 --- a/Makefile +++ b/Makefile @@ -58,7 +58,7 @@ test-act-ete-train: --dataset.episodes="[0]" \ --batch_size=2 \ --steps=4 \ - --eval_freq=2 \ + --env_eval_freq=2 \ --eval.n_episodes=1 \ --eval.batch_size=1 \ --save_freq=2 \ @@ -96,7 +96,7 @@ test-diffusion-ete-train: --dataset.episodes="[0]" \ --batch_size=2 \ --steps=2 \ - --eval_freq=2 \ + --env_eval_freq=2 \ --eval.n_episodes=1 \ --eval.batch_size=1 \ --save_checkpoint=true \ @@ -126,7 +126,7 @@ test-tdmpc-ete-train: --dataset.episodes="[0]" \ --batch_size=2 \ --steps=2 \ - --eval_freq=2 \ + --env_eval_freq=2 \ --eval.n_episodes=1 \ --eval.batch_size=1 \ --save_checkpoint=true \ @@ -161,7 +161,7 @@ test-smolvla-ete-train: --dataset.episodes="[0]" \ --batch_size=2 \ --steps=4 \ - --eval_freq=2 \ + --env_eval_freq=2 \ --eval.n_episodes=1 \ --eval.batch_size=1 \ --save_freq=2 \ diff --git a/README.md b/README.md index fa3e9e1a3..53d92f96e 100644 --- a/README.md +++ b/README.md @@ -87,7 +87,7 @@ Learn more about it in the [LeRobotDataset Documentation](https://huggingface.co ## SoTA Models -LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models, with more coming soon. It also provides you with the tools to instrument and inspect your training process. +LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, Vision-Language-Action (VLA) models, World Models, and Reward Models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.

Gr00t Architecture @@ -97,17 +97,17 @@ Training a policy is as simple as running a script configuration: ```bash lerobot-train \ - --policy=act \ + --policy.type=act \ --dataset.repo_id=lerobot/aloha_mobile_cabinet ``` -| Category | Models | -| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| **Imitation Learning** | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md) | -| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon) | -| **VLAs Models** | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx) | -| **World Models** | [VLA-JEPA](./docs/source/vla_jepa.mdx) (more coming soon) | -| **Reward Models** | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx) | +| Category | Models | +| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | +| **Imitation Learning** | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md) | +| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon) | +| **VLAs Models** | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx), [EVO1](./docs/source/evo1.mdx) | +| **World Models** | [VLA-JEPA](./docs/source/vla_jepa.mdx), [LingBot-VA](./docs/source/lingbot_va.mdx), [FastWAM](./docs/source/fastwam.mdx) | +| **Reward Models** | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx) | Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub @@ -136,6 +136,7 @@ Learn how to implement your own simulation environment or benchmark and distribu - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments. - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot. - **[T-Shirt Folding Experiment](https://huggingface.co/spaces/lerobot/robot-folding):** An end-to-end demonstration of folding t-shirts with LeRobot. +- **[LeLab](https://github.com/huggingface/leLab):** A web interface for LeRobot — teleoperate, calibrate, record datasets, replay, and train your SO arm from the browser, no CLI required. ## Citation diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml index 5d847a94d..92a9b22b2 100644 --- a/docs/source/_toctree.yml +++ b/docs/source/_toctree.yml @@ -69,8 +69,14 @@ title: VLA-JEPA - local: eo1 title: EO-1 + - local: lingbot_va + title: LingBot-VA + - local: fastwam + title: FastWAM + - local: evo1 + title: EVO1 - local: groot - title: NVIDIA GR00T N1.5 + title: NVIDIA GR00T - local: xvla title: X-VLA - local: multi_task_dit diff --git a/docs/source/bring_your_own_policies.mdx b/docs/source/bring_your_own_policies.mdx index 1b3871516..c3cc040e3 100644 --- a/docs/source/bring_your_own_policies.mdx +++ b/docs/source/bring_your_own_policies.mdx @@ -295,11 +295,12 @@ The file names are load-bearing: the factory does lazy imports by name, and the ### Wiring -Three places need to know about your policy. All by name. +Four places need to know about your policy. All by name. 1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast). 2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import. 3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches. +4. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page. Mirror an existing policy that's structurally similar to yours; the diff is small. @@ -371,6 +372,8 @@ The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingfa - [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard. - [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests. - [ ] `src/lerobot/policies//README.md` symlinked into `docs/source/policy__README.md`; user-facing `docs/source/.mdx` written and added to `_toctree.yml`. +- [ ] `templates/lerobot_modelcard_template.md` has a description entry and a `policy_docs` link for your policy. +- [ ] The models table in the root `README.md` lists your policy in the right category, linking to your doc page. - [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint). The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one. diff --git a/docs/source/cameras.mdx b/docs/source/cameras.mdx index 2dc2859dd..02714d591 100644 --- a/docs/source/cameras.mdx +++ b/docs/source/cameras.mdx @@ -157,6 +157,14 @@ finally: +### Working with depth + +The Intel RealSense and Reachy 2 cameras can capture both color and depth in lockstep. Calling `read()` returns the **color** frame as `(H, W, 3)` `uint8`. Calling `read_depth()` returns the **depth map** as `(H, W, 1)` `uint16`, where each pixel value is the distance from the sensor expressed in **millimetres**. A pixel value of `0` typically means "no measurement available" (out-of-range, occluded, or low-confidence). + +During recording, the control loop peeks the freshest buffered frames non-blockingly via `read_latest()` (color) and `read_latest_depth()` (depth), adding the depth map as a sibling feature (e.g. `front_depth` next to `front`). + +For how depth streams are stored and encoded when recording a dataset, see the [Depth streams](./video_encoding_parameters#depth-streams) section of the video encoding guide. + ## Use your phone's camera diff --git a/docs/source/cheat-sheet.mdx b/docs/source/cheat-sheet.mdx index a6afa14c2..0531c95bf 100644 --- a/docs/source/cheat-sheet.mdx +++ b/docs/source/cheat-sheet.mdx @@ -89,6 +89,36 @@ Control the data recording flow using keyboard shortcuts: - Press **Left Arrow (`←`)**: Delete current episode and retry. - Press **Escape (`ESC`)**: Stop, encode videos, and upload. +### Recording depth + +Intel RealSense cameras (`type: intelrealsense`) record a depth stream when you set `use_depth: true`. Depth is quantized to 12-bit codes and stored as its own video. + +```bash +lerobot-record \ + ... \ + --robot.cameras="{ head: {type: intelrealsense, serial_number_or_name: \"0123456789\", width: 640, height: 480, fps: 30, use_depth: true} }" \ + --dataset.repo_id=${HF_USER}/so101_depth_test \ + --dataset.single_task="put the red brick in a bowl" \ + --dataset.depth_encoder.depth_min=0.01 \ + --dataset.depth_encoder.depth_max=10.0 \ + --dataset.depth_encoder.shift=0.0 \ + --dataset.depth_encoder.use_log=true +``` + +### Video encoding parameters + +RGB and depth streams are encoded independently via the `--dataset.rgb_encoder.*` and `--dataset.depth_encoder.*` keys. + +```bash +lerobot-record \ + ... \ + --dataset.rgb_encoder.vcodec=h264 \ + --dataset.rgb_encoder.pix_fmt=yuv420p \ + --dataset.rgb_encoder.crf=23 \ + --dataset.depth_encoder.vcodec=hevc \ + --dataset.depth_encoder.extra_options='{"x265-params": "lossless=1"}' +``` + ### Training Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy: @@ -120,6 +150,14 @@ lerobot-train \ --steps=20000 ``` +No local GPU? Add `--job.target=` (e.g. `a10g-small`) to either command and `lerobot-train` runs it on [Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) instead — it uploads a local-only dataset for you and pushes the trained model. List flavors with `hf jobs hardware`. + +To resume, point `--config_path` at a checkpoint and add `--resume=true`. It accepts a local path or a Hub repo id (the latest checkpoint is fetched), and works locally or on a job by adding `--job.target=`: + +```bash +lerobot-train --config_path=${HF_USER}/policy_test --resume=true --job.target=a10g-small +``` + ### Inference Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever. diff --git a/docs/source/earthrover_mini_plus.mdx b/docs/source/earthrover_mini_plus.mdx index 508c0e3a9..f3b324093 100644 --- a/docs/source/earthrover_mini_plus.mdx +++ b/docs/source/earthrover_mini_plus.mdx @@ -194,7 +194,7 @@ lerobot-record \ --dataset.single_task="Navigate around obstacles" \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --display_data=true ``` diff --git a/docs/source/envhub_isaaclab_arena.mdx b/docs/source/envhub_isaaclab_arena.mdx index b934240d6..cd077806d 100644 --- a/docs/source/envhub_isaaclab_arena.mdx +++ b/docs/source/envhub_isaaclab_arena.mdx @@ -193,7 +193,7 @@ To learn more about training policies with LeRobot, please refer to the training - [SmolVLA](./smolvla) - [Pi0.5](./pi05) -- [GR00T N1.5](./groot) +- [GR00T N1.7](./groot) Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation: diff --git a/docs/source/evo1.mdx b/docs/source/evo1.mdx new file mode 100644 index 000000000..3f8e42798 --- /dev/null +++ b/docs/source/evo1.mdx @@ -0,0 +1,191 @@ +# EVO1 + +EVO1 is a Vision-Language-Action policy for robot control built around an InternVL3 backbone and a continuous flow-matching action head. This LeRobot integration exposes EVO1 as a standard policy type so it can be trained and evaluated with the usual LeRobot dataset, checkpoint, and processor APIs. + +## Model Overview + +The policy embeds one or more camera images and the language task prompt with InternVL3, pads robot state/action vectors to fixed maximum dimensions, and predicts future action chunks with a flow-matching action head. During inference, the policy samples an action chunk and returns `n_action_steps` actions from that chunk before sampling again. + +### What the LeRobot Integration Covers + +- Standard `policy.type=evo1` configuration through LeRobot +- InternVL3 image/text embedding with optional FlashAttention fallback +- Stage-based finetuning controls for action-head-only and VLM finetuning runs +- Continuous flow-matching action prediction +- Checkpoint save/load through LeRobot policy APIs +- Training with `lerobot-train` and evaluation with standard policy inference APIs + +The broader EVO1 project may include additional training scripts and dataset tooling. This page focuses on the LeRobot robot-control policy path. + +## Installation Requirements + +1. Install LeRobot by following the [Installation Guide](./installation). +2. Install EVO1 dependencies: + + ```bash + pip install -e ".[evo1]" + ``` + + For LIBERO evaluation, install the LIBERO extra as well: + + ```bash + pip install -e ".[evo1,libero]" + ``` + +3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available. + +EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory. + +## Data Requirements + +EVO1 expects a LeRobot dataset with: + +- One to `policy.max_views` visual observations, for example `observation.images.image` +- `observation.state` +- `action` +- A language task instruction in the dataset `task` field, or another field configured with `policy.task_field` + +State and action vectors are padded to `policy.max_state_dim` and `policy.max_action_dim`. Predictions are cropped back to the dataset action dimension before being returned. + +## Usage + +To use EVO1 in a LeRobot configuration, specify: + +```python +policy.type=evo1 +``` + +By default, a new EVO1 policy initializes its VLM from: + +```python +policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf +``` + +Once a LeRobot-format EVO1 checkpoint is available, load it with: + +```python +policy.path=your-org/your-evo1-checkpoint +``` + +## Training + +### Stage 1 + +Stage 1 freezes the VLM and trains the action head: + +```bash +lerobot-train \ + --dataset.repo_id=your_org/your_dataset \ + --policy.type=evo1 \ + --policy.training_stage=stage1 \ + --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \ + --policy.device=cuda \ + --policy.chunk_size=50 \ + --policy.n_action_steps=50 \ + --policy.max_state_dim=24 \ + --policy.max_action_dim=24 \ + --policy.optimizer_lr=1e-5 \ + --batch_size=4 \ + --steps=5000 \ + --output_dir=./outputs/evo1_stage1 +``` + +### Stage 2 + +Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint: + +```bash +lerobot-train \ + --dataset.repo_id=your_org/your_dataset \ + --policy.path=./outputs/evo1_stage1/checkpoints/005000/pretrained_model \ + --policy.training_stage=stage2 \ + --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \ + --policy.device=cuda \ + --policy.chunk_size=50 \ + --policy.n_action_steps=50 \ + --policy.max_state_dim=24 \ + --policy.max_action_dim=24 \ + --policy.optimizer_lr=1e-5 \ + --batch_size=4 \ + --steps=80000 \ + --output_dir=./outputs/evo1_stage2 +``` + +By default, `policy.training_stage` reapplies the finetuning defaults for that stage. This is important when +starting Stage 2 from a Stage 1 checkpoint, because the Stage 1 checkpoint config stores the VLM finetuning +flags as disabled. These stage defaults take precedence over saved or manually supplied `policy.finetune_*` +flags unless `policy.apply_training_stage_defaults=false`, so set that flag only when manually controlling +every finetuning flag. + +### Key Training Parameters + +| Parameter | Default | Description | +| --------------------------------------------- | --------------------------- | ----------------------------------------------------------------- | +| `policy.vlm_model_name` | `OpenGVLab/InternVL3-1B-hf` | Natively converted InternVL3 checkpoint or local model directory | +| `policy.training_stage` | `stage1` | `stage1` trains the action head; `stage2` finetunes VLM branches | +| `policy.apply_training_stage_defaults` | `true` | Reapplies stage finetuning defaults after loading a checkpoint | +| `policy.vlm_num_layers` | `14` | Number of InternVL3 language layers kept for the policy | +| `policy.vlm_dtype` | `bfloat16` | Requested VLM dtype | +| `policy.use_flash_attn` | `true` | Requests FlashAttention when installed; otherwise falls back | +| `policy.enable_gradient_checkpointing` | `true` | Enables checkpointing on supported InternVL3 modules | +| `policy.gradient_checkpointing_use_reentrant` | `false` | Reentrant setting passed to gradient checkpointing when supported | +| `policy.chunk_size` | `50` | Number of future actions predicted per chunk | +| `policy.n_action_steps` | `50` | Number of actions consumed from a sampled chunk | +| `policy.max_state_dim` | `24` | State padding dimension | +| `policy.max_action_dim` | `24` | Action padding dimension | +| `policy.postprocess_action_dim` | `null` | Optional action dimension returned after EVO1 postprocessing | +| `policy.binarize_gripper` | `false` | Binarizes the postprocessed gripper channel for LIBERO-style eval | +| `policy.task_field` | `task` | Batch field used as the language prompt | + +## Inference + +Try it out with a trained EVO1 checkpoint: + +```bash +lerobot-rollout \ + --policy.path=your-org/your-evo1-checkpoint \ + --inference.type=rtc \ # optional + ... +``` + +## Results + +### LIBERO Evaluation + +> [!NOTE] +> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation +> will be added once training completes. + +The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names +(`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every +14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor +can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To +evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead +of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags: + +```bash +lerobot-eval \ + --policy.path=your-org/your-evo1-libero-checkpoint \ + --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \ + --policy.device=cuda \ + --policy.use_flash_attn=true \ + --policy.n_action_steps=14 \ + --policy.postprocess_action_dim=7 \ + --policy.binarize_gripper=true \ + --env.type=libero \ + --env.task=libero_object \ + --env.camera_name_mapping="{agentview_image: agentview_image, robot0_eye_in_hand_image: robot0_eye_in_hand_image}" \ + --env.observation_height=448 \ + --env.observation_width=448 \ + --eval.batch_size=1 \ + --eval.n_episodes=1 +``` + +## References + +- [EVO1 repository](https://github.com/MINT-SJTU/Evo-1) +- [InternVL3-1B-hf](https://huggingface.co/OpenGVLab/InternVL3-1B-hf) + +## License + +This LeRobot integration follows the Apache 2.0 License used by LeRobot. Check the upstream EVO1 and InternVL3 model pages for the licenses of released checkpoints and data. diff --git a/docs/source/fastwam.mdx b/docs/source/fastwam.mdx new file mode 100644 index 000000000..18b4775f8 --- /dev/null +++ b/docs/source/fastwam.mdx @@ -0,0 +1,167 @@ +# FastWAM + +FastWAM is a World Action Model policy for robot control. The LeRobot integration exposes FastWAM through the standard policy API so it can be configured with `policy.type=fastwam`, trained with `lerobot-train`, and loaded through the LeRobot pretrained policy interface. + +## Model Overview + +FastWAM keeps video modeling during training, but uses direct action prediction at inference time instead of iteratively generating future observations. This LeRobot policy wraps the FastWAM action model, adapts LeRobot batches to FastWAM training samples, and provides the standard processor pipeline for normalization and action postprocessing. + +The implementation initializes the visual world-model components from `Wan-AI/Wan2.2-TI2V-5B` by default and predicts action chunks with shape `[batch, action_horizon, action_dim]`. + +### What the LeRobot Integration Covers + +- Standard `policy.type=fastwam` configuration through LeRobot +- Image, state, action, and language-task batch adaptation +- Action chunk inference through `select_action` and `predict_action_chunk` +- Checkpoint save/load through the LeRobot policy APIs +- Configurable LIBERO gripper action postprocessing + +## Installation Requirements + +Install LeRobot from source, then install FastWAM dependencies: + +```bash +pip install -e ".[fastwam]" +``` + +This installs the FastWAM policy extra from `pyproject.toml`: `transformers`, +`diffusers`, `ftfy`, and `regex`, plus LeRobot's base dependencies. + +For LIBERO evaluation, install the benchmark dependencies too: + +```bash +pip install -e ".[fastwam,libero]" +``` + +This installs both extras. In addition to the FastWAM dependencies above, the +`libero` extra installs LeRobot dataset dependencies, `hf-libero` on Linux, and +`scipy`. + +FastWAM uses the Wan2.2 TI2V backbone. The default model id is: + +```python +policy.model_id=Wan-AI/Wan2.2-TI2V-5B +``` + +## Data Requirements + +FastWAM expects a LeRobot dataset with: + +- one or more visual observations whose widths concatenate to `policy.image_size[1]` +- `observation.state` when `policy.proprio_dim` is not `None` +- `action` +- a language task instruction through the dataset task field, or precomputed `context` and `context_mask` tensors + +The default visual setup is one image feature named `observation.images.image` with shape `(3, 224, 448)`. If the dataset uses two cameras, configure `policy.input_features` so their heights match `224` and their widths sum to `448`. + +## Usage + +Create a new FastWAM policy with: + +```bash +lerobot-train \ + --dataset.repo_id=your-org/your-dataset \ + --policy.type=fastwam \ + --policy.action_dim=7 \ + --policy.proprio_dim=8 \ + --policy.action_horizon=32 \ + --policy.n_action_steps=10 \ + --policy.image_size='[224,448]' \ + --output_dir=./outputs/fastwam_training \ + --job_name=fastwam_training \ + --steps=300000 \ + --batch_size=8 \ + --policy.device=cuda +``` + +Evaluate an existing LeRobot-format checkpoint on LIBERO-10 with: + +```bash +lerobot-eval \ + --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \ + --policy.device=cuda \ + --policy.torch_dtype=float32 \ + --policy.n_action_steps=10 \ + --env.type=libero \ + --env.task=libero_10 \ + --env.observation_height=224 \ + --env.observation_width=224 \ + --eval.batch_size=1 \ + --eval.n_episodes=50 \ + --seed=0 \ + --env.episode_length=600 +``` + +For `libero_goal`, `libero_spatial`, and `libero_object`, use +`--env.episode_length=300`. + +For real-robot rollout, use the same checkpoint path: + +```bash +lerobot-rollout \ + --robot.type=so101_follower \ + --robot.port=/dev/ttyACM0 \ + --policy.path=your-org/fastwam-real-robot +``` + +## Configuration Notes + +### Image Features + +`policy.image_size` is the size of the concatenated FastWAM image tensor as `(height, width)`. Each configured image feature must have shape `(3, height, camera_width)`, and all camera widths must sum to the configured width. + +### Action Chunking + +`policy.action_horizon` controls the number of future actions supervised during training and predicted during inference. `policy.n_action_steps` controls how many actions are consumed before the policy predicts a fresh chunk. `policy.n_action_steps` must be less than or equal to `policy.action_horizon`. + +### Wan Components + +FastWAM loads the Wan VAE, video DiT, text encoder, and tokenizer from the configured Wan model directory or Hugging Face Hub model id. LeRobot-format FastWAM checkpoints saved by `save_pretrained` also copy the local Wan component files needed by `from_pretrained`. + +### Attention Backend + +FastWAM's DiT uses PyTorch's `scaled_dot_product_attention` (SDPA) for all attention. It does **not** use FlashAttention: its Mixture-of-Transformers (MoT) routing needs arbitrary boolean `[query, key]` attention masks, which the FlashAttention varlen API cannot express. Installing the `flash-attn` package therefore has no effect on the FastWAM path. (Note that SDPA itself may still select PyTorch's own flash / memory-efficient / math kernel internally — this is unrelated to the `flash-attn` package.) + +### LIBERO Action Toggle + +FastWAM LIBERO checkpoints use `policy.toggle_action_dimensions=[-1]` by +default to match the gripper action convention used by the original FastWAM +evaluation pipeline: + +```bash +--policy.toggle_action_dimensions='[-1]' +``` + +## Results + +Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224): + +| Suite | Success rate | n_episodes | +| -------------- | -----------: | ---------: | +| libero_spatial | 97.6% | 500 | +| libero_object | 99.0% | 500 | +| libero_goal | 95.0% | 500 | +| libero_10 | 94.0% | 500 | +| **average** | **96.4%** | 2000 | + +Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300` (1x H20 140 GB). + +## References + +- [Fast-WAM paper](https://arxiv.org/abs/2603.16666) +- [Fast-WAM project page](https://yuantianyuan01.github.io/FastWAM/) +- [Fast-WAM code](https://github.com/yuantianyuan01/FastWAM) +- [Released upstream checkpoints](https://huggingface.co/yuanty/fastwam) +- [Wan2.2 TI2V 5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) + +## Citation + +```bibtex +@article{yuan2026fastwam, + title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?}, + author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao}, + journal = {arXiv preprint arXiv:2603.16666}, + year = {2026}, + url = {https://arxiv.org/abs/2603.16666} +} +``` diff --git a/docs/source/groot.mdx b/docs/source/groot.mdx index a10b5e369..c6dcff2d7 100644 --- a/docs/source/groot.mdx +++ b/docs/source/groot.mdx @@ -1,16 +1,19 @@ -# GR00T N1.5 Policy +# GR00T Policy -GR00T N1.5 is an open foundation model from NVIDIA designed for generalized humanoid robot reasoning and skills. It is a cross-embodiment model that accepts multimodal input, including language and images, to perform manipulation tasks in diverse environments. +GR00T is an NVIDIA foundation model family for generalized humanoid robot reasoning and skills. It is a cross-embodiment policy that accepts multimodal input, including language, images, and proprioception, to perform manipulation tasks in diverse environments. -This document outlines the specifics of its integration and usage within the LeRobot framework. +LeRobot integrates GR00T N1.7 through the `groot` policy type. + +> [!WARNING] +> **Breaking change:** GR00T N1.5 support was removed from LeRobot, and current releases support GR00T N1.7 only. N1.5 checkpoints and configs are rejected with a migration note. To keep using an N1.5 checkpoint, pin the last release that supports it: `pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 (base model [`nvidia/GR00T-N1.7-3B`](https://huggingface.co/nvidia/GR00T-N1.7-3B)). ## Model Overview -NVIDIA Isaac GR00T N1.5 is an upgraded version of the GR00T N1 foundation model. It is built to improve generalization and language-following abilities for humanoid robots. +GR00T N1.7 uses a Cosmos-Reason2/Qwen3-VL backbone and provides checkpoints for SimplerEnv, DROID, and LIBERO. -Developers and researchers can post-train GR00T N1.5 with their own real or synthetic data to adapt it for specific humanoid robots or tasks. +Developers and researchers can post-train GR00T with their own real or synthetic data to adapt it for specific humanoid robots or tasks. -GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception. +GR00T uses pre-trained vision and language encoders with a flow matching action transformer to model a chunk of actions conditioned on vision, language, and proprioception. =2.2.1,<2.8.0" "torchvision>=0.21.0,<0.23.0" # --index-url https://download.pytorch.org/whl/cu1XX -pip install ninja "packaging>=24.2,<26.0" # flash attention dependencies -pip install "flash-attn>=2.5.9,<3.0.0" --no-build-isolation -python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} imported successfully')" +pip install "lerobot[groot]" ``` -3. Install LeRobot by running: +For a source checkout: ```bash -pip install lerobot[groot] +pip install -e ".[groot]" ``` ## Usage -To use GR00T in your LeRobot configuration, specify the policy type as: +To use GR00T N1.7: -```python -policy.type=groot +```bash +--policy.type=groot ``` ## Training @@ -63,72 +57,171 @@ policy.type=groot Here's a complete training command for finetuning the base GR00T model on your own dataset: +This command is using the `new_embodiment` flag, which is used for the SO-101 robot, [read more about how GR00T handles different embodiments.](https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md#--embodiment-tag). + ```bash -# Using a multi-GPU setup -accelerate launch \ - --multi_gpu \ - --num_processes=$NUM_GPUS \ - $(which lerobot-train) \ - --output_dir=$OUTPUT_DIR \ - --save_checkpoint=true \ - --batch_size=$BATCH_SIZE \ - --steps=$NUM_STEPS \ - --save_freq=$SAVE_FREQ \ - --log_freq=$LOG_FREQ \ - --policy.push_to_hub=true \ +# install extra deps for training +pip install "lerobot[training]" + +hf auth login +wandb login + +export DATASET_NAME=your_data_set +export HF_USER=your_hf_username +export DATASET=$HF_USER/$DATASET_NAME +export REPO_ID="${DATASET}_GR00T17" #this is the model that will be uploaded to huggingface +export OUTPUT_DIR=outputs/train/$REPO_ID + +lerobot-train \ + --dataset.repo_id=$DATASET \ + --dataset.image_transforms.enable=true \ --policy.type=groot \ + --policy.device=cuda \ + --policy.base_model_path=nvidia/GR00T-N1.7-3B \ + --policy.embodiment_tag=new_embodiment \ + --policy.chunk_size=16 \ + --policy.n_action_steps=16 \ + --policy.use_relative_actions=true \ + --policy.relative_exclude_joints='["gripper"]' \ + --policy.use_bf16=true \ + --policy.push_to_hub=true \ --policy.repo_id=$REPO_ID \ - --policy.tune_diffusion_model=false \ - --dataset.repo_id=$DATASET_ID \ + --seed=42 \ + --batch_size=64 \ + --steps=20000 \ + --save_checkpoint=true \ + --save_freq=5000 \ + --use_policy_training_preset=true \ + --env_eval_freq=0 \ + --eval_steps=0 \ + --log_freq=10 \ + --output_dir=$OUTPUT_DIR \ + --job_name=$DATASET \ --wandb.enable=true \ - --wandb.disable_artifact=true \ - --job_name=$JOB_NAME + --wandb.disable_artifact=true + ``` ## Performance Results -### Libero Benchmark Results +### LIBERO Benchmark Results > [!NOTE] -> Follow our instructions for Libero usage: [Libero](./libero) +> Follow the [LIBERO](./libero) setup instructions before running `lerobot-eval`. -GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results. +GR00T N1.7 has demonstrated strong performance on the LIBERO benchmark suite. To reproduce LeRobot results, follow the instructions in the [LIBERO](./libero) section. -| Benchmark | LeRobot Implementation | GR00T Reference | -| ------------------ | ---------------------- | --------------- | -| **Libero Spatial** | 82.0% | 92.0% | -| **Libero Object** | 99.0% | 92.0% | -| **Libero Long** | 82.0% | 76.0% | -| **Average** | 87.0% | 87.0% | +### Train on LIBERO -These results demonstrate GR00T's strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section. +Example training command for a LIBERO suite (here `libero_spatial`): + +```bash +IMAGE_TRANSFORMS='{ + "brightness": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"brightness": [0.7, 1.3]}}, + "contrast": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"contrast": [0.6, 1.4]}}, + "saturation": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"saturation": [0.5, 1.5]}}, + "hue": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"hue": [-0.08, 0.08]}} +}' + +lerobot-train \ + --dataset.repo_id=IPEC-COMMUNITY/libero_spatial_no_noops_1.0.0_lerobot \ + --dataset.root=/datasets/libero_spatial \ + --dataset.revision=main \ + --dataset.video_backend=pyav \ + --dataset.image_transforms.enable=true \ + --dataset.image_transforms.max_num_transforms=4 \ + --dataset.image_transforms.tfs="$IMAGE_TRANSFORMS" \ + --policy.type=groot \ + --policy.base_model_path=nvidia/GR00T-N1.7-3B \ + --policy.embodiment_tag=libero_sim \ + --policy.push_to_hub=false \ + --policy.use_relative_actions=false \ + --policy.max_steps=20000 \ + --batch_size=320 \ + --steps=20000 \ + --save_freq=2000 \ + --env_eval_freq=0 \ + --eval_steps=0 \ + --log_freq=10 \ + --wandb.enable=true \ + --wandb.project=lerobot \ + --wandb.mode=online \ + --wandb.disable_artifact=true \ + --num_workers=4 \ + --prefetch_factor=2 \ + --persistent_workers=true \ + --output_dir=$OUTPUT_DIR \ + --job_name=$JOB_NAME +``` + +This will follow the recipe found [here](https://github.com/NVIDIA/Isaac-GR00T/blob/main/examples/LIBERO/README.md). + +### GR00T N1.7 LIBERO Results + +Preliminary LeRobot integration results (GR00T-LeRobot, `eval.n_episodes >= 50` per suite): + +| Suite | Success rate | Checkpoint | +| ---------------- | -----------: | ------------------------------------------------------------------------------------------------------------- | +| LIBERO Spatial | 91% | [nvidia/gr00t17-lerobot-libero_spatial-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640) | +| LIBERO Object | 81% | [nvidia/gr00t17-lerobot-libero_object-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640) | +| LIBERO Goal | 97% | [nvidia/gr00t17-lerobot-libero_goal-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640) | +| LIBERO 10 (Long) | 84% | [nvidia/gr00t17-lerobot-libero_10-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640) | +| **Average** | **88.25%** | | + +```bash +export MODEL_ID=your_trained_model_on_huggingface + +lerobot-eval \ + --policy.type=groot \ + --policy.base_model_path=$MODEL_ID \ + --policy.embodiment_tag=libero_sim \ + --env.type=libero \ + --env.task=libero_spatial \ + --eval.n_episodes=50 +``` + +Use `eval.n_episodes >= 50` per suite when reporting success rates. ### Evaluate in your hardware setup Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example: ```bash -lerobot-rollout\ - --strategy.type=sentry \ - --strategy.upload_every_n_episodes=5 \ - --robot.type=bi_so_follower \ - --robot.left_arm_port=/dev/ttyACM1 \ - --robot.right_arm_port=/dev/ttyACM0 \ - --robot.id=bimanual_follower \ - --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}, - left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}, - top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}, - }' \ +# install extra deps for roullout and real hardware +pip install "lerobot[feetech,viz]" + +export MODEL_ID=your_trained_model_on_huggingface + +# make sure that camera index matches your setup! +# find index using `uv run lerobot-find-cameras opencv` +WRIST_CAM='wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: "MJPG"}' +FRONT_CAM='front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"}' +export ROBOT_CAMERAS="{ $WRIST_CAM, $FRONT_CAM }" +export ROBOT_ID=follower_robot +export ROBOT_PORT=/dev/ttyACM0 + +uv run lerobot-rollout \ + --strategy.type=base \ + --policy.path=$MODEL_ID \ + --policy.base_model_path=nvidia/GR00T-N1.7-3B \ + --policy.n_action_steps=8 \ + --robot.type=so101_follower \ + --robot.port=$ROBOT_PORT \ + --robot.id=$ROBOT_ID \ + --robot.cameras="$ROBOT_CAMERAS" \ + --task="place the vial in the rack" \ + --duration=60 \ + --device=cuda \ --display_data=true \ - --dataset.repo_id=/eval_groot-bimanual \ - --dataset.single_task="Grab and handover the red cube to the other arm" \ - --dataset.streaming_encoding=true \ - --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ - --policy.path=/groot-bimanual \ # your trained model - --duration=600 + --inference.type=rtc \ + --inference.rtc.enabled=True \ # set to False if it causes inference instability + --inference.rtc.execution_horizon=8 \ + --inference.queue_threshold=0 ``` +> [!NOTE] +> Value of `inference.queue_threshold` should not exceed 5 to ensure stable inference. + ## License -This model follows NVIDIA's proprietary license, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T). Future versions (starting from N1.7) will follow **Apache 2.0 License**. +GR00T N1.7 is released under the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). diff --git a/docs/source/hardware_guide.mdx b/docs/source/hardware_guide.mdx index 0998344ec..79c2de98c 100644 --- a/docs/source/hardware_guide.mdx +++ b/docs/source/hardware_guide.mdx @@ -82,17 +82,18 @@ VRAM is the first filter. Within a tier, pick by budget and availability — the ### Hugging Face Jobs -[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: **`huggingface/lerobot-gpu:latest`**, rebuilt **every night at 02:00 UTC from `main`** ([`docker_publish.yml`](https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml)) — so it tracks the current state of the repo, not a tagged release. +[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second, without owning a GPU. `lerobot-train` submits and streams the job for you — just add `--job.target=` to a normal training command: ```bash -hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \ - bash -c "nvidia-smi && lerobot-train \ - --policy.type=act --dataset.repo_id=/ \ - --policy.repo_id=/act_ --batch_size=8 --steps=50000" +lerobot-train \ + --policy.type=act --dataset.repo_id=/ \ + --policy.repo_id=/act_ \ + --job.target=a10g-large ``` Notes: -- The leading `nvidia-smi` is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched. -- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training. -- `--flavor` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). For the current full catalogue + pricing see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs). +- Run `hf auth login` once before submitting, the job runs under your token. +- `--job.target` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). List the current catalogue with pricing via `hf jobs hardware`, or see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs). +- The job defaults to a `2d` (48h) timeout. Override it with `--job.timeout=4h` (or any other valid duration string) to shorten or extend the timeout. The job automatically stops when the command completes. +- For the full walkthrough — dataset upload, checkpoint streaming, resuming a run on a job — see the [imitation-learning training guide](./il_robots#train-using-hugging-face-jobs). diff --git a/docs/source/hilserl.mdx b/docs/source/hilserl.mdx index 76e985cfe..09a370f3d 100644 --- a/docs/source/hilserl.mdx +++ b/docs/source/hilserl.mdx @@ -719,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c "num_workers": 4, "steps": 5000, "log_freq": 10, - "eval_freq": 1000, + "env_eval_freq": 1000, "save_freq": 1000, "save_checkpoint": true, "seed": 2, diff --git a/docs/source/hope_jr.mdx b/docs/source/hope_jr.mdx index 1f3b08fd7..c29a9f216 100644 --- a/docs/source/hope_jr.mdx +++ b/docs/source/hope_jr.mdx @@ -232,7 +232,7 @@ lerobot-record \ --dataset.private=true \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --display_data=true ``` @@ -278,6 +278,6 @@ lerobot-record \ --dataset.num_episodes=10 \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model ``` diff --git a/docs/source/il_robots.mdx b/docs/source/il_robots.mdx index 53ae5af82..5893b93f4 100644 --- a/docs/source/il_robots.mdx +++ b/docs/source/il_robots.mdx @@ -126,7 +126,7 @@ import time from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig from lerobot.cameras.opencv import OpenCVCameraConfig -from lerobot.utils.visualization_utils import init_rerun, log_rerun_data, shutdown_rerun +from lerobot.utils.visualization_utils import init_visualization, log_visualization_data, shutdown_visualization robot_config = SO101FollowerConfig( port="/dev/tty.usbmodem5AB90687491", @@ -142,7 +142,7 @@ teleop_config = SO101LeaderConfig( id="my_leader_arm", ) -init_rerun(session_name="teleoperation") +init_visualization("rerun", session_name="teleoperation") # pass "foxglove" to stream to Foxglove instead robot = SO101Follower(robot_config) teleop_device = SO101Leader(teleop_config) @@ -158,7 +158,7 @@ while True: observation = robot.get_observation() action = teleop_device.get_action() robot.send_action(action) - log_rerun_data(observation=observation, action=action) + log_visualization_data("rerun", observation=observation, action=action) elapsed_time = time.perf_counter() - start_time sleep_time = TIME_PER_FRAME - elapsed_time @@ -207,7 +207,7 @@ lerobot-record \ --dataset.num_episodes=5 \ --dataset.single_task="Grab the black cube" \ --dataset.streaming_encoding=true \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --dataset.encoder_threads=2 ``` @@ -223,7 +223,7 @@ from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig from lerobot.teleoperators.so_leader.so_leader import SO101Leader from lerobot.common.control_utils import init_keyboard_listener from lerobot.utils.utils import log_say -from lerobot.utils.visualization_utils import init_rerun +from lerobot.utils.visualization_utils import init_visualization from lerobot.scripts.lerobot_record import record_loop from lerobot.processor import make_default_processors @@ -270,7 +270,7 @@ def main(): # Initialize the keyboard listener and rerun visualization _, events = init_keyboard_listener() - init_rerun(session_name="recording") + init_visualization("rerun", session_name="recording") # Connect the robot and teleoperator robot.connect() @@ -390,9 +390,17 @@ Set the flow of data recording using command-line arguments: Control the data recording flow using keyboard shortcuts: -- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next. -- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it. -- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset. +- Press **Right Arrow (`→`)** or **`n`**: Early stop the current episode or reset time and move to the next. +- Press **Left Arrow (`←`)** or **`r`**: Cancel the current episode and re-record it. +- Press **Escape (`ESC`)** or **`q`**: Immediately stop the session, encode videos, and upload the dataset. + + + +These control-flow shortcuts work on **X11, Wayland, and headless/SSH** sessions. When a global keyboard backend isn't available (Wayland, a headless machine, or macOS without Accessibility permission), `lerobot-record` automatically reads the same keys from the terminal — launch it from an interactive terminal and keep it focused. You can also use the letter equivalents **`n`** (next, same as `→`), **`r`** (re-record, same as `←`) and **`q`** (quit, same as `ESC`). No `$DISPLAY` setup is required. + +This applies to the recording control flow only. Keyboard **teleoperation** (driving the robot with the keyboard) still needs a global key backend, so it works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless sessions. + + #### Tips for gathering data @@ -406,7 +414,7 @@ If you want to dive deeper into this important topic, you can check out the [blo #### Troubleshooting: -- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux). +- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as `lerobot-record` runs in an interactive terminal — no `$DISPLAY` setup is needed. If the keys have no effect, make sure you are in an interactive (TTY) terminal, not a piped/non-TTY session, and that it is focused; the letter equivalents `n` / `r` / `q` also work. Keyboard _teleoperation_ (as opposed to the recording control flow) still requires a global key backend — an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — and is unavailable on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux). ## Visualize a dataset @@ -506,6 +514,12 @@ lerobot-train \ --resume=true ``` +`--config_path` also accepts a **Hub repo id**: if a run pushed its checkpoints to the Hub (with `--save_checkpoint_to_hub=true`), you can resume straight from the repo — its latest checkpoint is downloaded and training continues, restoring the optimizer, scheduler, step counter and data order: + +```bash +lerobot-train --config_path=${HF_USER}/my_policy --resume=true +``` + If you do not want to push your model to the hub after training use `--policy.push_to_hub=false`. Additionally you can provide extra `tags` or specify a `license` for your model or make the model repo `private` by adding this: `--policy.private=true --policy.tags=\[ppo,rl\] --policy.license=mit` @@ -518,78 +532,48 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs). -To run the training use this command: +`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name: - - ```bash -hf jobs run \ - --flavor a10g-small \ - --timeout 4h \ - --secrets HF_TOKEN \ - huggingface/lerobot-gpu:latest \ - -- \ - python -m lerobot.scripts.lerobot_train \ - --dataset.repo_id=username/dataset \ - --policy.type=act \ - --steps=5000 \ - --batch_size=16 \ - --policy.device=cuda \ - --policy.repo_id=username/your_policy \ - --log_freq=100 +lerobot-train \ + --dataset.repo_id=${HF_USER}/so101_test \ + --policy.type=act \ + --policy.repo_id=${HF_USER}/my_policy \ + --job.target=a10g-small ``` - - - -```python -from huggingface_hub import run_job, get_token +List available flavors and pricing with `hf jobs hardware`. The run streams its logs to your terminal; press Ctrl-C to detach (the job keeps running in the cloud). Re-attach or cancel with: -run_name = "act_so101_hf_jobs" -dataset_id = "username/dataset" -user_hub_id = "username" - -command_args = [ - "python", "-m", "lerobot.scripts.lerobot_train", - "--dataset.repo_id", dataset_id, - "--policy.type", "act", - "--steps", "5000", - "--batch_size", "16", - "--num_workers", "4", - "--policy.device", "cuda", - "--log_freq", "100", - "--save_freq", "1000", - "--save_checkpoint", "true", - "--wandb.enable", "false", - "--policy.repo_id", f"{user_hub_id}/{run_name}" -] - -print(f"Submitting job '{run_name}' to Hugging Face Infrastructure...") - -job_info = run_job( - image="huggingface/lerobot-gpu:latest", - command=command_args, - flavor="a10g-small", - timeout="4h", - secrets={"HF_TOKEN": get_token()} -) - -print("\n🚀 Job successfully launched!") -print(f"🔹 Job ID: {job_info.id}") -print(f"🔗 Live UI Dashboard & Logs: {job_info.url}") +```bash +hf jobs logs +hf jobs cancel ``` - - - +If your dataset exists only locally (not yet on the Hub), it is automatically pushed to a **private** Hub repo so the job can download it by `repo_id` (nothing is made public). The trained model is pushed to the model repo at the end of the run. To also push every intermediate checkpoint to the Hub as it is saved (so you can monitor progress mid-run), add `--save_checkpoint_to_hub=true` — this requires a runtime image that includes this feature. -You can modify the `--flavor` to use different hardware, for example: `t4-small`, `a100-large`, `h200`. Use `hf jobs hardware` to see the full list with pricing. -Depending on the model you want to train and the hardware you selected you can also modify the `--batch_size` and `--number_of_workers`. -For longer training sessions increase the timeout. +Every job (and any dataset pushed by the run) is tagged `lerobot` so it's easy to find on the Hub. Add your own with `--job.tags '["my-tag"]'`. -Once the training is started you can go to [Jobs](https://huggingface.co/settings/jobs) and see if your jobs is running as well as all the outputs. Sometimes it takes a few minutes to schedule your job so be patient. +By default the job is capped at `2d` (48h) of wall-clock. Override it with an HF Jobs duration string, e.g. `--job.timeout=4h` to fail faster or `--job.timeout=7d` for a longer run. -After training the model will be pushed to hub and you can use it as any other model with LeRobot. +> **Note:** the model repo is created up front (it holds the staged training config the job runs from). If a run fails before the model is pushed, that repo is left on the Hub so you can inspect it — it is not deleted automatically, so repeated failures can leave empty repos behind. Remove one with `hf repo delete `. + +**Prerequisites:** run `hf auth login` before submitting. For Weights & Biases integration, run `wandb login` or set `WANDB_API_KEY` on your machine — the key is forwarded to the job automatically. + +**Resuming on a job.** Adding `--job.target` to a resume command runs the resume in the cloud — the same command works locally or remotely. The checkpoint repo is the source of truth, and new checkpoints continue the lineage in the same repo: + +```bash +# resume a Hub run on a job (its checkpoints are already on the Hub) +lerobot-train --config_path=${HF_USER}/my_policy --resume=true --job.target=a10g-small + +# resume a LOCAL run on a job — the checkpoint is uploaded to a private Hub repo first, +# then the job resumes from it (a local-only dataset is uploaded the same way) +lerobot-train \ + --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \ + --resume=true \ + --job.target=a10g-small +``` + +Job settings come from the current command, so override `--job.target`, `--job.timeout`, etc. as needed; for the resumed run to itself be resumable later, keep `--save_checkpoint_to_hub=true`. #### Upload policy checkpoints @@ -612,6 +596,8 @@ hf upload ${HF_USER}/act_so101_test${CKPT} \ Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs: +The examples below load the model from `--policy.path`. To pin a specific pushed version — useful once `--save_checkpoint_to_hub=true` has committed several checkpoints — add `--policy.pretrained_revision` with a commit hash, branch, or tag. Each pushed checkpoint is tagged with its step (e.g. `--policy.pretrained_revision=010000`), so you can recover a checkpoint by step without looking up its commit sha. + ```bash diff --git a/docs/source/lekiwi.mdx b/docs/source/lekiwi.mdx index 7e7c1a680..739073b65 100644 --- a/docs/source/lekiwi.mdx +++ b/docs/source/lekiwi.mdx @@ -319,7 +319,7 @@ If you want to dive deeper into this important topic, you can check out the [blo #### Troubleshooting: -- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux). +- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as you run the recording from an interactive terminal (keep it focused) — no `$DISPLAY` setup is needed; the letter equivalents `n` / `r` / `q` also work. Note that **keyboard teleoperation of the LeKiwi base** is different: it relies on a global key backend and therefore works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux). ## Replay an episode diff --git a/docs/source/lerobot-dataset-v3.mdx b/docs/source/lerobot-dataset-v3.mdx index 21cb232d3..0647af0b0 100644 --- a/docs/source/lerobot-dataset-v3.mdx +++ b/docs/source/lerobot-dataset-v3.mdx @@ -44,7 +44,7 @@ lerobot-record \ --dataset.num_episodes=5 \ --dataset.single_task="Grab the black cube" \ --dataset.streaming_encoding=true \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --dataset.encoder_threads=2 ``` diff --git a/docs/source/libero.mdx b/docs/source/libero.mdx index 043348690..b95af1d27 100644 --- a/docs/source/libero.mdx +++ b/docs/source/libero.mdx @@ -143,7 +143,7 @@ lerobot-train \ --batch_size=4 \ --eval.batch_size=1 \ --eval.n_episodes=1 \ - --eval_freq=1000 + --env_eval_freq=1000 ``` ## Reproducing published results diff --git a/docs/source/libero_plus.mdx b/docs/source/libero_plus.mdx index 4249bf49e..b065649fa 100644 --- a/docs/source/libero_plus.mdx +++ b/docs/source/libero_plus.mdx @@ -173,7 +173,7 @@ lerobot-train \ --batch_size=4 \ --eval.batch_size=1 \ --eval.n_episodes=1 \ - --eval_freq=1000 + --env_eval_freq=1000 ``` ## Relationship to LIBERO diff --git a/docs/source/lingbot_va.mdx b/docs/source/lingbot_va.mdx new file mode 100644 index 000000000..d33e90340 --- /dev/null +++ b/docs/source/lingbot_va.mdx @@ -0,0 +1,187 @@ +# LingBot-VA + +LingBot-VA is an **autoregressive video-action world-model policy** built on the **Wan2.2** +video-diffusion stack. It interleaves, in one autoregressive sequence, the prediction of +future **video latents** and **robot actions** ("VA" = Video-Action). The LeRobot +integration wires LingBot-VA into the standard training, evaluation and processor +interfaces. + +## Model Overview + +LingBot-VA is a **dual-stream "mixture-of-transformers"**: a video/latent stream +(`patch_embedding_mlp → blocks → proj_out`) and an action stream +(`action_embedder → blocks → action_proj_out`) share the same 30 transformer blocks and +text conditioning. + +| Component | Class | Role | +| ------------------------ | ----------------------- | ----------------------------------------------------------- | +| DiT backbone (trainable) | `WanTransformer3DModel` | ~5B-param dual-stream transformer. | +| VAE (frozen) | `AutoencoderKLWan` | Wan2.2 VAE, `z_dim=48`. Lazy-pulled from the source repo. | +| Text encoder (frozen) | `UMT5EncoderModel` | UMT5-XXL, `d_model=4096`. Lazy-pulled from the source repo. | + +At inference the policy runs an autoregressive loop per chunk: it denoises the video-latent +stream (CFG, ~20 steps) and the action stream (~50 steps) with two independent +flow-matching schedulers, maintaining a KV cache across chunks. Real observed keyframes are +fed back into the KV cache as the chunk is executed (closed-loop world modeling). + +### What the LeRobot Integration Covers + +- Standard `policy.type=lingbot_va` configuration through LeRobot. +- Ready-to-use LeRobot-format checkpoints on the Hub (converted from the released upstream ones). +- Autoregressive dual-stream inference behind the standard `select_action` interface + (single-environment eval, `--eval.batch_size=1`). +- Opt-in saving of the policy's **predicted (imagined) videos** during eval / training. +- Evaluation with `lerobot-eval` on LIBERO and RoboTwin. +- Training / fine-tuning via the dual-stream flow-matching loss (`policy.forward`), see below. + +## Installation + +1. Install LeRobot by following the [Installation Guide](./installation). +2. Install the LingBot-VA extra: + +```bash +pip install -e ".[lingbot_va]" +``` + +## Checkpoints + +The released upstream checkpoints have been converted to LeRobot format and pushed to the Hub: + +| Variant | LeRobot checkpoint | +| ---------------------- | -------------------------------- | +| LIBERO-Long post-train | `lerobot/lingbot_va_libero_long` | +| RoboTwin post-train | `lerobot/lingbot_va_robotwin` | +| Pretrained base | `lerobot/lingbot_va_base` | + +Only the trainable ~5B transformer is stored in the LeRobot +`model.safetensors`. The frozen VAE + UMT5 + tokenizer (~20 GB) are pulled from +`config.wan_pretrained_path` at load time (defaults to the source `robbyant/*` repo). The +UMT5-XXL text encoder runs on CPU by default (`config.text_encoder_device`) so the 5B +transformer + VAE fit on a single 24–32 GB GPU. + +## Evaluation (LIBERO) + +```bash +lerobot-eval \ + --policy.path=lerobot/lingbot_va_libero_long \ + --policy.device=cuda \ + --env.type=libero --env.task=libero_10 \ + --env.observation_height=128 --env.observation_width=128 \ + --eval.n_episodes=50 --eval.batch_size=1 \ + --output_dir=outputs/eval/lingbot_va_libero +``` + +LingBot-VA's streaming inference (KV cache + observed-keyframe feedback) is implemented for +single-environment eval; use `--eval.batch_size=1`. + +## Evaluation (RoboTwin) + +RoboTwin 2.0 needs the SAPIEN + CuRobo simulator stack. You can use the benchmark Docker image +(`docker/Dockerfile.benchmark.robotwin`, which also needs `warp-lang==1.3.1` and CuRobo built +with the GPU's compute capability in `TORCH_CUDA_ARCH_LIST`). RoboTwin uses **end-effector-pose +control**, so run with `--env.action_mode=ee`: the policy predicts per-arm `xyz+quaternion+gripper` +deltas (`robotwin_tshape` latent layout) that are composed onto the episode's initial eef pose and +executed via CuRobo IK. + +```bash +lerobot-eval \ + --policy.path=lerobot/lingbot_va_robotwin \ + --policy.device=cuda \ + --env.type=robotwin --env.task=beat_block_hammer --env.action_mode=ee \ + --eval.n_episodes=10 --eval.batch_size=1 \ + --output_dir=outputs/eval/lingbot_va_robotwin +``` + +### Saving predicted (imagined) videos + +Set `--policy.save_predicted_video=true` to additionally VAE-decode the predicted video +latents and write `pred_episode_*.mp4` next to the env-rendered `eval_episode_*.mp4` videos. +The same flag works for the periodic eval during `lerobot-train`. + +## Training / fine-tuning + +`LingBotVAPolicy.forward(batch)` implements the dual-stream **flow-matching** loss +(`latent_loss + action_loss`, timestep-weighted, action-masked) from the paper: it VAE-encodes +the camera clips into video latents, UMT5-encodes the task, noises both streams, runs the +transformer's block-causal training pass and returns `(loss, metrics)`. Optimizer preset is AdamW +with a linear-warmup-then-constant schedule (matching upstream). + +Requirements: + +- The block-causal masks use PyTorch **flex-attention**, so build the policy with + `--policy.attn_mode=flex` for training (the default `torch` SDPA is inference-only). +- The full 5B DiT does not fit a single 24–32 GB GPU under AdamW; fine-tune with **LoRA** + (`--policy.use_peft=true`) and/or optimizer offload. `get_optim_params` returns only the + trainable (e.g. adapter) parameters; the VAE + UMT5 text encoder stay frozen. + +```bash +lerobot-train \ + --policy.path=lerobot/lingbot_va_libero_long --policy.attn_mode=flex \ + --policy.use_peft=true \ + --dataset.repo_id= \ + --batch_size=1 --steps=... --output_dir=outputs/train/lingbot_va +``` + +The dataset must provide camera clips (a temporal window per camera, VAE-encoded to +`frame_chunk_size` latent frames) and `frame_chunk_size * action_per_frame` action steps per item. + +## Data format (action channels & camera order) + +LingBot-VA is an **end-effector (Cartesian) pose** policy, it predicts EEF poses + gripper, not +joint positions. Actions live in a fixed multi-embodiment **30-dim** layout; map your robot's +action dimensions into these channels and pad the rest with `0` (`used_action_channel_ids` selects +the channels a given checkpoint actually uses): + +| channels | meaning | +| -------- | ----------------------------------------------------- | +| 0–6 | Left-arm end-effector pose | +| 7–13 | Right-arm end-effector pose | +| 14–20 | Left-arm joints (unused by the released checkpoints) | +| 21–27 | Right-arm joints (unused by the released checkpoints) | +| 28 | Left gripper | +| 29 | Right gripper | + +- **LIBERO** uses channels `0–6`: a 6-DoF EEF delta (xyz + rotation) + gripper (single arm). +- **RoboTwin** uses channels `[0–6, 28, 7–13, 29]`: left EEF (xyz + quaternion) + left gripper + + right EEF + right gripper (16 dims). The env converts these poses to joint trajectories via + CuRobo IK — joints are never predicted. + +Joint-space datasets (or a different EEF convention) must be remapped into this schema before +fine-tuning these checkpoints. + +**Camera order is fixed and order-sensitive**, per-camera latents are concatenated spatially in +`obs_cam_keys` order, so the physical camera→slot mapping must match training: + +| benchmark | `obs_cam_keys` (in order) | `camera_layout` | +| --------- | ----------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- | +| LIBERO | `observation.images.image` (agentview / 3rd-person), `observation.images.image2` (eye-in-hand wrist) | `width_concat` (latents concatenated on width) | +| RoboTwin | `observation.images.head_camera`, `observation.images.left_camera`, `observation.images.right_camera` | `robotwin_tshape` (full-res head below, two half-res wrists on top) | + +The first camera is the exterior/head view and the rest are wrist views. + +## Inference Hyperparameters (LIBERO) + +| Key | Value | +| -------------------------------------- | --------------------------------------------------------------------------------- | +| height × width | 128 × 128 | +| cameras | `observation.images.image` (agentview), `observation.images.image2` (eye-in-hand) | +| action channels used | 0–6 (7-DoF arm + gripper) | +| action_per_frame / frame_chunk_size | 4 / 4 | +| attn_window | 30 | +| video / action denoising steps | 20 / 50 | +| guidance_scale / action_guidance_scale | 5 / 1 | +| snr_shift / action_snr_shift | 5.0 / 0.05 | + +These are the defaults of `LingBotVAConfig`; override any of them via `--policy.=...`. + +## Notes + +- **Attention backend:** inference uses the `torch` SDPA backend (always available). The + `flashattn` and `flex` backends are optional; `flex` is only needed for training. +- **Model size:** the DiT is ~5B params and the frozen VAE+UMT5 add ~20 GB; inference needs + roughly 18–24 GB of VRAM. + +## License + +LingBot-VA is released under Apache-2.0. See the +[upstream repository](https://github.com/Robbyant/lingbot-va). diff --git a/docs/source/metaworld.mdx b/docs/source/metaworld.mdx index 8e629dea9..b7accdfa2 100644 --- a/docs/source/metaworld.mdx +++ b/docs/source/metaworld.mdx @@ -120,11 +120,11 @@ lerobot-train \ --batch_size=4 \ --eval.batch_size=1 \ --eval.n_episodes=1 \ - --eval_freq=1000 + --env_eval_freq=1000 ``` ## Practical tips - Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context. - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark. -- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget. +- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget. diff --git a/docs/source/molmoact2.mdx b/docs/source/molmoact2.mdx index c6ae24e9e..9eb449ca9 100644 --- a/docs/source/molmoact2.mdx +++ b/docs/source/molmoact2.mdx @@ -17,7 +17,7 @@ the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2). Install LeRobot with the MolmoAct2 optional dependencies: ```bash -pip install -e ".[molmoact2]" +uv sync --locked --extra molmoact2 ``` To run the models in this repository, you need an NVIDIA GPU. The measurements @@ -46,8 +46,8 @@ The repo has been tested with Ubuntu 22.04. To use MolmoAct2 in a LeRobot training config, set: -```python -policy.type=molmoact2 +```bash +--policy.type=molmoact2 ``` ## Training @@ -103,7 +103,7 @@ accelerate launch \ --batch_size=32 \ --num_workers=4 \ --log_freq=20 \ - --eval_freq=-1 \ + --env_eval_freq=-1 \ --save_checkpoint=true \ --save_freq=2000 ``` @@ -142,7 +142,7 @@ accelerate launch \ --batch_size=32 \ --num_workers=4 \ --log_freq=20 \ - --eval_freq=-1 \ + --env_eval_freq=-1 \ --save_checkpoint=true \ --save_freq=2000 ``` @@ -386,6 +386,68 @@ These results demonstrate MolmoAct2's strong performance across diverse robotic manipulation tasks. To reproduce them, follow the instructions in the LIBERO evaluation section. +## Hardware Deployment (lerobot-rollout) + +LeRobot-format checkpoints are available on the Hub for direct use with +`lerobot-rollout`. Each checkpoint uses specific camera names that must +match your robot's camera configuration. + +### Camera naming convention + +Each checkpoint expects specific `observation.images.*` keys. +If your robot cameras have different names, use `--rename_map` to map them: + +| Checkpoint | Camera keys | Description | +| ----------------------------- | ---------------------- | ------------------------ | +| MolmoAct2-LIBERO-LeRobot | `image`, `wrist_image` | LIBERO sim cameras | +| MolmoAct2-BimanualYAM-LeRobot | `top`, `left`, `right` | YAM 3-camera setup | +| MolmoAct2-DROID-LeRobot | `cam0`, `cam1` | External + wrist | +| MolmoAct2-SO100_101-LeRobot | `cam0`, `cam1` | Primary + secondary view | + +Example with an SO-100 robot using top and side cameras: + +```bash +lerobot-rollout \ + --policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \ + --rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.side": "observation.images.cam1"}' \ + --robot.type=so100_follower \ + --robot.port=/dev/ttyACM0 \ + --robot.cameras='{ + top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, + side: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30} + }' \ + --task="pick up the red cube" --duration=30 +``` + +To use a wrist camera instead, just change the rename mapping: + +```bash +--rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.wrist": "observation.images.cam1"}' +``` + +### Joint frame transform (SO-100/101 zero-shot) + + +The MolmoAct2-SO100_101 checkpoint was trained on data that uses a different +joint calibration convention than LeRobot >= 0.5.0. Without a frame +correction, the arm may move in the wrong direction. + +This affects both **zero-shot deployment** and **fine-tuning** from the +original checkpoint. The pretrained weights expect the old convention, so +all joint data (observations and actions) must be transformed to match. + +The converted LeRobot checkpoint (`lerobot/MolmoAct2-SO100_101-LeRobot`) +already includes this correction in its processor pipeline. If you convert +or fine-tune the checkpoint yourself, set the following in the policy config (`configuration_molmoact2.py`): + +- `joint_signs`: `[1, -1, 1, 1, 1, 1]` (flips shoulder_lift direction) +- `joint_offsets`: `[0, 90, 90, 0, 0, 0]` (shifts shoulder_lift and elbow_flex by 90°) + +See the [backward compatibility guide](./backwardcomp) for details on the +calibration change. + + + ## Differences From the Original Implementation This LeRobot port is intended to match MolmoAct2 behavior while using LeRobot's diff --git a/docs/source/multi_gpu_training.mdx b/docs/source/multi_gpu_training.mdx index d7369e8f8..7c212364e 100644 --- a/docs/source/multi_gpu_training.mdx +++ b/docs/source/multi_gpu_training.mdx @@ -95,7 +95,7 @@ If you want to scale your hyperparameters when using multiple GPUs, you should d accelerate launch --num_processes=2 $(which lerobot-train) \ --optimizer.lr=2e-4 \ --dataset.repo_id=lerobot/pusht \ - --policy=act + --policy.type=act ``` **Training Steps Scaling:** @@ -110,9 +110,64 @@ accelerate launch --num_processes=2 $(which lerobot-train) \ --batch_size=8 \ --steps=50000 \ --dataset.repo_id=lerobot/pusht \ - --policy=act + --policy.type=act ``` +## Training Large Models with FSDP + +DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under +DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters, +gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background. + +An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine): + +```bash +accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \ + --dataset.repo_id=${HF_USER}/my_dataset \ + --policy.type= \ + --output_dir=outputs/train/my_policy_fsdp +``` + +A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent): + +```yaml +compute_environment: LOCAL_MACHINE +distributed_type: FSDP +mixed_precision: bf16 +num_machines: 1 +num_processes: 4 +fsdp_config: + fsdp_version: 1 + fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3) + fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP + fsdp_transformer_layer_cls_to_wrap: # repeated block class to shard + fsdp_use_orig_params: true # required: optimizer is built pre-prepare + fsdp_state_dict_type: FULL_STATE_DICT +``` + +Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each +block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the +optimizer before `accelerator.prepare()`. + +### FSDP checkpoints + +LeRobot gathers the full state dict across all ranks and the main process writes it as a single +`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for: + +- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master + copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently + with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific + caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting, + so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU + first, or cast `model.safetensors` to the deployment dtype offline. +- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved + alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json` + format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`. + Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can + resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only + sample-exact when the world size and batch size match the original run (a warning is logged + otherwise); the optimizer/model state itself is unaffected. + ## Notes - The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration. diff --git a/docs/source/multi_task_dit.mdx b/docs/source/multi_task_dit.mdx index 450d8a9f2..ebe46489a 100644 --- a/docs/source/multi_task_dit.mdx +++ b/docs/source/multi_task_dit.mdx @@ -314,7 +314,7 @@ lerobot-train \ --steps=30000 \ --save_freq=1000 \ --log_freq=100 \ - --eval_freq=1000 \ + --env_eval_freq=1000 \ --policy.type=multi_task_dit \ --policy.device=cuda \ --policy.horizon=32 \ diff --git a/docs/source/pi0fast.mdx b/docs/source/pi0fast.mdx index f7272acc5..15dff8071 100644 --- a/docs/source/pi0fast.mdx +++ b/docs/source/pi0fast.mdx @@ -96,7 +96,7 @@ lerobot-train \ --policy.type=pi0_fast \ --output_dir=./outputs/pi0fast_training \ --job_name=pi0fast_training \ - --policy.pretrained_path=lerobot/pi0_fast_base \ + --policy.pretrained_path=lerobot/pi0fast-base \ --policy.dtype=bfloat16 \ --policy.gradient_checkpointing=true \ --policy.chunk_size=10 \ @@ -187,7 +187,7 @@ lerobot-train \ --dataset.repo_id=lerobot/libero \ --output_dir=outputs/libero_pi0fast \ --job_name=libero_pi0fast \ - --policy.path=lerobot/pi0fast_base \ + --policy.path=lerobot/pi0fast-base \ --policy.dtype=bfloat16 \ --steps=100000 \ --save_freq=20000 \ diff --git a/docs/source/policy_evo1_README.md b/docs/source/policy_evo1_README.md new file mode 100644 index 000000000..dc8b75344 --- /dev/null +++ b/docs/source/policy_evo1_README.md @@ -0,0 +1,18 @@ +# EVO1 + +EVO1 is a Vision-Language-Action policy for robot control. The LeRobot +integration uses an InternVL3 vision-language backbone with a flow-matching +action head, and supports staged training through the standard LeRobot policy +APIs. + +The upstream EVO1 project is available at +[MINT-SJTU/Evo-1](https://github.com/MINT-SJTU/Evo-1). + +```bibtex +@misc{evo1, + title = {EVO1}, + author = {{MINT-SJTU}}, + year = {2025}, + howpublished = {\url{https://github.com/MINT-SJTU/Evo-1}}, +} +``` diff --git a/docs/source/policy_fastwam_README.md b/docs/source/policy_fastwam_README.md new file mode 100644 index 000000000..6af0eaa79 --- /dev/null +++ b/docs/source/policy_fastwam_README.md @@ -0,0 +1,56 @@ +## Research Paper + +Paper: https://arxiv.org/abs/2603.16666 + +## Repository + +Code: https://github.com/yuantianyuan01/FastWAM + +Project page: https://yuantianyuan01.github.io/FastWAM/ + +## Citation + +```bibtex +@article{yuan2026fastwam, + title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?}, + author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao}, + journal = {arXiv preprint arXiv:2603.16666}, + year = {2026}, + url = {https://arxiv.org/abs/2603.16666} +} +``` + +## Additional Resources + +Base video model: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B + +Released upstream checkpoints: https://huggingface.co/yuanty/fastwam + +## Results + +Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224): + +| Suite | Success rate | n_episodes | +| -------------- | -----------: | ---------: | +| libero_spatial | 97.6% | 500 | +| libero_object | 99.0% | 500 | +| libero_goal | 95.0% | 500 | +| libero_10 | 94.0% | 500 | +| **average** | **96.4%** | 2000 | + +Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300`. + +For LIBERO-10, use `--env.task=libero_10 --env.episode_length=600`: + +```bash +lerobot-eval \ + --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \ + --policy.device=cuda \ + --policy.torch_dtype=float32 \ + --policy.n_action_steps=10 \ + --env.type=libero \ + --env.task=libero_10 --env.observation_height=256 --env.observation_width=256 \ + --eval.batch_size=1 \ + --eval.n_episodes=50 \ + --seed=0 --env.episode_length=600 +``` diff --git a/docs/source/policy_groot_README.md b/docs/source/policy_groot_README.md index efcd76ebe..4b256fb27 100644 --- a/docs/source/policy_groot_README.md +++ b/docs/source/policy_groot_README.md @@ -1,6 +1,13 @@ ## Research Paper -Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/ +GR00T N1 technical report (covers the GR00T N1.x family, including N1.7): https://arxiv.org/abs/2503.14734 + +GR00T N1.7 model card: https://huggingface.co/nvidia/GR00T-N1.7-3B + +GR00T N1.5 research page (earlier version): https://research.nvidia.com/labs/gear/gr00t-n1_5/ + +> GR00T N1.5 support was removed from LeRobot; the last release supporting it is `lerobot==0.5.1`. +> Current releases support GR00T N1.7 only. ## Repository @@ -24,4 +31,108 @@ Code: https://github.com/NVIDIA/Isaac-GR00T Blog: https://developer.nvidia.com/isaac/gr00t -Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B +Hugging Face Models: + +- GR00T N1.7: https://huggingface.co/nvidia/GR00T-N1.7-3B +- GR00T N1.7 LIBERO checkpoints: https://huggingface.co/nvidia/GR00T-N1.7-LIBERO + +

+Original-vs-LeRobot parity test + +## Original-vs-LeRobot parity test + +`tests/policies/groot/test_groot_vs_original.py` verifies this LeRobot +reimplementation of GR00T N1.7 (Qwen3-VL backbone + flow-matching action head) +against NVIDIA's original `gr00t` package with two comparisons, each parametrized +over every embodiment tag present in the checkpoint: + +1. **Model parity** — given byte-identical pre-processed inputs and the same + flow-matching seed (recorded in each artifact), both implementations must produce + the **same raw model output** (`get_action(...)["action_pred"]`, the normalized + flow-matching prediction). Output shapes must match exactly; any action-horizon + or action-dim mismatch fails the test. +2. **Preprocessor parity** — given the identical raw observations (per-camera + frames, state vectors, language instruction), LeRobot's own preprocessor pipeline + (real Qwen3-VL chat template / tokenizer / image packing + checkpoint-driven + state normalization, no mocks) must produce the **same collated model inputs** + (`input_ids`, `attention_mask`, `pixel_values`, `image_grid_thw`, `state`, + `embodiment_id`) as the original package's processor. + +### Why two environments + +The original `gr00t` package pins `transformers==4.57.3` (Python 3.10); this +integration requires `transformers>=5.x` (Qwen3-VL). Under 5.x, `PretrainedConfig` +is itself a defaulted dataclass, so the original config dataclasses fail to import +(`non-default argument follows default argument`). The two implementations therefore +**cannot be imported in the same Python process**. + +So the test uses a **producer / consumer** split across two venvs: + +1. **Producer** — `tests/policies/groot/utils/dump_original_n1_7.py`, run in the _original_ + gr00t venv. For each embodiment it builds dummy inputs generically from the + checkpoint metadata (state dims from `statistics.json`; camera/language keys from + the processor modality configs), runs the original model, and saves to one `.npz` + per tag: the raw observations (`raw::` keys), the exact collated inputs + (`in::` keys), the seed, and the raw `action_pred`. +2. **Consumer** — the pytest above, run in the _LeRobot_ venv. It discovers every + `.npz`; the model-parity case replays the byte-identical collated inputs through + the LeRobot model with the recorded seed and asserts the outputs match, and the + preprocessor-parity case replays the raw observations through LeRobot's full + preprocessor pipeline and asserts the collated tensors match. + +> Artifacts generated by older versions of the dump script contain no `raw::` +> fields; the preprocessor-parity case then **skips** with a regeneration hint. +> Re-run the producer to refresh them. + +### Fairness controls + +- **Same pre-processed inputs (model parity)** — the original processor's `input_ids`, + `pixel_values`, `image_grid_thw`, `attention_mask`, `state`, `embodiment_id` are + fed verbatim to the LeRobot model (no re-tokenization / re-normalization), so the + model comparison isolates the model. LeRobot's own tokenization / image packing is + covered separately by the preprocessor-parity case, which compares its output + against those same collated tensors from identical raw observations. +- **Same precision + attention kernel** — both sides run **fp32 + SDPA**. The + original defaults to `use_flash_attention=True` (flash_attention_2 + bf16); the + producer forces SDPA + fp32. (With the defaults the gap is ~3e-2 — pure + kernel/rounding noise, not an implementation difference.) +- **Same flow-matching seed** — fixed right before sampling on both sides; the + producer records it in each artifact (`--seed`, default 42) and the consumer + replays the recorded value. + +### How to run + +```bash +# Resolve a local checkpoint (GR00T-N1.7-LIBERO / libero_10) +CKPT=$(python - <<'PY' +import os +from huggingface_hub import snapshot_download +print(os.path.join(snapshot_download("nvidia/GR00T-N1.7-LIBERO", + allow_patterns=["libero_10/*"]), "libero_10")) +PY +) + +# 1) Produce the original-side artifacts for all embodiments (original gr00t venv, CUDA) +CUDA_VISIBLE_DEVICES=0 /path/to/Isaac-GR00T/.venv-original/bin/python \ + tests/policies/groot/utils/dump_original_n1_7.py \ + --ckpt "$CKPT" --out-dir tests/policies/groot/artifacts --device cuda --seed 42 + +# 2) Run the parity test (LeRobot venv) — one parametrized case per embodiment +CUDA_VISIBLE_DEVICES=0 GROOT_PARITY_DEVICE=cuda \ + uv run pytest tests/policies/groot/test_groot_vs_original.py -v -s +``` + +The `.npz` artifacts are local-only (gitignored, ~6–10 MB each) and are regenerated by +the producer; they are never committed. The tests **skip** (do not fail) on CI or +when the checkpoint / artifacts are absent. + +#### Env knobs (all optional) + +| Var | Default | Purpose | +| ----------------------------------------- | -------------------------------- | ------------------------------------- | +| `GROOT_N1_7_PARITY_DIR` | `tests/policies/groot/artifacts` | directory of per-tag `.npz` artifacts | +| `GROOT_N1_7_LIBERO_CKPT` | auto (HF cache) | override checkpoint dir | +| `GROOT_PARITY_DEVICE` | `cuda` if available | `cpu` or `cuda` | +| `GROOT_PARITY_ATOL` / `GROOT_PARITY_RTOL` | `1e-3` | comparison tolerance | + +
diff --git a/docs/source/reachy2.mdx b/docs/source/reachy2.mdx index 4b08569db..7f975af43 100644 --- a/docs/source/reachy2.mdx +++ b/docs/source/reachy2.mdx @@ -161,7 +161,7 @@ lerobot-record \ --dataset.private=true \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --display_data=true ``` @@ -203,7 +203,7 @@ lerobot-record \ --dataset.private=true \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - # --dataset.camera_encoder.vcodec=auto \ + # --dataset.rgb_encoder.vcodec=auto \ --display_data=true ``` diff --git a/docs/source/robocasa.mdx b/docs/source/robocasa.mdx index f6a784e72..5a335a484 100644 --- a/docs/source/robocasa.mdx +++ b/docs/source/robocasa.mdx @@ -166,7 +166,7 @@ lerobot-train \ --output_dir=./outputs/smolvla_robocasa_CloseFridge \ --steps=100000 \ --batch_size=4 \ - --eval_freq=5000 \ + --env_eval_freq=5000 \ --eval.batch_size=1 \ --eval.n_episodes=5 \ --save_freq=10000 diff --git a/docs/source/so101.mdx b/docs/source/so101.mdx index 1274b8282..5b4ed0985 100644 --- a/docs/source/so101.mdx +++ b/docs/source/so101.mdx @@ -122,7 +122,7 @@ The video below shows the sequence of steps for setting the motor ids. #### Follower -Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your leader arm a name with the `id` parameter. +Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your follower arm a name with the `id` parameter. diff --git a/docs/source/streaming_video_encoding.mdx b/docs/source/streaming_video_encoding.mdx index 96e049eb3..0be32b717 100644 --- a/docs/source/streaming_video_encoding.mdx +++ b/docs/source/streaming_video_encoding.mdx @@ -17,7 +17,7 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti | Parameter | CLI Flag | Type | Default | Description | | ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- | | `streaming_encoding` | `--dataset.streaming_encoding` | `bool` | `True` | Enable real-time encoding during capture | -| `vcodec` | `--dataset.camera_encoder.vcodec` | `str` | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder | +| `vcodec` | `--dataset.rgb_encoder.vcodec` | `str` | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder | | `encoder_threads` | `--dataset.encoder_threads` | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide | | `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int` | `30` | Max buffered frames per camera (~1s at 30fps). Consumes RAM | @@ -82,15 +82,15 @@ Use HW encoding when: ### Available HW Encoders -| Encoder | Platform | Hardware | CLI Value | -| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- | -| `h264_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.camera_encoder.vcodec=h264_videotoolbox` | -| `hevc_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.camera_encoder.vcodec=hevc_videotoolbox` | -| `h264_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.camera_encoder.vcodec=h264_nvenc` | -| `hevc_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.camera_encoder.vcodec=hevc_nvenc` | -| `h264_vaapi` | Linux | Intel/AMD GPU | `--dataset.camera_encoder.vcodec=h264_vaapi` | -| `h264_qsv` | Linux/Windows | Intel Quick Sync | `--dataset.camera_encoder.vcodec=h264_qsv` | -| `auto` | Any | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.camera_encoder.vcodec=auto` | +| Encoder | Platform | Hardware | CLI Value | +| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------ | +| `h264_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.rgb_encoder.vcodec=h264_videotoolbox` | +| `hevc_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.rgb_encoder.vcodec=hevc_videotoolbox` | +| `h264_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.rgb_encoder.vcodec=h264_nvenc` | +| `hevc_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.rgb_encoder.vcodec=hevc_nvenc` | +| `h264_vaapi` | Linux | Intel/AMD GPU | `--dataset.rgb_encoder.vcodec=h264_vaapi` | +| `h264_qsv` | Linux/Windows | Intel Quick Sync | `--dataset.rgb_encoder.vcodec=h264_qsv` | +| `auto` | Any | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.rgb_encoder.vcodec=auto` | > [!NOTE] > In order to use the HW accelerated encoders you might need to upgrade your GPU drivers. @@ -100,15 +100,15 @@ Use HW encoding when: ## 5. Troubleshooting -| Symptom | Likely Cause | Fix | -| ------------------------------------------------------------------ | -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage) | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.camera_encoder.vcodec=auto`) | -| "Encoder queue full" warnings or dropped frames in dataset | Encoder can't keep up (Queue overflow) | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.camera_encoder.vcodec=auto`). | -| High RAM usage | Queue filling faster than encoding | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding | -| Large video files | Using HW encoder or H.264 | Expected trade-off. Switch to `libsvtav1` if CPU allows | -| `save_episode()` still slow | `streaming_encoding` is `False` | Set `--dataset.streaming_encoding=true` | -| Encoder thread crash | Codec not available or invalid settings | Check `vcodec` is installed, try `--dataset.camera_encoder.vcodec=auto` | -| Recorded dataset is missing frames | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected. | +| Symptom | Likely Cause | Fix | +| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | +| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage) | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.rgb_encoder.vcodec=auto`) | +| "Encoder queue full" warnings or dropped frames in dataset | Encoder can't keep up (Queue overflow) | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.rgb_encoder.vcodec=auto`). | +| High RAM usage | Queue filling faster than encoding | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding | +| Large video files | Using HW encoder or H.264 | Expected trade-off. Switch to `libsvtav1` if CPU allows | +| `save_episode()` still slow | `streaming_encoding` is `False` | Set `--dataset.streaming_encoding=true` | +| Encoder thread crash | Codec not available or invalid settings | Check `vcodec` is installed, try `--dataset.rgb_encoder.vcodec=auto` | +| Recorded dataset is missing frames | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected. | ## 6. Recommended Configurations @@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the # 2camsx 640x480x3 @30fps: Requires some tuning. # Use H.264, disable streaming, consider batching encoding -lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ... +lerobot-record --dataset.rgb_encoder.vcodec=h264 --dataset.streaming_encoding=false ... ``` ## 7. Closing note diff --git a/docs/source/using_dataset_tools.mdx b/docs/source/using_dataset_tools.mdx index 49247a6c1..a6dcdb1a7 100644 --- a/docs/source/using_dataset_tools.mdx +++ b/docs/source/using_dataset_tools.mdx @@ -11,8 +11,9 @@ LeRobot provides several utilities for manipulating datasets: 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids` 4. **Add Features** - Add new features to a dataset 5. **Remove Features** - Remove features from a dataset -6. **Convert to Video** - Convert image-based datasets to video format for efficient storage -7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc. +6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders) +7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings +8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc. The core implementation is in `lerobot.datasets.dataset_tools`. An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`. @@ -117,10 +118,19 @@ lerobot-edit-dataset \ --repo_id lerobot/pusht_image \ --operation.type convert_image_to_video \ --operation.output_dir outputs/pusht_video \ - --operation.camera_encoder.vcodec libsvtav1 \ - --operation.camera_encoder.pix_fmt yuv420p \ - --operation.camera_encoder.g 2 \ - --operation.camera_encoder.crf 30 + --operation.rgb_encoder.vcodec libsvtav1 \ + --operation.rgb_encoder.pix_fmt yuv420p \ + --operation.rgb_encoder.g 2 \ + --operation.rgb_encoder.crf 30 + +# Convert a dataset that includes depth maps, customizing the depth encoder +lerobot-edit-dataset \ + --repo_id lerobot/pusht_image \ + --operation.type convert_image_to_video \ + --operation.output_dir outputs/pusht_video \ + --operation.depth_encoder.depth_min 0.01 \ + --operation.depth_encoder.depth_max 10.0 \ + --operation.depth_encoder.use_log true # Convert only specific episodes lerobot-edit-dataset \ @@ -147,11 +157,42 @@ lerobot-edit-dataset \ **Parameters:** - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`) -- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.. See [Video Encoding Parameters](./video_encoding_parameters) for more details. +- `rgb_encoder`: Video encoder settings applied to RGB cameras — all sub-fields accessible via `--operation.rgb_encoder.`. See [Video Encoding Parameters](./video_encoding_parameters) for more details. +- `depth_encoder`: Video encoder settings applied to depth-map cameras (e.g. from an Intel RealSense). In addition to the standard encoder fields it exposes the depth quantization knobs (`depth_min`, `depth_max`, `shift`, `use_log`), accessible via `--operation.depth_encoder.`. These quantization settings are persisted to the dataset metadata so depth can be dequantized back to physical units on load. See the [Depth streams](./video_encoding_parameters#depth-streams) section for details. - `episode_indices`: List of specific episodes to convert (default: all episodes) - `num_workers`: Number of parallel workers for processing (default: 4) -**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved. +**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). Depth-map cameras are detected automatically and routed to the `depth_encoder`, while RGB cameras use the `rgb_encoder`. All episodes, stats, and tasks are preserved. + +#### Re-encode Videos + +Re-encode the videos of an existing video dataset with different encoder settings, without going back to raw frames. RGB videos use the `rgb_encoder` and depth videos use the `depth_encoder`. Provide only the encoder(s) you want to re-encode; the other stream type is left untouched. + +```bash +# Re-encode all RGB videos with new settings (saves to lerobot/pusht_reencoded by default) +lerobot-edit-dataset \ + --repo_id lerobot/pusht \ + --operation.type reencode_videos \ + --operation.rgb_encoder.vcodec h264 \ + --operation.rgb_encoder.pix_fmt yuv420p \ + --operation.rgb_encoder.crf 23 + +# Re-encode both RGB and depth videos in a dataset with depth maps +lerobot-edit-dataset \ + --repo_id lerobot/pusht_depth \ + --operation.type reencode_videos \ + --operation.rgb_encoder.vcodec h264 \ + --operation.depth_encoder.crf 50 +``` + +**Parameters:** + +- `rgb_encoder`: Encoder settings applied to every RGB video. Omit to skip re-encoding RGB videos. +- `depth_encoder`: Encoder settings applied to every depth video. Omit to skip re-encoding depth videos. +- `num_workers`: Number of parallel workers for processing. + +> [!NOTE] +> When re-encoding depth videos, the existing depth quantization parameters (`depth_min`, `depth_max`, `shift`, `use_log`) and the `is_depth_map` flag are **preserved** — re-encoding only changes the codec/quality of the stored stream, not how depth is dequantized on load. ### Show the information of datasets @@ -224,6 +265,8 @@ lerobot-dataset-viz \ Once executed, the tool opens `rerun.io` and displays the camera streams, robot states, and actions for the selected episode. +To use [Foxglove](https://foxglove.dev) instead of Rerun, install the extra add `--display-mode foxglove`. This starts a WebSocket server (connect the Foxglove app to `ws://127.0.0.1:8765`) that serves the episode as a seekable timeline you can play/pause and scrub. + For advanced usage—including visualizing datasets stored on a remote server—run: ```bash diff --git a/docs/source/video_encoding_parameters.mdx b/docs/source/video_encoding_parameters.mdx index 0b5b99b2b..132d25056 100644 --- a/docs/source/video_encoding_parameters.mdx +++ b/docs/source/video_encoding_parameters.mdx @@ -2,15 +2,15 @@ When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands. -Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `camera_encoder`, a nested `VideoEncoderConfig` (`lerobot.configs.video.VideoEncoderConfig`) passed through PyAV. +Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `rgb_encoder`, a nested `RGBEncoderConfig` (`lerobot.configs.video.RGBEncoderConfig`) passed through PyAV. -You can set these parameters from the CLI with `--dataset.camera_encoder.` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run. +You can set these parameters from the CLI with `--dataset.rgb_encoder.` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run. - Video storage must be on for `camera_encoder` to have any effect — + Video storage must be on for `rgb_encoder` to have any effect — `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the - recording default). With video off, inputs stay as images and `camera_encoder` - is ignored. + recording default). With video off, inputs stay as images and `rgb_encoder` is + ignored. For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark). @@ -33,9 +33,9 @@ lerobot-record \ --dataset.single_task="Grab the cube" \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2 \ - --dataset.camera_encoder.vcodec=h264 \ - --dataset.camera_encoder.preset=fast \ - --dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \ + --dataset.rgb_encoder.vcodec=h264 \ + --dataset.rgb_encoder.preset=fast \ + --dataset.rgb_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \ --display_data=true ``` @@ -50,7 +50,7 @@ Only override these parameters if you have a specific reason to, and measure the -All flags below are prefixed with `--dataset.camera_encoder.` on the CLI. +All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI. | Parameter | Type | Default | Description | | --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | @@ -65,6 +65,77 @@ All flags below are prefixed with `--dataset.camera_encoder.` on the CLI. --- +## Depth streams + +Depth maps (Intel RealSense, Reachy 2) are stored as their **own video streams** alongside the RGB streams. Raw depth (`uint16` millimetres or `float32` metres) can't survive an 8-bit codec, so LeRobot **quantizes** each map to a 12-bit code (`[0, 4095]`) — logarithmically by default, to match the `1/depth` error profile of depth sensors — then packs it into a high-bit-depth pixel format (`gray12le`) and encodes it with a 12-bit codec. + +```mermaid +flowchart LR + A["Raw depth (uint16 mm / float32 m)"] --> B["Clip to depth_min, depth_max"] + B --> C["Quantize to 12-bit code 0–4095 (log or linear)"] + C --> D["Pack into gray12le"] + D --> E["Encode video (hevc Main 12)"] + E --> F[("MP4 + metadata: depth_min/max, shift, use_log")] + F -. "load time (depth_output_unit)" .-> G["Dequantize to mm or m"] + + classDef input fill:#e3f2fd,stroke:#1565c0,color:#0d47a1; + classDef encode fill:#ede7f6,stroke:#5e35b1,color:#311b92; + classDef store fill:#fff8e1,stroke:#f9a825,color:#e65100; + classDef load fill:#e8f5e9,stroke:#2e7d32,color:#1b5e20; + + class A input; + class B,C,D,E encode; + class F store; + class G load; +``` + +Configure the depth pipeline through a parallel **`depth_encoder`** block (`DepthEncoderConfig`). It shares every `RGBEncoderConfig` field (`vcodec`, `pix_fmt`, `crf`, …) and adds four quantizer knobs, set via `--dataset.depth_encoder.`: + +```bash +lerobot-record \ + ... \ + --dataset.depth_encoder.vcodec=hevc \ + --dataset.depth_encoder.depth_min=0.05 \ + --dataset.depth_encoder.depth_max=5.0 \ + --dataset.depth_encoder.use_log=true +``` + +| Parameter | Type | Default | Description | +| --------------- | ------- | ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | +| `vcodec` | `str` | `"hevc"` | HEVC Main 12 (a 12-bit-capable codec, MP4-compatible). | +| `extra_options` | `dict` | `{"x265-params": "lossless=1"}` | **Depth defaults to lossless** (exact round-trip); `crf` is ignored. Pass `extra_options={}` and set `crf` for a smaller lossy stream. | +| `pix_fmt` | `str` | `"gray12le"` | Single-channel 12-bit pixel format used to carry the quantized codes. | +| `depth_min` | `float` | `0.01` | Depth in metres mapped to quantum `0`. Values below are clipped on decode. | +| `depth_max` | `float` | `10.0` | Depth in metres mapped to quantum `4095`. Values above are clipped on decode. | +| `shift` | `float` | `3.5` | Pre-log offset (metres) used in logarithmic quantization for numerical stability near zero. Must satisfy `depth_min + shift > 0`. | +| `use_log` | `bool` | `True` | If `true`, quantize in log-space (recommended for typical depth sensors). Set to `false` for uniform/linear quantization. | + +> [!TIP] +> `depth_min`, `depth_max`, and `shift` are always interpreted in **metres**, regardless of the input depth's unit. Inputs are auto-detected: integer arrays (e.g. `uint16` millimetres straight from a RealSense) are treated as millimetres, floating arrays as metres. +> Pick `depth_min` / `depth_max` to bracket the actual working range of your sensor — quanta outside that range saturate, which can crush detail at the boundaries. + +Depth features are flagged with `"is_depth_map": true` in `meta/info.json`, and their quantizer settings (`video.depth_min`, `video.depth_max`, `video.shift`, `video.use_log`) are persisted — which is what lets depth be **dequantized back to physical units** on load. + +### Output unit at load time + +`depth_encoder` is a **record-time** concern. The unit that depth maps are dequantized to on _load_ (e.g. during training) is set separately by the read-time flag `--dataset.depth_output_unit`: + +```bash +lerobot-train \ + --dataset.repo_id=/ \ + --dataset.depth_output_unit=m \ + --policy.type=act +``` + +| Parameter | Type | Default | Description | +| ------------------- | ----- | ------- | -------------------------------------------------------------------------------------------- | +| `depth_output_unit` | `str` | `"mm"` | Physical unit depth maps are dequantized to on load: `"mm"` (millimetres) or `"m"` (metres). | + +> [!TIP] +> This is purely a decode-time presentation choice — it does **not** alter the stored video or its metadata, so the same dataset can be read as `mm` or `m` without re-encoding. It has no effect on datasets without depth cameras. + +--- + ## Persistence in dataset metadata After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.`, the layout in `info.json` is: @@ -82,7 +153,7 @@ After the first episode of a video stream is encoded, the encoder configuration "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, - "video.is_depth_map": false, + "is_depth_map": false, "video.g": 2, "video.crf": 30, "video.preset": "fast", @@ -97,12 +168,12 @@ After the first episode of a video stream is encoded, the encoder configuration Two sources contribute to the `info` block: -- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `video.is_depth_map`, plus `audio.*` if an audio stream is present. -- **Encoder-derived** (taken from `VideoEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`. +- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `is_depth_map`, plus `audio.*` if an audio stream is present. +- **Encoder-derived** (taken from `RGBEncoderConfig` or `DepthEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`. This block is populated **once**, from the **first** episode. It assumes every - episode in the dataset was encoded with the same `camera_encoder`. Changing + episode in the dataset was encoded with the same `rgb_encoder`. Changing encoder settings partway through a recording is not supported — the `info.json` will only reflect the parameters used for the first episode. diff --git a/docs/source/vlabench.mdx b/docs/source/vlabench.mdx index da579d674..9d45da4ec 100644 --- a/docs/source/vlabench.mdx +++ b/docs/source/vlabench.mdx @@ -165,7 +165,7 @@ lerobot-train \ --output_dir=./outputs/smolvla_vlabench_primitive \ --steps=100000 \ --batch_size=4 \ - --eval_freq=5000 \ + --env_eval_freq=5000 \ --eval.batch_size=1 \ --eval.n_episodes=1 \ --save_freq=10000 diff --git a/examples/lekiwi/evaluate.py b/examples/lekiwi/evaluate.py index 3ddcd1f14..13bb6ac28 100644 --- a/examples/lekiwi/evaluate.py +++ b/examples/lekiwi/evaluate.py @@ -17,7 +17,7 @@ import logging import time -from lerobot.common.control_utils import init_keyboard_listener, predict_action +from lerobot.common.control_utils import predict_action from lerobot.datasets import LeRobotDataset from lerobot.policies import make_pre_post_processors from lerobot.policies.act import ACTPolicy @@ -26,6 +26,7 @@ from lerobot.processor import make_default_processors from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig from lerobot.utils.constants import ACTION, OBS_STR from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.robot_utils import precise_sleep from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun, log_rerun_data diff --git a/examples/lekiwi/record.py b/examples/lekiwi/record.py index 2c581f5ff..f62a9eb49 100644 --- a/examples/lekiwi/record.py +++ b/examples/lekiwi/record.py @@ -14,7 +14,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from lerobot.common.control_utils import init_keyboard_listener from lerobot.datasets import LeRobotDataset from lerobot.processor import make_default_processors from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig @@ -23,6 +22,7 @@ from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig from lerobot.utils.constants import ACTION, OBS_STR from lerobot.utils.feature_utils import hw_to_dataset_features +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun diff --git a/examples/phone_to_so100/evaluate.py b/examples/phone_to_so100/evaluate.py index e859123d0..d1fb4de67 100644 --- a/examples/phone_to_so100/evaluate.py +++ b/examples/phone_to_so100/evaluate.py @@ -18,7 +18,7 @@ import logging import time from lerobot.cameras.opencv import OpenCVCameraConfig -from lerobot.common.control_utils import init_keyboard_listener, predict_action +from lerobot.common.control_utils import predict_action from lerobot.configs import FeatureType, PolicyFeature from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features from lerobot.model.kinematics import RobotKinematics @@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import ( from lerobot.types import RobotAction, RobotObservation from lerobot.utils.constants import ACTION, OBS_STR from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.robot_utils import precise_sleep from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun, log_rerun_data diff --git a/examples/phone_to_so100/record.py b/examples/phone_to_so100/record.py index 87b8e49fd..612e94ab9 100644 --- a/examples/phone_to_so100/record.py +++ b/examples/phone_to_so100/record.py @@ -15,7 +15,6 @@ # limitations under the License. from lerobot.cameras.opencv import OpenCVCameraConfig -from lerobot.common.control_utils import init_keyboard_listener from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features from lerobot.model.kinematics import RobotKinematics from lerobot.processor import ( @@ -39,6 +38,7 @@ from lerobot.teleoperators.phone.config_phone import PhoneOS from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction from lerobot.types import RobotAction, RobotObservation from lerobot.utils.feature_utils import combine_feature_dicts +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun diff --git a/examples/so100_to_so100_EE/evaluate.py b/examples/so100_to_so100_EE/evaluate.py index 63def68d0..2a2022623 100644 --- a/examples/so100_to_so100_EE/evaluate.py +++ b/examples/so100_to_so100_EE/evaluate.py @@ -18,7 +18,7 @@ import logging import time from lerobot.cameras.opencv import OpenCVCameraConfig -from lerobot.common.control_utils import init_keyboard_listener, predict_action +from lerobot.common.control_utils import predict_action from lerobot.configs import FeatureType, PolicyFeature from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features from lerobot.model.kinematics import RobotKinematics @@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import ( from lerobot.types import RobotAction, RobotObservation from lerobot.utils.constants import ACTION, OBS_STR from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.robot_utils import precise_sleep from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun, log_rerun_data diff --git a/examples/so100_to_so100_EE/record.py b/examples/so100_to_so100_EE/record.py index a0b92da3b..3706ee4f5 100644 --- a/examples/so100_to_so100_EE/record.py +++ b/examples/so100_to_so100_EE/record.py @@ -16,7 +16,6 @@ from lerobot.cameras.opencv import OpenCVCameraConfig -from lerobot.common.control_utils import init_keyboard_listener from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features from lerobot.model.kinematics import RobotKinematics from lerobot.processor import ( @@ -36,6 +35,7 @@ from lerobot.scripts.lerobot_record import record_loop from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig from lerobot.types import RobotAction, RobotObservation from lerobot.utils.feature_utils import combine_feature_dicts +from lerobot.utils.keyboard_input import init_keyboard_listener from lerobot.utils.utils import log_say from lerobot.utils.visualization_utils import init_rerun diff --git a/pyproject.toml b/pyproject.toml index 0dc86d7ff..5f9e0adc5 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -124,7 +124,8 @@ hardware = [ "lerobot[deepdiff-dep]", ] viz = [ - "rerun-sdk>=0.24.0,<0.27.0", + "rerun-sdk>=0.24.0,<0.34.0", + "foxglove-sdk>=0.25.1,<0.26.0", ] # ── User-facing composite extras (map to CLI scripts) ───── # lerobot-record, lerobot-replay, lerobot-calibrate, lerobot-teleoperate, etc. @@ -140,7 +141,14 @@ av-dep = ["av>=15.0.0,<16.0.0"] pygame-dep = ["pygame>=2.5.1,<2.7.0"] # NOTE: 0.9.16 links against liburdfdom_sensor.so.4, which is unavailable on Ubuntu 24.04 # (noble ships urdfdom 3.x). Cap below 0.9.16 until system urdfdom 4.x is broadly available. -placo-dep = ["placo>=0.9.6,<0.9.16"] +# +# NOTE: placo pulls in pin (Pinocchio), whose binary wheels dlopen specific cmeel sonames +# (liburdfdom_sensor.so.4.0, libtinyxml2.so.10) but declare only `>=` floors on their cmeel +# packages. The 2026-05-21 major bumps (cmeel-urdfdom 6.0.0 -> .so.6, cmeel-tinyxml2 11.0.0 +# -> .so.11) ship newer sonames, so left unpinned the resolver grabs them and `import placo` +# fails at load with "liburdfdom_sensor.so.4.0: cannot open shared object file" (see #3755). +# There is no cmeel-urdfdom 5.x; <5 selects the 4.x ABI the placo/pin wheels are built against. +placo-dep = ["placo>=0.9.6,<0.9.16", "cmeel-urdfdom>=4,<5", "cmeel-tinyxml2<11"] transformers-dep = ["transformers>=5.4.0,<5.6.0"] grpcio-dep = ["grpcio>=1.73.1,<2.0.0", "protobuf>=6.31.1,<8.0.0"] accelerate-dep = ["accelerate>=1.14.0,<2.0.0"] @@ -156,6 +164,7 @@ pynput-dep = ["pynput>=1.7.8,<1.9.0"] pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"] motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"] motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"] +timm-dep = ["timm>=1.0.0,<1.1.0"] # Motors feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"] @@ -211,19 +220,24 @@ groot = [ "lerobot[transformers-dep]", "lerobot[peft-dep]", "lerobot[diffusers-dep]", + "lerobot[dataset]", # NOTE: processor_groot builds a LeRobotDataset for relative-action training stats "dm-tree>=0.1.8,<1.0.0", - "timm>=1.0.0,<1.1.0", + "lerobot[timm-dep]", "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')", - "ninja>=1.11.1,<2.0.0", - "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'" ] sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"] robometer = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]", "lerobot[peft-dep]"] topreward = ["lerobot[transformers-dep]"] xvla = ["lerobot[transformers-dep]"] eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"] +fastwam = [ + "lerobot[transformers-dep]", + "lerobot[diffusers-dep]", +] +evo1 = ["lerobot[transformers-dep]"] hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.14,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"] vla_jepa = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[qwen-vl-utils-dep]"] +lingbot_va = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[accelerate-dep]"] # Features async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"] @@ -301,10 +315,13 @@ all = [ "lerobot[pi]", "lerobot[molmoact2]", "lerobot[smolvla]", - # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn + "lerobot[fastwam]", + "lerobot[groot]", "lerobot[xvla]", + "lerobot[evo1]", "lerobot[hilserl]", "lerobot[vla_jepa]", + "lerobot[lingbot_va]", "lerobot[async]", "lerobot[dev]", "lerobot[test]", @@ -437,7 +454,8 @@ default.extend-ignore-identifiers-re = [ "is_compileable", "ROBOTIS", "OT_VALUE", - "VanderBilt" + "VanderBilt", + "seperated_timestep", ] # TODO: Uncomment when ready to use diff --git a/requirements-macos.txt b/requirements-macos.txt deleted file mode 100644 index c5bbe1c8a..000000000 --- a/requirements-macos.txt +++ /dev/null @@ -1,729 +0,0 @@ -# -# This file is autogenerated by pip-compile with Python 3.12 -# by the following command: -# -# pip-compile --output-file=requirements-macos.txt requirements.in -# --e .[all] - # via -[all] -absl-py==2.4.0 - # via - # dm-control - # dm-env - # dm-tree - # labmaze - # mujoco -accelerate==1.13.0 - # via - # lerobot - # peft -aiohappyeyeballs==2.6.1 - # via aiohttp -aiohttp==3.13.3 - # via fsspec -aiosignal==1.4.0 - # via aiohttp -annotated-doc==0.0.4 - # via - # fastapi - # typer -annotated-types==0.7.0 - # via pydantic -anyio==4.12.1 - # via - # httpx - # starlette - # watchfiles -asttokens==3.0.1 - # via stack-data -attrs==25.4.0 - # via - # aiohttp - # dm-tree - # jsonlines - # rerun-sdk -av==15.1.0 - # via - # lerobot - # qwen-vl-utils -certifi==2026.2.25 - # via - # httpcore - # httpx - # requests - # sentry-sdk -cffi==2.0.0 - # via pymunk -cfgv==3.5.0 - # via pre-commit -charset-normalizer==3.4.5 - # via requests -click==8.3.1 - # via - # typer - # uvicorn - # wandb -cloudpickle==3.1.2 - # via gymnasium -cmake==4.1.3 - # via lerobot -cmeel==0.59.0 - # via - # cmeel-assimp - # cmeel-boost - # cmeel-console-bridge - # cmeel-octomap - # cmeel-qhull - # cmeel-tinyxml2 - # cmeel-urdfdom - # cmeel-zlib - # coal-library - # eigenpy - # eiquadprog - # pin - # placo - # rhoban-cmeel-jsoncpp -cmeel-assimp==5.4.3.1 - # via coal-library -cmeel-boost==1.87.0.1 - # via - # coal-library - # eigenpy - # eiquadprog - # pin -cmeel-console-bridge==1.0.2.3 - # via cmeel-urdfdom -cmeel-octomap==1.10.0 - # via coal-library -cmeel-qhull==8.0.2.1 - # via coal-library -cmeel-tinyxml2==10.0.0 - # via cmeel-urdfdom -cmeel-urdfdom==4.0.1 - # via pin -cmeel-zlib==1.3.1 - # via cmeel-assimp -coal-library==3.0.1 - # via pin -contourpy==1.3.3 - # via - # lerobot - # matplotlib -coverage[toml]==7.13.4 - # via pytest-cov -cycler==0.12.1 - # via matplotlib -datasets==4.6.1 - # via lerobot -debugpy==1.8.20 - # via lerobot -decorator==5.2.1 - # via ipython -deepdiff==8.6.1 - # via lerobot -diffusers==0.35.2 - # via lerobot -dill==0.4.0 - # via - # datasets - # multiprocess -distlib==0.4.0 - # via virtualenv -dm-control==1.0.37 - # via gym-aloha -dm-env==1.6 - # via dm-control -dm-tree==0.1.9 - # via - # dm-control - # dm-env -docopt==0.6.2 - # via num2words -draccus==0.10.0 - # via lerobot -dynamixel-sdk==3.8.4 - # via lerobot -eigenpy==3.10.3 - # via coal-library -einops==0.8.2 - # via lerobot -eiquadprog==1.2.9 - # via placo -etils[epath,epy]==1.14.0 - # via mujoco -executing==2.2.1 - # via stack-data -faker==34.0.2 - # via lerobot -farama-notifications==0.0.4 - # via gymnasium -fastapi==0.135.1 - # via - # lerobot - # teleop -feetech-servo-sdk==1.0.0 - # via lerobot -filelock==3.25.0 - # via - # datasets - # diffusers - # huggingface-hub - # python-discovery - # torch - # virtualenv -fonttools==4.61.1 - # via matplotlib -frozenlist==1.8.0 - # via - # aiohttp - # aiosignal -fsspec[http]==2026.2.0 - # via - # datasets - # etils - # huggingface-hub - # torch -gitdb==4.0.12 - # via gitpython -gitpython==3.1.46 - # via wandb -glfw==2.10.0 - # via - # dm-control - # mujoco -grpcio==1.73.1 - # via - # grpcio-tools - # lerobot - # reachy2-sdk - # reachy2-sdk-api -grpcio-tools==1.73.1 - # via - # lerobot - # reachy2-sdk-api -gym-aloha==0.1.3 - # via lerobot -gym-hil==0.1.13 - # via lerobot -gym-pusht==0.1.6 - # via lerobot -gymnasium==1.2.3 - # via - # gym-aloha - # gym-hil - # gym-pusht - # lerobot - # metaworld -h11==0.16.0 - # via - # httpcore - # uvicorn -hebi-py==2.11.0 - # via lerobot -hf-xet==1.3.2 - # via huggingface-hub -hidapi==0.14.0.post4 - # via - # gym-hil - # lerobot -httpcore==1.0.9 - # via httpx -httptools==0.7.1 - # via uvicorn -httpx==0.28.1 - # via - # datasets - # huggingface-hub -huggingface-hub==1.6.0 - # via - # accelerate - # datasets - # diffusers - # lerobot - # peft - # tokenizers - # transformers -identify==2.6.17 - # via pre-commit -idna==3.11 - # via - # anyio - # httpx - # requests - # yarl -imageio[ffmpeg]==2.37.2 - # via - # gym-aloha - # gym-hil - # lerobot - # metaworld - # scikit-image -imageio-ffmpeg==0.6.0 - # via imageio -importlib-metadata==8.7.1 - # via diffusers -iniconfig==2.3.0 - # via pytest -ipython==9.11.0 - # via meshcat -ipython-pygments-lexers==1.1.1 - # via ipython -ischedule==1.2.7 - # via placo -jedi==0.19.2 - # via ipython -jinja2==3.1.6 - # via torch -jsonlines==4.0.0 - # via lerobot -kiwisolver==1.4.9 - # via matplotlib -labmaze==1.0.6 - # via dm-control -lazy-loader==0.5 - # via scikit-image -librt==0.8.1 - # via mypy -lxml==6.0.2 - # via dm-control -markdown-it-py==4.0.0 - # via rich -markupsafe==3.0.3 - # via jinja2 -matplotlib==3.10.8 - # via lerobot -matplotlib-inline==0.2.1 - # via ipython -mdurl==0.1.2 - # via markdown-it-py -mergedeep==1.3.4 - # via draccus -meshcat==0.3.2 - # via placo -metaworld==3.0.0 - # via lerobot -mock-serial==0.0.1 - # via lerobot -mpmath==1.3.0 - # via sympy -mujoco==3.5.0 - # via - # dm-control - # gym-aloha - # gym-hil - # metaworld -multidict==6.7.1 - # via - # aiohttp - # yarl -multiprocess==0.70.18 - # via datasets -mypy==1.19.1 - # via lerobot -mypy-extensions==1.1.0 - # via - # mypy - # typing-inspect -networkx==3.6.1 - # via - # scikit-image - # torch -nodeenv==1.10.0 - # via pre-commit -num2words==0.5.14 - # via lerobot -numpy==2.2.6 - # via - # accelerate - # cmeel-boost - # contourpy - # datasets - # diffusers - # dm-control - # dm-env - # dm-tree - # gymnasium - # hebi-py - # imageio - # labmaze - # lerobot - # matplotlib - # meshcat - # metaworld - # mujoco - # opencv-python - # opencv-python-headless - # pandas - # peft - # pyquaternion - # reachy2-sdk - # rerun-sdk - # scikit-image - # scipy - # shapely - # teleop - # tifffile - # torchvision - # transformers - # transforms3d -opencv-python==4.13.0.92 - # via - # gym-pusht - # reachy2-sdk -opencv-python-headless==4.12.0.88 - # via lerobot -orderly-set==5.5.0 - # via deepdiff -packaging==25.0 - # via - # accelerate - # datasets - # huggingface-hub - # lazy-loader - # lerobot - # matplotlib - # peft - # pytest - # qwen-vl-utils - # reachy2-sdk - # scikit-image - # transformers - # wandb -pandas==2.3.3 - # via - # datasets - # lerobot -parso==0.8.6 - # via jedi -pathspec==1.0.4 - # via mypy -peft==0.18.1 - # via lerobot -pexpect==4.9.0 - # via ipython -pillow==12.1.1 - # via - # diffusers - # imageio - # matplotlib - # meshcat - # qwen-vl-utils - # rerun-sdk - # scikit-image - # torchvision -pin==3.4.0 - # via placo -placo==0.9.16 - # via lerobot -platformdirs==4.9.4 - # via - # python-discovery - # virtualenv - # wandb -pluggy==1.6.0 - # via - # pytest - # pytest-cov -pre-commit==4.5.1 - # via lerobot -prompt-toolkit==3.0.52 - # via ipython -propcache==0.4.1 - # via - # aiohttp - # yarl -protobuf==6.31.1 - # via - # dm-control - # grpcio-tools - # lerobot - # reachy2-sdk - # reachy2-sdk-api - # wandb -psutil==7.2.2 - # via - # accelerate - # imageio - # peft -ptyprocess==0.7.0 - # via pexpect -pure-eval==0.2.3 - # via stack-data -pyarrow==23.0.1 - # via - # datasets - # rerun-sdk -pycparser==3.0 - # via cffi -pydantic==2.12.5 - # via - # fastapi - # wandb -pydantic-core==2.41.5 - # via pydantic -pygame==2.6.1 - # via - # gym-hil - # gym-pusht - # lerobot -pygments==2.19.2 - # via - # ipython - # ipython-pygments-lexers - # pytest - # rich -pymunk==6.11.1 - # via - # gym-pusht - # lerobot -pyngrok==7.5.1 - # via meshcat -pynput==1.8.1 - # via - # gym-hil - # lerobot -pyobjc-core==12.1 - # via - # pyobjc-framework-applicationservices - # pyobjc-framework-cocoa - # pyobjc-framework-coretext - # pyobjc-framework-quartz -pyobjc-framework-applicationservices==12.1 - # via pynput -pyobjc-framework-cocoa==12.1 - # via - # pyobjc-framework-applicationservices - # pyobjc-framework-coretext - # pyobjc-framework-quartz -pyobjc-framework-coretext==12.1 - # via pyobjc-framework-applicationservices -pyobjc-framework-quartz==12.1 - # via - # pynput - # pyobjc-framework-applicationservices - # pyobjc-framework-coretext -pyopengl==3.1.10 - # via - # dm-control - # mujoco -pyparsing==3.3.2 - # via - # dm-control - # matplotlib -pyquaternion==0.9.9 - # via reachy2-sdk -pyrealsense2-macosx==2.56.5 - # via lerobot -pyserial==3.5 - # via - # dynamixel-sdk - # feetech-servo-sdk - # lerobot -pytest==8.4.2 - # via - # lerobot - # pytest-cov - # pytest-timeout - # teleop -pytest-cov==7.0.0 - # via lerobot -pytest-timeout==2.4.0 - # via lerobot -python-dateutil==2.9.0.post0 - # via - # faker - # matplotlib - # pandas -python-discovery==1.1.1 - # via virtualenv -python-dotenv==1.2.2 - # via uvicorn -pytz==2026.1.post1 - # via pandas -pyyaml==6.0.3 - # via - # accelerate - # datasets - # draccus - # hebi-py - # huggingface-hub - # peft - # pre-commit - # pyngrok - # pyyaml-include - # transformers - # uvicorn - # wandb -pyyaml-include==1.4.1 - # via draccus -pyzmq==27.1.0 - # via - # lerobot - # meshcat -qwen-vl-utils==0.0.14 - # via lerobot -reachy2-sdk==1.0.15 - # via lerobot -reachy2-sdk-api==1.0.21 - # via reachy2-sdk -regex==2026.2.28 - # via - # diffusers - # transformers -requests==2.32.5 - # via - # datasets - # diffusers - # dm-control - # qwen-vl-utils - # teleop - # wandb -rerun-sdk==0.26.2 - # via lerobot -rhoban-cmeel-jsoncpp==1.9.4.9 - # via placo -rich==14.3.3 - # via typer -safetensors==0.7.0 - # via - # accelerate - # diffusers - # lerobot - # peft - # transformers -scikit-image==0.25.2 - # via - # gym-pusht - # lerobot -scipy==1.17.1 - # via - # dm-control - # lerobot - # metaworld - # scikit-image - # torchdiffeq -sentry-sdk==2.54.0 - # via wandb -shapely==2.1.2 - # via gym-pusht -shellingham==1.5.4 - # via typer -six==1.17.0 - # via - # pynput - # python-dateutil -smmap==5.0.3 - # via gitdb -stack-data==0.6.3 - # via ipython -starlette==0.52.1 - # via fastapi -sympy==1.14.0 - # via torch -teleop==0.1.4 - # via lerobot -termcolor==3.3.0 - # via lerobot -tifffile==2026.3.3 - # via scikit-image -tokenizers==0.22.2 - # via transformers -toml==0.10.2 - # via draccus -torch==2.10.0 - # via - # accelerate - # lerobot - # peft - # torchdiffeq - # torchvision -torchcodec==0.10.0 - # via lerobot -torchdiffeq==0.2.5 - # via lerobot -torchvision==0.25.0 - # via lerobot -tornado==6.5.4 - # via meshcat -tqdm==4.67.3 - # via - # datasets - # dm-control - # huggingface-hub - # peft - # transformers -traitlets==5.14.3 - # via - # ipython - # matplotlib-inline -transformers==5.3.0 - # via - # lerobot - # peft -transforms3d==0.4.2 - # via teleop -typer==0.24.1 - # via - # huggingface-hub - # transformers -typing-extensions==4.15.0 - # via - # aiosignal - # anyio - # etils - # faker - # fastapi - # gymnasium - # huggingface-hub - # mypy - # pydantic - # pydantic-core - # rerun-sdk - # starlette - # torch - # typing-inspect - # typing-inspection - # wandb -typing-inspect==0.9.0 - # via draccus -typing-inspection==0.4.2 - # via - # fastapi - # pydantic -tzdata==2025.3 - # via pandas -u-msgpack-python==2.8.0 - # via meshcat -urllib3==2.6.3 - # via - # requests - # sentry-sdk -uvicorn[standard]==0.41.0 - # via teleop -uvloop==0.22.1 - # via uvicorn -virtualenv==21.1.0 - # via pre-commit -wandb==0.24.2 - # via lerobot -watchfiles==1.1.1 - # via uvicorn -wcwidth==0.6.0 - # via prompt-toolkit -websocket-client==1.9.0 - # via teleop -websockets==16.0 - # via uvicorn -wrapt==2.1.2 - # via dm-tree -xxhash==3.6.0 - # via datasets -yarl==1.23.0 - # via aiohttp -zipp==3.23.0 - # via - # etils - # importlib-metadata - -# The following packages are considered to be unsafe in a requirements file: -# setuptools diff --git a/requirements-ubuntu.txt b/requirements-ubuntu.txt deleted file mode 100644 index 0cdc54190..000000000 --- a/requirements-ubuntu.txt +++ /dev/null @@ -1,882 +0,0 @@ -# -# This file is autogenerated by pip-compile with Python 3.12 -# by the following command: -# -# pip-compile --output-file=requirements-ubuntu.txt requirements.in -# --e .[all] - # via -[all] -absl-py==2.4.0 - # via - # dm-control - # dm-env - # dm-tree - # labmaze - # mujoco - # tensorboard -accelerate==1.13.0 - # via - # lerobot - # peft -aiohappyeyeballs==2.6.1 - # via aiohttp -aiohttp==3.13.3 - # via fsspec -aiosignal==1.4.0 - # via aiohttp -annotated-doc==0.0.4 - # via - # fastapi - # typer -annotated-types==0.7.0 - # via pydantic -antlr4-python3-runtime==4.9.3 - # via - # hydra-core - # omegaconf -anyio==4.12.1 - # via - # httpx - # starlette - # watchfiles -asttokens==3.0.1 - # via stack-data -attrs==25.4.0 - # via - # aiohttp - # dm-tree - # jsonlines - # jsonschema - # referencing - # rerun-sdk -av==15.1.0 - # via - # lerobot - # qwen-vl-utils -bddl==1.0.1 - # via hf-libero -certifi==2026.2.25 - # via - # httpcore - # httpx - # requests - # sentry-sdk -cffi==2.0.0 - # via pymunk -cfgv==3.5.0 - # via pre-commit -charset-normalizer==3.4.5 - # via requests -click==8.3.1 - # via - # typer - # uvicorn - # wandb -cloudpickle==3.1.2 - # via - # gymnasium - # hf-libero -cmake==4.1.3 - # via lerobot -cmeel==0.59.0 - # via - # cmeel-assimp - # cmeel-boost - # cmeel-console-bridge - # cmeel-octomap - # cmeel-qhull - # cmeel-tinyxml2 - # cmeel-urdfdom - # cmeel-zlib - # coal-library - # eigenpy - # eiquadprog - # pin - # placo - # rhoban-cmeel-jsoncpp -cmeel-assimp==5.4.3.1 - # via coal-library -cmeel-boost==1.87.0.1 - # via - # coal-library - # eigenpy - # eiquadprog - # pin -cmeel-console-bridge==1.0.2.3 - # via cmeel-urdfdom -cmeel-octomap==1.10.0 - # via coal-library -cmeel-qhull==8.0.2.1 - # via coal-library -cmeel-tinyxml2==10.0.0 - # via cmeel-urdfdom -cmeel-urdfdom==4.0.1 - # via pin -cmeel-zlib==1.3.1 - # via cmeel-assimp -coal-library==3.0.1 - # via pin -contourpy==1.3.3 - # via - # lerobot - # matplotlib -coverage[toml]==7.13.4 - # via pytest-cov -cuda-bindings==12.9.4 - # via torch -cuda-pathfinder==1.4.1 - # via cuda-bindings -cycler==0.12.1 - # via matplotlib -datasets==4.6.1 - # via lerobot -debugpy==1.8.20 - # via lerobot -decorator==5.2.1 - # via ipython -deepdiff==8.6.1 - # via lerobot -diffusers==0.35.2 - # via lerobot -dill==0.4.0 - # via - # datasets - # multiprocess -distlib==0.4.0 - # via virtualenv -dm-control==1.0.37 - # via gym-aloha -dm-env==1.6 - # via dm-control -dm-tree==0.1.9 - # via - # dm-control - # dm-env -docopt==0.6.2 - # via num2words -draccus==0.10.0 - # via lerobot -dynamixel-sdk==3.8.4 - # via lerobot -easydict==1.13 - # via hf-libero -egl-probe==1.0.2 - # via robomimic -eigenpy==3.10.3 - # via coal-library -einops==0.8.2 - # via - # hf-libero - # lerobot -eiquadprog==1.2.9 - # via placo -etils[epath,epy]==1.14.0 - # via mujoco -evdev==1.9.3 - # via pynput -executing==2.2.1 - # via stack-data -faker==34.0.2 - # via lerobot -farama-notifications==0.0.4 - # via gymnasium -fastapi==0.135.1 - # via - # lerobot - # teleop -fastjsonschema==2.21.2 - # via nbformat -feetech-servo-sdk==1.0.0 - # via lerobot -filelock==3.25.0 - # via - # datasets - # diffusers - # huggingface-hub - # python-discovery - # torch - # virtualenv -fonttools==4.61.1 - # via matplotlib -frozenlist==1.8.0 - # via - # aiohttp - # aiosignal -fsspec[http]==2026.2.0 - # via - # datasets - # etils - # huggingface-hub - # torch -future==1.0.0 - # via hf-libero -gitdb==4.0.12 - # via gitpython -gitpython==3.1.46 - # via wandb -glfw==2.10.0 - # via - # dm-control - # mujoco -grpcio==1.73.1 - # via - # grpcio-tools - # lerobot - # reachy2-sdk - # reachy2-sdk-api - # tensorboard -grpcio-tools==1.73.1 - # via - # lerobot - # reachy2-sdk-api -gym-aloha==0.1.3 - # via lerobot -gym-hil==0.1.13 - # via lerobot -gym-pusht==0.1.6 - # via lerobot -gymnasium==1.2.3 - # via - # gym-aloha - # gym-hil - # gym-pusht - # hf-libero - # lerobot - # metaworld -h11==0.16.0 - # via - # httpcore - # uvicorn -h5py==3.16.0 - # via robomimic -hebi-py==2.11.0 - # via lerobot -hf-egl-probe==1.0.2 - # via hf-libero -hf-libero==0.1.3 - # via lerobot -hf-xet==1.3.2 - # via huggingface-hub -hidapi==0.14.0.post4 - # via - # gym-hil - # lerobot -httpcore==1.0.9 - # via httpx -httptools==0.7.1 - # via uvicorn -httpx==0.28.1 - # via - # datasets - # huggingface-hub -huggingface-hub==1.6.0 - # via - # accelerate - # datasets - # diffusers - # lerobot - # peft - # tokenizers - # transformers -hydra-core==1.3.2 - # via hf-libero -identify==2.6.17 - # via pre-commit -idna==3.11 - # via - # anyio - # httpx - # requests - # yarl -imageio[ffmpeg]==2.37.2 - # via - # gym-aloha - # gym-hil - # lerobot - # metaworld - # robomimic - # scikit-image -imageio-ffmpeg==0.6.0 - # via - # imageio - # robomimic -importlib-metadata==8.7.1 - # via diffusers -iniconfig==2.3.0 - # via pytest -ipython==9.11.0 - # via meshcat -ipython-pygments-lexers==1.1.1 - # via ipython -ischedule==1.2.7 - # via placo -jedi==0.19.2 - # via ipython -jinja2==3.1.6 - # via torch -jsonlines==4.0.0 - # via lerobot -jsonschema==4.26.0 - # via nbformat -jsonschema-specifications==2025.9.1 - # via jsonschema -jupyter-core==5.9.1 - # via nbformat -jupytext==1.19.1 - # via bddl -kiwisolver==1.4.9 - # via matplotlib -labmaze==1.0.6 - # via dm-control -lazy-loader==0.5 - # via scikit-image -librt==0.8.1 - # via mypy -llvmlite==0.46.0 - # via numba -lxml==6.0.2 - # via dm-control -markdown==3.10.2 - # via tensorboard -markdown-it-py==4.0.0 - # via - # jupytext - # mdit-py-plugins - # rich -markupsafe==3.0.3 - # via - # jinja2 - # werkzeug -matplotlib==3.10.8 - # via - # hf-libero - # lerobot -matplotlib-inline==0.2.1 - # via ipython -mdit-py-plugins==0.5.0 - # via jupytext -mdurl==0.1.2 - # via markdown-it-py -mergedeep==1.3.4 - # via draccus -meshcat==0.3.2 - # via placo -metaworld==3.0.0 - # via lerobot -mock-serial==0.0.1 - # via lerobot -mpmath==1.3.0 - # via sympy -mujoco==3.5.0 - # via - # dm-control - # gym-aloha - # gym-hil - # hf-libero - # metaworld - # robosuite -multidict==6.7.1 - # via - # aiohttp - # yarl -multiprocess==0.70.18 - # via datasets -mypy==1.19.1 - # via lerobot -mypy-extensions==1.1.0 - # via - # mypy - # typing-inspect -nbformat==5.10.4 - # via jupytext -networkx==3.6.1 - # via - # bddl - # scikit-image - # torch -nodeenv==1.10.0 - # via pre-commit -num2words==0.5.14 - # via lerobot -numba==0.64.0 - # via robosuite -numpy==2.2.6 - # via - # accelerate - # bddl - # cmeel-boost - # contourpy - # datasets - # diffusers - # dm-control - # dm-env - # dm-tree - # gymnasium - # h5py - # hebi-py - # hf-libero - # imageio - # labmaze - # lerobot - # matplotlib - # meshcat - # metaworld - # mujoco - # numba - # opencv-python - # opencv-python-headless - # pandas - # peft - # pyquaternion - # reachy2-sdk - # rerun-sdk - # robomimic - # robosuite - # scikit-image - # scipy - # shapely - # teleop - # tensorboard - # tensorboardx - # tifffile - # torchvision - # transformers - # transforms3d -nvidia-cublas-cu12==12.8.4.1 - # via - # nvidia-cudnn-cu12 - # nvidia-cusolver-cu12 - # torch -nvidia-cuda-cupti-cu12==12.8.90 - # via torch -nvidia-cuda-nvrtc-cu12==12.8.93 - # via torch -nvidia-cuda-runtime-cu12==12.8.90 - # via torch -nvidia-cudnn-cu12==9.10.2.21 - # via torch -nvidia-cufft-cu12==11.3.3.83 - # via torch -nvidia-cufile-cu12==1.13.1.3 - # via torch -nvidia-curand-cu12==10.3.9.90 - # via torch -nvidia-cusolver-cu12==11.7.3.90 - # via torch -nvidia-cusparse-cu12==12.5.8.93 - # via - # nvidia-cusolver-cu12 - # torch -nvidia-cusparselt-cu12==0.7.1 - # via torch -nvidia-nccl-cu12==2.27.5 - # via torch -nvidia-nvjitlink-cu12==12.8.93 - # via - # nvidia-cufft-cu12 - # nvidia-cusolver-cu12 - # nvidia-cusparse-cu12 - # torch -nvidia-nvshmem-cu12==3.4.5 - # via torch -nvidia-nvtx-cu12==12.8.90 - # via torch -omegaconf==2.3.0 - # via hydra-core -opencv-python==4.13.0.92 - # via - # gym-pusht - # hf-libero - # reachy2-sdk - # robosuite -opencv-python-headless==4.12.0.88 - # via lerobot -orderly-set==5.5.0 - # via deepdiff -packaging==25.0 - # via - # accelerate - # datasets - # huggingface-hub - # hydra-core - # jupytext - # lazy-loader - # lerobot - # matplotlib - # peft - # pytest - # qwen-vl-utils - # reachy2-sdk - # scikit-image - # tensorboard - # tensorboardx - # transformers - # wandb -pandas==2.3.3 - # via - # datasets - # lerobot -parso==0.8.6 - # via jedi -pathspec==1.0.4 - # via mypy -peft==0.18.1 - # via lerobot -pexpect==4.9.0 - # via ipython -pillow==12.1.1 - # via - # diffusers - # imageio - # matplotlib - # meshcat - # qwen-vl-utils - # rerun-sdk - # robosuite - # scikit-image - # tensorboard - # torchvision -pin==3.4.0 - # via placo -placo==0.9.16 - # via lerobot -platformdirs==4.9.4 - # via - # jupyter-core - # python-discovery - # virtualenv - # wandb -pluggy==1.6.0 - # via - # pytest - # pytest-cov -pre-commit==4.5.1 - # via lerobot -prompt-toolkit==3.0.52 - # via ipython -propcache==0.4.1 - # via - # aiohttp - # yarl -protobuf==6.31.1 - # via - # dm-control - # grpcio-tools - # lerobot - # reachy2-sdk - # reachy2-sdk-api - # tensorboard - # tensorboardx - # wandb -psutil==7.2.2 - # via - # accelerate - # imageio - # peft - # robomimic -ptyprocess==0.7.0 - # via pexpect -pure-eval==0.2.3 - # via stack-data -pyarrow==23.0.1 - # via - # datasets - # rerun-sdk -pycparser==3.0 - # via cffi -pydantic==2.12.5 - # via - # fastapi - # wandb -pydantic-core==2.41.5 - # via pydantic -pygame==2.6.1 - # via - # gym-hil - # gym-pusht - # lerobot -pygments==2.19.2 - # via - # ipython - # ipython-pygments-lexers - # pytest - # rich -pymunk==6.11.1 - # via - # gym-pusht - # lerobot -pyngrok==7.5.1 - # via meshcat -pynput==1.8.1 - # via - # gym-hil - # lerobot -pyopengl==3.1.10 - # via - # dm-control - # mujoco -pyparsing==3.3.2 - # via - # dm-control - # matplotlib -pyquaternion==0.9.9 - # via reachy2-sdk -pyrealsense2==2.56.5.9235 - # via lerobot -pyserial==3.5 - # via - # dynamixel-sdk - # feetech-servo-sdk - # lerobot -pytest==8.4.2 - # via - # bddl - # lerobot - # pytest-cov - # pytest-timeout - # teleop -pytest-cov==7.0.0 - # via lerobot -pytest-timeout==2.4.0 - # via lerobot -python-dateutil==2.9.0.post0 - # via - # faker - # matplotlib - # pandas -python-discovery==1.1.1 - # via virtualenv -python-dotenv==1.2.2 - # via uvicorn -python-xlib==0.33 - # via pynput -pytz==2026.1.post1 - # via pandas -pyyaml==6.0.3 - # via - # accelerate - # datasets - # draccus - # hebi-py - # huggingface-hub - # jupytext - # omegaconf - # peft - # pre-commit - # pyngrok - # pyyaml-include - # transformers - # uvicorn - # wandb -pyyaml-include==1.4.1 - # via draccus -pyzmq==27.1.0 - # via - # lerobot - # meshcat -qwen-vl-utils==0.0.14 - # via lerobot -reachy2-sdk==1.0.15 - # via lerobot -reachy2-sdk-api==1.0.21 - # via reachy2-sdk -referencing==0.37.0 - # via - # jsonschema - # jsonschema-specifications -regex==2026.2.28 - # via - # diffusers - # transformers -requests==2.32.5 - # via - # datasets - # diffusers - # dm-control - # qwen-vl-utils - # teleop - # wandb -rerun-sdk==0.26.2 - # via lerobot -rhoban-cmeel-jsoncpp==1.9.4.9 - # via placo -rich==14.3.3 - # via typer -robomimic==0.2.0 - # via hf-libero -robosuite==1.4.0 - # via hf-libero -rpds-py==0.30.0 - # via - # jsonschema - # referencing -safetensors==0.7.0 - # via - # accelerate - # diffusers - # lerobot - # peft - # transformers -scikit-image==0.25.2 - # via - # gym-pusht - # lerobot -scipy==1.17.1 - # via - # dm-control - # lerobot - # metaworld - # robosuite - # scikit-image - # torchdiffeq -sentry-sdk==2.54.0 - # via wandb -shapely==2.1.2 - # via gym-pusht -shellingham==1.5.4 - # via typer -six==1.17.0 - # via - # pynput - # python-dateutil - # python-xlib -smmap==5.0.3 - # via gitdb -stack-data==0.6.3 - # via ipython -starlette==0.52.1 - # via fastapi -sympy==1.14.0 - # via torch -teleop==0.1.4 - # via lerobot -tensorboard==2.20.0 - # via robomimic -tensorboard-data-server==0.7.2 - # via tensorboard -tensorboardx==2.6.4 - # via robomimic -termcolor==3.3.0 - # via - # lerobot - # robomimic -thop==0.1.1.post2209072238 - # via hf-libero -tifffile==2026.3.3 - # via scikit-image -tokenizers==0.22.2 - # via transformers -toml==0.10.2 - # via draccus -torch==2.10.0 - # via - # accelerate - # lerobot - # peft - # robomimic - # thop - # torchdiffeq - # torchvision -torchcodec==0.10.0 - # via lerobot -torchdiffeq==0.2.5 - # via lerobot -torchvision==0.25.0 - # via - # lerobot - # robomimic -tornado==6.5.4 - # via meshcat -tqdm==4.67.3 - # via - # datasets - # dm-control - # huggingface-hub - # peft - # robomimic - # transformers -traitlets==5.14.3 - # via - # ipython - # jupyter-core - # matplotlib-inline - # nbformat -transformers==5.3.0 - # via - # hf-libero - # lerobot - # peft -transforms3d==0.4.2 - # via teleop -triton==3.6.0 - # via torch -typer==0.24.1 - # via - # huggingface-hub - # transformers -typing-extensions==4.15.0 - # via - # aiosignal - # anyio - # etils - # faker - # fastapi - # gymnasium - # huggingface-hub - # mypy - # pydantic - # pydantic-core - # referencing - # rerun-sdk - # starlette - # torch - # typing-inspect - # typing-inspection - # wandb -typing-inspect==0.9.0 - # via draccus -typing-inspection==0.4.2 - # via - # fastapi - # pydantic -tzdata==2025.3 - # via pandas -u-msgpack-python==2.8.0 - # via meshcat -urllib3==2.6.3 - # via - # requests - # sentry-sdk -uvicorn[standard]==0.41.0 - # via teleop -uvloop==0.22.1 - # via uvicorn -virtualenv==21.1.0 - # via pre-commit -wandb==0.24.2 - # via - # hf-libero - # lerobot -watchfiles==1.1.1 - # via uvicorn -wcwidth==0.6.0 - # via prompt-toolkit -websocket-client==1.9.0 - # via teleop -websockets==16.0 - # via uvicorn -werkzeug==3.1.6 - # via tensorboard -wrapt==2.1.2 - # via dm-tree -xxhash==3.6.0 - # via datasets -yarl==1.23.0 - # via aiohttp -zipp==3.23.0 - # via - # etils - # importlib-metadata - -# The following packages are considered to be unsafe in a requirements file: -# setuptools diff --git a/requirements.in b/requirements.in deleted file mode 100644 index b39632f71..000000000 --- a/requirements.in +++ /dev/null @@ -1,9 +0,0 @@ -# requirements.in - -# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64). -# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64 - -# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64). -# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux - --e .[all] diff --git a/src/lerobot/annotations/steerable_pipeline/frames.py b/src/lerobot/annotations/steerable_pipeline/frames.py index a6c904673..5a6a5879c 100644 --- a/src/lerobot/annotations/steerable_pipeline/frames.py +++ b/src/lerobot/annotations/steerable_pipeline/frames.py @@ -36,7 +36,7 @@ from typing import Any, Protocol import PIL.Image import torch -from lerobot.configs.video import VideoEncoderConfig +from lerobot.configs import RGBEncoderConfig from lerobot.datasets.video_utils import decode_video_frames, reencode_video from .reader import EpisodeRecord, snap_to_frame @@ -164,7 +164,9 @@ class VideoFrameProvider: # only for video-stored cameras. Image-stored cameras (also in # ``camera_keys``) would KeyError, so restrict the list — and the # default — to video keys. - keys = list(self._meta.video_keys) + # Depth cameras are excluded from the annotation pipeline for now. + depth_keys = set(self._meta.depth_keys) + keys = [key for key in self._meta.video_keys if key not in depth_keys] # Last-resort fallback: if metadata didn't surface any video keys but # the caller explicitly named a camera (``--vlm.camera_key=...``), # trust them — the key is by definition known to exist on the dataset. @@ -276,12 +278,12 @@ class VideoFrameProvider: from_timestamp = float(ep[f"videos/{self.camera_key}/from_timestamp"]) to_timestamp = float(ep[f"videos/{self.camera_key}/to_timestamp"]) src = self.root / self._meta.get_video_file_path(record.episode_index, self.camera_key) - encoder = VideoEncoderConfig(vcodec="h264", pix_fmt="yuv420p", g=None, crf=23, preset="ultrafast") + encoder = RGBEncoderConfig(vcodec="h264", pix_fmt="yuv420p", g=None, crf=23, preset="ultrafast") try: reencode_video( src, out_path, - camera_encoder=encoder, + video_encoder=encoder, overwrite=True, start_time_s=from_timestamp, end_time_s=to_timestamp, diff --git a/src/lerobot/async_inference/helpers.py b/src/lerobot/async_inference/helpers.py index 4931c68c5..54f0ca69f 100644 --- a/src/lerobot/async_inference/helpers.py +++ b/src/lerobot/async_inference/helpers.py @@ -105,8 +105,9 @@ def raw_observation_to_observation( def prepare_image(image: torch.Tensor) -> torch.Tensor: - """Minimal preprocessing to turn int8 images to float32 in [0, 1], and create a memory-contiguous tensor""" - image = image.type(torch.float32) / 255 + """Minimal preprocessing to turn RGB uint8 images to float32 in [0, 1], and create a memory-contiguous tensor""" + if image.dtype == torch.uint8: + image = image.type(torch.float32) / 255 image = image.contiguous() return image diff --git a/src/lerobot/cameras/opencv/camera_opencv.py b/src/lerobot/cameras/opencv/camera_opencv.py index 3e92eaf06..e50d24c01 100644 --- a/src/lerobot/cameras/opencv/camera_opencv.py +++ b/src/lerobot/cameras/opencv/camera_opencv.py @@ -436,17 +436,18 @@ class OpenCVCamera(Camera): Internal loop run by the background thread for asynchronous reading. On each iteration: - 1. Reads a color frame + 1. Reads a color frame (blocking call) 2. Stores result in latest_frame and updates timestamp (thread-safe) 3. Sets new_frame_event to notify listeners Stops on DeviceNotConnectedError, logs other errors and continues. """ - if self.stop_event is None: + stop_event = self.stop_event + if stop_event is None: raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.") failure_count = 0 - while not self.stop_event.is_set(): + while not stop_event.is_set(): try: raw_frame = self._read_from_hardware() processed_frame = self._postprocess_image(raw_frame) @@ -484,6 +485,8 @@ class OpenCVCamera(Camera): if self.thread is not None and self.thread.is_alive(): self.thread.join(timeout=2.0) + if self.thread.is_alive(): + logger.warning(f"{self} read thread did not terminate within timeout.") self.thread = None self.stop_event = None diff --git a/src/lerobot/cameras/realsense/camera_realsense.py b/src/lerobot/cameras/realsense/camera_realsense.py index e156e6d14..29cb1e5e0 100644 --- a/src/lerobot/cameras/realsense/camera_realsense.py +++ b/src/lerobot/cameras/realsense/camera_realsense.py @@ -128,6 +128,7 @@ class RealSenseCamera(Camera): self.fps = config.fps self.color_mode = config.color_mode + self.use_rgb = config.use_rgb self.use_depth = config.use_depth self.warmup_s = config.warmup_s @@ -195,12 +196,15 @@ class RealSenseCamera(Camera): # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise. self.warmup_s = max(self.warmup_s, 1) + warmup_read = self.async_read if self.use_rgb else self.async_read_depth start_time = time.time() while time.time() - start_time < self.warmup_s: - self.async_read(timeout_ms=self.warmup_s * 1000) + warmup_read(timeout_ms=self.warmup_s * 1000) time.sleep(0.1) with self.frame_lock: - if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None: + if (self.use_rgb and self.latest_color_frame is None) or ( + self.use_depth and self.latest_depth_frame is None + ): raise ConnectionError(f"{self} failed to capture frames during warmup.") logger.info(f"{self} connected.") @@ -268,13 +272,13 @@ class RealSenseCamera(Camera): ) if len(found_devices) > 1: - serial_numbers = [dev["serial_number"] for dev in found_devices] + serial_numbers = [dev["id"] for dev in found_devices] raise ValueError( f"Multiple RealSense cameras found with name '{name}'. " f"Please use a unique serial number instead. Found SNs: {serial_numbers}" ) - serial_number = str(found_devices[0]["serial_number"]) + serial_number = str(found_devices[0]["id"]) return serial_number def _configure_rs_pipeline_config(self, rs_config: Any) -> None: @@ -282,15 +286,17 @@ class RealSenseCamera(Camera): rs.config.enable_device(rs_config, self.serial_number) if self.width and self.height and self.fps: - rs_config.enable_stream( - rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps - ) + if self.use_rgb: + rs_config.enable_stream( + rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps + ) if self.use_depth: rs_config.enable_stream( rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps ) else: - rs_config.enable_stream(rs.stream.color) + if self.use_rgb: + rs_config.enable_stream(rs.stream.color) if self.use_depth: rs_config.enable_stream(rs.stream.depth) @@ -298,8 +304,9 @@ class RealSenseCamera(Camera): def _configure_capture_settings(self) -> None: """Sets fps, width, and height from device stream if not already configured. - Uses the color stream profile to update unset attributes. Handles rotation by - swapping width/height when needed. Original capture dimensions are always stored. + Uses the color stream profile (or the depth stream profile when the color + stream is disabled) to update unset attributes. Handles rotation by swapping + width/height when needed. Original capture dimensions are always stored. Raises: DeviceNotConnectedError: If device is not connected. @@ -308,7 +315,8 @@ class RealSenseCamera(Camera): if self.rs_profile is None: raise RuntimeError(f"{self}: rs_profile must be initialized before use.") - stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile() + rs_stream = rs.stream.color if self.use_rgb else rs.stream.depth + stream = self.rs_profile.get_stream(rs_stream).as_video_stream_profile() if self.fps is None: self.fps = stream.fps() @@ -323,6 +331,14 @@ class RealSenseCamera(Camera): self.width, self.height = actual_width, actual_height self.capture_width, self.capture_height = actual_width, actual_height + def _read(self, read_depth: bool = False) -> NDArray[Any]: + """Shared helper for :meth:`read`/:meth:`read_depth`: wait for a fresh color or depth frame.""" + if self.thread is None or not self.thread.is_alive(): + raise RuntimeError(f"{self} read thread is not running.") + + self.new_frame_event.clear() + return self._async_read(timeout_ms=10000, read_depth=read_depth) + @check_if_not_connected def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]: """ @@ -332,8 +348,8 @@ class RealSenseCamera(Camera): from the camera hardware via the RealSense pipeline. Returns: - np.ndarray: The depth map as a NumPy array (height, width) - of type `np.uint16` (raw depth values in millimeters) and rotation. + np.ndarray: The depth map as a NumPy array (height, width, 1) + of type `np.uint16` (raw depth values in millimeters). Raises: DeviceNotConnectedError: If the camera is not connected. @@ -349,20 +365,7 @@ class RealSenseCamera(Camera): f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}." ) - if self.thread is None or not self.thread.is_alive(): - raise RuntimeError(f"{self} read thread is not running.") - - self.new_frame_event.clear() - - _ = self.async_read(timeout_ms=10000) - - with self.frame_lock: - depth_map = self.latest_depth_frame - - if depth_map is None: - raise RuntimeError("No depth frame available. Ensure camera is streaming.") - - return depth_map + return self._read(read_depth=True) def _read_from_hardware(self): if self.rs_pipeline is None: @@ -405,12 +408,10 @@ class RealSenseCamera(Camera): f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions." ) - if self.thread is None or not self.thread.is_alive(): - raise RuntimeError(f"{self} read thread is not running.") + if not self.use_rgb: + raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.") - self.new_frame_event.clear() - - frame = self.async_read(timeout_ms=10000) + frame = self._read() read_duration_ms = (time.perf_counter() - start_time) * 1e3 logger.debug(f"{self} read took: {read_duration_ms:.1f}ms") @@ -465,32 +466,38 @@ class RealSenseCamera(Camera): Internal loop run by the background thread for asynchronous reading. On each iteration: - 1. Reads a color frame with 500ms timeout - 2. Stores result in latest_frame and updates timestamp (thread-safe) + 1. Reads a color/depth frame (blocking call with 10s timeout) + 2. Stores result in latest_color_frame/latest_depth_frame and updates timestamp (thread-safe) 3. Sets new_frame_event to notify listeners Stops on DeviceNotConnectedError, logs other errors and continues. """ - if self.stop_event is None: + stop_event = self.stop_event + if stop_event is None: raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.") failure_count = 0 - while not self.stop_event.is_set(): + while not stop_event.is_set(): try: frame = self._read_from_hardware() - color_frame_raw = frame.get_color_frame() - color_frame = np.asanyarray(color_frame_raw.get_data()) - processed_color_frame = self._postprocess_image(color_frame) + + if self.use_rgb: + color_frame_raw = frame.get_color_frame() + color_frame = np.asanyarray(color_frame_raw.get_data()) + processed_color_frame = self._postprocess_image(color_frame) if self.use_depth: depth_frame_raw = frame.get_depth_frame() depth_frame = np.asanyarray(depth_frame_raw.get_data()) processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True) + if processed_depth_frame.ndim == 2: # (H, W) -> (H, W, 1) + processed_depth_frame = processed_depth_frame[..., np.newaxis] capture_time = time.perf_counter() with self.frame_lock: - self.latest_color_frame = processed_color_frame + if self.use_rgb: + self.latest_color_frame = processed_color_frame if self.use_depth: self.latest_depth_frame = processed_depth_frame self.latest_timestamp = capture_time @@ -522,6 +529,8 @@ class RealSenseCamera(Camera): if self.thread is not None and self.thread.is_alive(): self.thread.join(timeout=2.0) + if self.thread.is_alive(): # pragma: no cover + logger.warning(f"{self} read thread did not terminate within timeout.") self.thread = None self.stop_event = None @@ -532,7 +541,26 @@ class RealSenseCamera(Camera): self.latest_timestamp = None self.new_frame_event.clear() - # NOTE(Steven): Missing implementation for depth for now + def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]: + """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame.""" + if self.thread is None or not self.thread.is_alive(): + raise RuntimeError(f"{self} read thread is not running.") + + if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0): + raise TimeoutError( + f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. " + f"Read thread alive: {self.thread.is_alive()}." + ) + + with self.frame_lock: + frame = self.latest_depth_frame if read_depth else self.latest_color_frame + self.new_frame_event.clear() + + if frame is None: + raise RuntimeError(f"Internal error: Event set but no frame available for {self}.") + + return frame + @check_if_not_connected def async_read(self, timeout_ms: float = 200) -> NDArray[Any]: """ @@ -557,25 +585,31 @@ class RealSenseCamera(Camera): RuntimeError: If the background thread died unexpectedly or another error occurs. """ + if not self.use_rgb: + raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.") + + return self._async_read(timeout_ms=timeout_ms) + + def _read_latest(self, max_age_ms: int, read_depth: bool = False) -> NDArray[Any]: + """Shared helper for :meth:`read_latest`/:meth:`read_latest_depth`: peek the latest buffered frame.""" if self.thread is None or not self.thread.is_alive(): raise RuntimeError(f"{self} read thread is not running.") - if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0): - raise TimeoutError( - f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. " - f"Read thread alive: {self.thread.is_alive()}." - ) - with self.frame_lock: - frame = self.latest_color_frame - self.new_frame_event.clear() + frame = self.latest_depth_frame if read_depth else self.latest_color_frame + timestamp = self.latest_timestamp - if frame is None: - raise RuntimeError(f"Internal error: Event set but no frame available for {self}.") + if frame is None or timestamp is None: + raise RuntimeError(f"{self} has not captured any frames yet.") + + age_ms = (time.perf_counter() - timestamp) * 1e3 + if age_ms > max_age_ms: + raise TimeoutError( + f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)." + ) return frame - # NOTE(Steven): Missing implementation for depth for now @check_if_not_connected def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]: """Return the most recent (color) frame captured immediately (Peeking). @@ -592,24 +626,48 @@ class RealSenseCamera(Camera): DeviceNotConnectedError: If the camera is not connected. RuntimeError: If the camera is connected but has not captured any frames yet. """ + if not self.use_rgb: + raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.") - if self.thread is None or not self.thread.is_alive(): - raise RuntimeError(f"{self} read thread is not running.") + return self._read_latest(max_age_ms=max_age_ms) - with self.frame_lock: - frame = self.latest_color_frame - timestamp = self.latest_timestamp + @check_if_not_connected + def async_read_depth(self, timeout_ms: float = 200) -> NDArray[np.uint16]: + """Read the latest depth frame asynchronously, in millimeters. - if frame is None or timestamp is None: - raise RuntimeError(f"{self} has not captured any frames yet.") + Mirrors :meth:`async_read` but returns the depth stream rather than the + color stream. Output is ``np.uint16`` of shape ``(H, W, 1)``, where each + pixel is the distance from the sensor in millimeters. - age_ms = (time.perf_counter() - timestamp) * 1e3 - if age_ms > max_age_ms: - raise TimeoutError( - f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)." - ) + Raises: + DeviceNotConnectedError: If the camera is not connected. + RuntimeError: If ``use_depth`` is ``False`` for this camera, or if + the background read thread is not running. + TimeoutError: If no frame becomes available within ``timeout_ms``. + """ + if not self.use_depth: + raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.") - return frame + return self._async_read(timeout_ms=timeout_ms, read_depth=True) + + @check_if_not_connected + def read_latest_depth(self, max_age_ms: int = 500) -> NDArray[Any]: + """Return the most recent depth frame in millimeters (peeking). + + Non-blocking counterpart of :meth:`read_latest` for the depth stream. + Output is ``np.uint16`` of shape ``(H, W, 1)``, where each pixel is the + distance from the sensor in millimeters. + + Raises: + DeviceNotConnectedError: If the camera is not connected. + RuntimeError: If ``use_depth`` is ``False`` for this camera, or if + no depth frame has been captured yet. + TimeoutError: If the latest depth frame is older than ``max_age_ms``. + """ + if not self.use_depth: + raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.") + + return self._read_latest(max_age_ms=max_age_ms, read_depth=True) def disconnect(self) -> None: """ diff --git a/src/lerobot/cameras/realsense/configuration_realsense.py b/src/lerobot/cameras/realsense/configuration_realsense.py index 71b083b00..018675195 100644 --- a/src/lerobot/cameras/realsense/configuration_realsense.py +++ b/src/lerobot/cameras/realsense/configuration_realsense.py @@ -42,12 +42,14 @@ class RealSenseCameraConfig(CameraConfig): height: Requested frame height in pixels for the color stream. serial_number_or_name: Unique serial number or human-readable name to identify the camera. color_mode: Color mode for image output (RGB or BGR). Defaults to RGB. + use_rgb: Whether to enable the color stream. Defaults to True. use_depth: Whether to enable depth stream. Defaults to False. rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation. warmup_s: Time reading frames before returning from connect (in seconds) Note: - Either name or serial_number must be specified. + - At least one of `use_rgb` or `use_depth` must be enabled. - Depth stream configuration (if enabled) will use the same FPS as the color stream. - The actual resolution and FPS may be adjusted by the camera to the nearest supported mode. - For `fps`, `width` and `height`, either all of them need to be set, or none of them. @@ -55,6 +57,7 @@ class RealSenseCameraConfig(CameraConfig): serial_number_or_name: str color_mode: ColorMode = ColorMode.RGB + use_rgb: bool = True use_depth: bool = False rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION warmup_s: int = 1 @@ -63,6 +66,9 @@ class RealSenseCameraConfig(CameraConfig): self.color_mode = ColorMode(self.color_mode) self.rotation = Cv2Rotation(self.rotation) + if not self.use_rgb and not self.use_depth: + raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.") + values = (self.fps, self.width, self.height) if any(v is not None for v in values) and any(v is None for v in values): raise ValueError( diff --git a/src/lerobot/cameras/zmq/camera_zmq.py b/src/lerobot/cameras/zmq/camera_zmq.py index 1b0be5de6..cd32a117b 100644 --- a/src/lerobot/cameras/zmq/camera_zmq.py +++ b/src/lerobot/cameras/zmq/camera_zmq.py @@ -246,11 +246,12 @@ class ZMQCamera(Camera): """ Internal loop run by the background thread for asynchronous reading. """ - if self.stop_event is None: + stop_event = self.stop_event + if stop_event is None: raise RuntimeError(f"{self}: stop_event is not initialized.") failure_count = 0 - while not self.stop_event.is_set(): + while not stop_event.is_set(): try: frame = self._read_from_hardware() capture_time = time.perf_counter() @@ -292,6 +293,8 @@ class ZMQCamera(Camera): if self.thread is not None and self.thread.is_alive(): self.thread.join(timeout=2.0) + if self.thread.is_alive(): + logger.warning(f"{self} read thread did not terminate within timeout.") self.thread = None self.stop_event = None diff --git a/src/lerobot/common/control_utils.py b/src/lerobot/common/control_utils.py index ddaf77d26..e3130643d 100644 --- a/src/lerobot/common/control_utils.py +++ b/src/lerobot/common/control_utils.py @@ -17,12 +17,9 @@ from __future__ import annotations ######################################################################################## # Utilities ######################################################################################## -import logging import time -import traceback from contextlib import nullcontext from copy import copy -from functools import cache from typing import TYPE_CHECKING, Any import numpy as np @@ -43,34 +40,6 @@ from lerobot.robots import Robot from lerobot.types import PolicyAction -@cache -def is_headless(): - """ - Detects if the Python script is running in a headless environment (e.g., without a display). - - This function attempts to import `pynput`, a library that requires a graphical environment. - If the import fails, it assumes the environment is headless. The result is cached to avoid - re-running the check. - - Returns: - True if the environment is determined to be headless, False otherwise. - """ - try: - import pynput # noqa - - return False - except Exception: - print( - "Error trying to import pynput. Switching to headless mode. " - "As a result, the video stream from the cameras won't be shown, " - "and you won't be able to change the control flow with keyboards. " - "For more info, see traceback below.\n" - ) - traceback.print_exc() - print() - return True - - def predict_action( observation: dict[str, np.ndarray], policy: PreTrainedPolicy, @@ -122,59 +91,6 @@ def predict_action( return action -def init_keyboard_listener(): - """ - Initializes a non-blocking keyboard listener for real-time user interaction. - - This function sets up a listener for specific keys (right arrow, left arrow, escape) to control - the program flow during execution, such as stopping recording or exiting loops. It gracefully - handles headless environments where keyboard listening is not possible. - - Returns: - A tuple containing: - - The `pynput.keyboard.Listener` instance, or `None` if in a headless environment. - - A dictionary of event flags (e.g., `exit_early`) that are set by key presses. - """ - # Allow to exit early while recording an episode or resetting the environment, - # by tapping the right arrow key '->'. This might require a sudo permission - # to allow your terminal to monitor keyboard events. - events = {} - events["exit_early"] = False - events["rerecord_episode"] = False - events["stop_recording"] = False - - if is_headless(): - logging.warning( - "Headless environment detected. On-screen cameras display and keyboard inputs will not be available." - ) - listener = None - return listener, events - - # Only import pynput if not in a headless environment - from pynput import keyboard - - def on_press(key): - try: - if key == keyboard.Key.right: - print("Right arrow key pressed. Exiting loop...") - events["exit_early"] = True - elif key == keyboard.Key.left: - print("Left arrow key pressed. Exiting loop and rerecord the last episode...") - events["rerecord_episode"] = True - events["exit_early"] = True - elif key == keyboard.Key.esc: - print("Escape key pressed. Stopping data recording...") - events["stop_recording"] = True - events["exit_early"] = True - except Exception as e: - print(f"Error handling key press: {e}") - - listener = keyboard.Listener(on_press=on_press) - listener.start() - - return listener, events - - def sanity_check_dataset_name(repo_id, policy_cfg): """ Validates the dataset repository name against the presence of a policy configuration. diff --git a/src/lerobot/common/train_utils.py b/src/lerobot/common/train_utils.py index 2d23b4003..b26196f14 100644 --- a/src/lerobot/common/train_utils.py +++ b/src/lerobot/common/train_utils.py @@ -15,12 +15,14 @@ # limitations under the License. from pathlib import Path +from huggingface_hub import HfApi, snapshot_download from torch.optim import Optimizer from torch.optim.lr_scheduler import LRScheduler from lerobot.configs.train import TrainPipelineConfig from lerobot.optim import ( load_optimizer_state, + load_optimizer_state_dict, load_scheduler_state, save_optimizer_state, save_scheduler_state, @@ -34,6 +36,7 @@ from lerobot.utils.constants import ( TRAINING_STATE_DIR, TRAINING_STEP, ) +from lerobot.utils.hub import find_latest_hub_checkpoint from lerobot.utils.io_utils import load_json, write_json from lerobot.utils.random_utils import load_rng_state, save_rng_state @@ -98,6 +101,8 @@ def save_checkpoint( postprocessor: PolicyProcessorPipeline | None = None, num_processes: int | None = None, batch_size: int | None = None, + model_state_dict: dict | None = None, + optim_state_dict: dict | None = None, ) -> None: """This function creates the following directory structure: @@ -127,9 +132,18 @@ def save_checkpoint( resume. Defaults to None (not recorded). batch_size (int | None, optional): Per-process batch size to record for sample-exact resume. Defaults to None (not recorded). + model_state_dict: Pre-gathered full (unsharded) model state dict. Required under FSDP, + where `policy.state_dict()` would return sharded tensors; the caller gathers it via a + cross-rank collective and passes it here so rank 0 can write it directly. It holds + FSDP's fp32 master weights and is saved as-is (the loader casts to the policy dtype on + read). When None (DDP / single-GPU), the model is saved the normal way. Defaults to None. + optim_state_dict: Pre-gathered full (unsharded) optimizer state dict. Required under FSDP + (gathered alongside `model_state_dict` via `gather_fsdp_state_dicts`); saved in the same + safetensors format as the single-GPU path. When None, `optimizer.state_dict()` is used. + Defaults to None. """ pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR - policy.save_pretrained(pretrained_dir) + policy.save_pretrained(pretrained_dir, state_dict=model_state_dict) cfg.save_pretrained(pretrained_dir) if cfg.peft is not None: # When using PEFT, policy.save_pretrained will only write the adapter weights + config, not the @@ -140,7 +154,13 @@ def save_checkpoint( if postprocessor is not None: postprocessor.save_pretrained(pretrained_dir) save_training_state( - checkpoint_dir, step, optimizer, scheduler, num_processes=num_processes, batch_size=batch_size + checkpoint_dir, + step, + optimizer, + scheduler, + num_processes=num_processes, + batch_size=batch_size, + optim_state_dict=optim_state_dict, ) @@ -151,6 +171,7 @@ def save_training_state( scheduler: LRScheduler | None = None, num_processes: int | None = None, batch_size: int | None = None, + optim_state_dict: dict | None = None, ) -> None: """ Saves the training step, optimizer state, scheduler state, and rng state. @@ -164,19 +185,21 @@ def save_training_state( Defaults to None. num_processes (int | None, optional): Distributed world size to record. Defaults to None. batch_size (int | None, optional): Per-process batch size to record. Defaults to None. + optim_state_dict: Pre-gathered full optimizer state dict (for FSDP). Saved instead of + `optimizer.state_dict()` when provided. Defaults to None. """ save_dir = checkpoint_dir / TRAINING_STATE_DIR save_dir.mkdir(parents=True, exist_ok=True) save_training_step(train_step, save_dir, num_processes=num_processes, batch_size=batch_size) save_rng_state(save_dir) if optimizer is not None: - save_optimizer_state(optimizer, save_dir) + save_optimizer_state(optimizer, save_dir, optim_state_dict=optim_state_dict) if scheduler is not None: save_scheduler_state(scheduler, save_dir) def load_training_state( - checkpoint_dir: Path, optimizer: Optimizer, scheduler: LRScheduler | None + checkpoint_dir: Path, optimizer: Optimizer, scheduler: LRScheduler | None, load_optimizer: bool = True ) -> tuple[int, Optimizer, LRScheduler | None]: """ Loads the training step, optimizer state, scheduler state, and rng state. @@ -186,6 +209,10 @@ def load_training_state( checkpoint_dir (Path): The checkpoint directory. Should contain a 'training_state' dir. optimizer (Optimizer): The optimizer to load the state_dict to. scheduler (LRScheduler | None): The scheduler to load the state_dict to (can be None). + load_optimizer (bool, optional): Whether to load the optimizer state from disk. Defaults to + True. Set to False under FSDP, where the sharded optimizer state must be loaded after + `accelerator.prepare()` via `load_fsdp_optimizer_state` (the optimizer is returned + untouched here). Raises: NotADirectoryError: If 'checkpoint_dir' doesn't contain a 'training_state' dir @@ -200,8 +227,119 @@ def load_training_state( load_rng_state(training_state_dir) step = load_training_step(training_state_dir) - optimizer = load_optimizer_state(optimizer, training_state_dir) + if load_optimizer: + optimizer = load_optimizer_state(optimizer, training_state_dir) if scheduler is not None: scheduler = load_scheduler_state(scheduler, training_state_dir) return step, optimizer, scheduler + + +def gather_fsdp_state_dicts(model, optimizer) -> tuple[dict, dict]: + """Gather the full (unsharded) model and optimizer state dicts under FSDP. + + `model.state_dict()` and `FSDP.optim_state_dict(...)` are cross-rank collectives, so this must be + called on *every* rank with the prepared (FSDP-wrapped) `model` and `optimizer`. With + `rank0_only=True` and `offload_to_cpu=True`, every rank runs the all-gather but only rank 0 + materializes the full dicts (the others get empty dicts) and they are kept on CPU to bound GPU + memory. The returned optimizer state dict is keyed by parameter FQNs and is world-size + independent; `load_fsdp_optimizer_state` reshards it on resume. + + Returns: + (model_state_dict, optim_state_dict): full dicts on rank 0, empty dicts on other ranks. + """ + from torch.distributed.fsdp import ( + FullOptimStateDictConfig, + FullStateDictConfig, + FullyShardedDataParallel as FSDP, # noqa F401 + StateDictType, + ) + + state_cfg = FullStateDictConfig(offload_to_cpu=True, rank0_only=True) + optim_cfg = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True) + with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg): + model_state_dict = model.state_dict() + optim_state_dict = FSDP.optim_state_dict(model, optimizer) + return model_state_dict, optim_state_dict + + +def load_fsdp_optimizer_state(model, optimizer, checkpoint_dir: Path) -> None: + """Load the FSDP optimizer state (saved as safetensors) and reshard it into the optimizer. + + This is a cross-rank collective and must be called on every rank *after* `accelerator.prepare()` + with the prepared (FSDP-wrapped) `model` and `optimizer`. The saved state is the full, + world-size-independent optimizer state (keyed by parameter FQNs); `FSDP.optim_state_dict_to_load` + reshards it to the current FSDP topology, so resume on a different number of GPUs works. + """ + from torch.distributed.fsdp import ( + FullOptimStateDictConfig, + FullStateDictConfig, + FullyShardedDataParallel as FSDP, # noqa F401 + StateDictType, + ) + + # Every rank reads the same full state from the (shared) checkpoint dir, so rank0_only=False. + full_osd = load_optimizer_state_dict(checkpoint_dir / TRAINING_STATE_DIR) + state_cfg = FullStateDictConfig(rank0_only=False) + optim_cfg = FullOptimStateDictConfig(rank0_only=False) + with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg): + sharded_osd = FSDP.optim_state_dict_to_load(model=model, optim=optimizer, optim_state_dict=full_osd) + optimizer.load_state_dict(sharded_osd) + + +def push_checkpoint_to_hub( + checkpoint_dir: Path, + repo_id: str, + *, + private: bool | None = None, +) -> None: + """Upload a saved checkpoint directory to the Hub under checkpoints//. + + Called once per save step when save_checkpoint_to_hub is enabled, so a + timed-out or crashed run still leaves recoverable checkpoints on the Hub. + The model repo is created idempotently, and the commit is tagged with the + checkpoint step so a checkpoint can be recovered with + --policy.pretrained_revision= instead of a commit sha. + """ + api = HfApi() + api.create_repo(repo_id=repo_id, repo_type="model", private=private, exist_ok=True) + commit = api.upload_folder( + folder_path=str(checkpoint_dir), + repo_id=repo_id, + repo_type="model", + path_in_repo=f"checkpoints/{checkpoint_dir.name}", + commit_message=f"checkpoint {checkpoint_dir.name}", + ) + api.create_tag( + repo_id=repo_id, + tag=checkpoint_dir.name, + revision=commit.oid, + repo_type="model", + exist_ok=True, + ) + + +def resolve_resume_checkpoint(repo_id: str, output_dir: Path) -> Path: + """Download the latest checkpoint of a Hub training repo into a local run dir. + + The symmetric counterpart to `push_checkpoint_to_hub`: given a model repo holding + `checkpoints//{pretrained_model,training_state}` subtrees, download the highest-numbered step + into `output_dir/checkpoints//`, recreate the local `last` symlink, and return that local + checkpoint dir. Used to resume training from the Hub on a machine (or HF Jobs pod) that does not + have the original local run dir. + """ + latest = find_latest_hub_checkpoint(repo_id) + if latest is None: + raise FileNotFoundError( + f"No checkpoint found in '{repo_id}' under '{CHECKPOINTS_DIR}/'. " + "Was the run trained with --save_checkpoint_to_hub?" + ) + snapshot_download( + repo_id=repo_id, + repo_type="model", + allow_patterns=f"{latest}/*", + local_dir=str(output_dir), + ) + checkpoint_dir = output_dir / latest + update_last_checkpoint(checkpoint_dir) + return checkpoint_dir diff --git a/src/lerobot/common/wandb_utils.py b/src/lerobot/common/wandb_utils.py index b782cd751..c229b5eaa 100644 --- a/src/lerobot/common/wandb_utils.py +++ b/src/lerobot/common/wandb_utils.py @@ -180,24 +180,26 @@ class WandBLogger: self._wandb_custom_step_key.add(new_custom_key) self._wandb.define_metric(new_custom_key, hidden=True) + batch_data = {} for k, v in d.items(): + # Skip the custom step key here, it's added to the batch below. + if custom_step_key is not None and k == custom_step_key: + continue + if not isinstance(v, (int | float | str)): logging.warning( f'WandB logging of key "{k}" was ignored as its type "{type(v)}" is not handled by this wrapper.' ) continue - # Do not log the custom step key itself. - if self._wandb_custom_step_key is not None and k in self._wandb_custom_step_key: - continue + batch_data[f"{mode}/{k}"] = v + if batch_data: if custom_step_key is not None: - value_custom_step = d[custom_step_key] - data = {f"{mode}/{k}": v, f"{mode}/{custom_step_key}": value_custom_step} - self._wandb.log(data) - continue - - self._wandb.log(data={f"{mode}/{k}": v}, step=step) + batch_data[f"{mode}/{custom_step_key}"] = d[custom_step_key] + self._wandb.log(batch_data) + else: + self._wandb.log(data=batch_data, step=step) def log_video(self, video_path: str, step: int, mode: str = "train"): if mode not in {"train", "eval"}: diff --git a/src/lerobot/configs/__init__.py b/src/lerobot/configs/__init__.py index be4491811..c32e3368b 100644 --- a/src/lerobot/configs/__init__.py +++ b/src/lerobot/configs/__init__.py @@ -22,7 +22,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig`` """ from .dataset import DatasetRecordConfig -from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig +from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig from .policies import PreTrainedConfig from .recipe import MessageTurn, TrainingRecipe, load_recipe from .types import ( @@ -33,10 +33,18 @@ from .types import ( RTCAttentionSchedule, ) from .video import ( + DEFAULT_DEPTH_UNIT, + DEPTH_METER_UNIT, + DEPTH_MILLIMETER_UNIT, VALID_VIDEO_CODECS, VIDEO_ENCODER_INFO_KEYS, + DepthEncoderConfig, + RGBEncoderConfig, VideoEncoderConfig, - camera_encoder_defaults, + depth_encoder_defaults, + encoder_config_from_video_info, + infer_depth_unit, + rgb_encoder_defaults, ) __all__ = [ @@ -50,6 +58,7 @@ __all__ = [ "DatasetRecordConfig", "DatasetConfig", "EvalConfig", + "JobConfig", "MessageTurn", "PeftConfig", "PreTrainedConfig", @@ -57,9 +66,18 @@ __all__ = [ "WandBConfig", "load_recipe", "VideoEncoderConfig", + "RGBEncoderConfig", + "DepthEncoderConfig", # Defaults - "camera_encoder_defaults", + "rgb_encoder_defaults", + "depth_encoder_defaults", + # Factories + "encoder_config_from_video_info", + "infer_depth_unit", # Constants + "DEFAULT_DEPTH_UNIT", + "DEPTH_METER_UNIT", + "DEPTH_MILLIMETER_UNIT", "VALID_VIDEO_CODECS", "VIDEO_ENCODER_INFO_KEYS", ] diff --git a/src/lerobot/configs/dataset.py b/src/lerobot/configs/dataset.py index c40c0fae2..7d30ca038 100644 --- a/src/lerobot/configs/dataset.py +++ b/src/lerobot/configs/dataset.py @@ -18,7 +18,7 @@ from dataclasses import dataclass, field from datetime import datetime from pathlib import Path -from .video import VideoEncoderConfig, camera_encoder_defaults +from .video import DepthEncoderConfig, RGBEncoderConfig, depth_encoder_defaults, rgb_encoder_defaults @dataclass @@ -58,8 +58,10 @@ class DatasetRecordConfig: # Set to 1 for immediate encoding (default behavior), or higher for batched encoding video_encoding_batch_size: int = 1 # Video encoder settings for camera MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys, - # e.g. ``--dataset.camera_encoder.vcodec=h264`` (see ``VideoEncoderConfig``). - camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults) + # e.g. ``--dataset.rgb_encoder.vcodec=h264`` (see ``RGBEncoderConfig``). + rgb_encoder: RGBEncoderConfig = field(default_factory=rgb_encoder_defaults) + # Video encoder settings for depth-map MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys. + depth_encoder: DepthEncoderConfig = field(default_factory=depth_encoder_defaults) # Enable streaming video encoding: encode frames in real-time during capture instead # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding streaming_encoding: bool = False diff --git a/src/lerobot/configs/default.py b/src/lerobot/configs/default.py index b809e71d9..38991a665 100644 --- a/src/lerobot/configs/default.py +++ b/src/lerobot/configs/default.py @@ -19,6 +19,8 @@ from dataclasses import dataclass, field from lerobot.transforms import ImageTransformsConfig from lerobot.utils.import_utils import get_safe_default_video_backend +from .video import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT + @dataclass class DatasetConfig: @@ -35,12 +37,23 @@ class DatasetConfig: revision: str | None = None use_imagenet_stats: bool = True video_backend: str = field(default_factory=get_safe_default_video_backend) - # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0). + # When True, RGB video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0). # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion. return_uint8: bool = False + # Physical unit depth maps are dequantized to at load time: "mm" (millimeters) or "m" (metres). + # Has no effect on datasets without depth cameras. + depth_output_unit: str = DEFAULT_DEPTH_UNIT streaming: bool = False + # Fraction of episodes held out per task for offline evaluation (0.0 = disabled). + eval_split: float = 0.0 def __post_init__(self) -> None: + if self.depth_output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT): + raise ValueError( + f"depth_output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {self.depth_output_unit!r}" + ) + if not (0.0 <= self.eval_split < 1.0): + raise ValueError(f"eval_split must be in [0.0, 1.0), got {self.eval_split}") if self.episodes is not None: if any(ep < 0 for ep in self.episodes): raise ValueError( @@ -73,8 +86,17 @@ class EvalConfig: # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing). # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1. use_async_envs: bool = True + # Whether to record eval rollouts as a LeRobot dataset on disk. + recording: bool = False + # If set, push recorded eval datasets to the Hub under this repo id (one repo per task, + # suffixed by task and env index). Requires recording=true. + recording_repo_id: str | None = None + # Whether the pushed recording repositories should be private. + recording_private: bool = False def __post_init__(self) -> None: + if self.recording_repo_id is not None and not self.recording: + raise ValueError("eval.recording_repo_id requires eval.recording=true.") if self.batch_size == 0: self.batch_size = self._auto_batch_size() if self.batch_size > self.n_episodes: @@ -123,3 +145,35 @@ class PeftConfig: # If None, the PEFT library defaults to alpha=8, which may dampen high-rank adapters. # Common values are r (alpha == rank) or 2*r. lora_alpha: int | None = None + + +@dataclass +class JobConfig: + # Where training runs. None (omitted) or "local" runs on this machine. + # Any other value is an HF Jobs flavor and submits the run to HF Jobs. + # List available flavors + pricing with `hf jobs hardware` command. + target: str | None = None + # Runtime image for the remote job (ignored for local runs). + image: str = "huggingface/lerobot-gpu:latest" + # Max wall-clock for the remote job as an HF Jobs duration string (e.g. "2h"). + # Defaults to "2d": We pass an explicit, generous cap instead. Set a smaller + # value to fail fast, or a larger one for long runs. + timeout: str | None = "2d" + # Submit and exit instead of streaming the job logs in the foreground. + detach: bool = False + # Extra tags attached to the HF job and to any dataset this run pushes to the + # Hub. A "lerobot" tag is always added; e.g. --job.tags '["lelab"]' adds more. + tags: list[str] = field(default_factory=list) + + # Two entry points to the same predicate: the staticmethod tests a raw target string + # straight from argv (before any JobConfig exists, to decide dispatch early), while the + # property is the ergonomic accessor for code that already holds a config instance. + @staticmethod + def is_remote_target(target: str | None) -> bool: + """True when `target` names an HF Jobs flavor rather than a local run.""" + return target not in (None, "local") + + @property + def is_remote(self) -> bool: + """True when training should run on HF Jobs rather than this machine.""" + return self.is_remote_target(self.target) diff --git a/src/lerobot/configs/policies.py b/src/lerobot/configs/policies.py index 91701af6d..b0f003519 100644 --- a/src/lerobot/configs/policies.py +++ b/src/lerobot/configs/policies.py @@ -79,6 +79,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC): # type: igno # Either the repo ID of a model hosted on the Hub or a path to a directory containing weights # saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch. pretrained_path: Path | None = None + # Optional Hub revision (commit hash, branch, or tag) to pin the pretrained model version. + pretrained_revision: str | None = None def __post_init__(self) -> None: if not self.device or not is_torch_device_available(self.device): diff --git a/src/lerobot/configs/rewards.py b/src/lerobot/configs/rewards.py index 7e99e7f71..92490bc9f 100644 --- a/src/lerobot/configs/rewards.py +++ b/src/lerobot/configs/rewards.py @@ -56,6 +56,8 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC): device: str | None = None pretrained_path: str | None = None + # Optional Hub revision (commit hash, branch, or tag) to pin the pretrained reward model version. + pretrained_revision: str | None = None push_to_hub: bool = False repo_id: str | None = None diff --git a/src/lerobot/configs/train.py b/src/lerobot/configs/train.py index bac1a946b..e3d354691 100644 --- a/src/lerobot/configs/train.py +++ b/src/lerobot/configs/train.py @@ -26,11 +26,12 @@ from huggingface_hub.errors import HfHubHTTPError from lerobot import envs from lerobot.optim import LRSchedulerConfig, OptimizerConfig -from lerobot.utils.hub import HubMixin +from lerobot.utils.constants import PRETRAINED_MODEL_DIR +from lerobot.utils.hub import HubMixin, find_latest_hub_checkpoint from lerobot.utils.sample_weighting import SampleWeightingConfig from . import parser -from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig +from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig from .policies import PreTrainedConfig from .rewards import RewardModelConfig @@ -83,10 +84,11 @@ class TrainPipelineConfig(HubMixin): # with the same value for `dir` its contents will be overwritten unless you set `resume` to true. output_dir: Path | None = None job_name: str | None = None - # Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure - # `dir` is the directory of an existing run with at least one checkpoint in it. - # Note that when resuming a run, the default behavior is to use the configuration from the checkpoint, - # regardless of what's provided with the training command at the time of resumption. + # Set `resume` to true to resume a previous run. Pass `--config_path` pointing at either a local + # checkpoint's train_config.json or a Hub repo id holding `checkpoints//` subtrees (the + # latest checkpoint is downloaded and resumed from). Note that when resuming, the default behavior + # is to use the configuration from the checkpoint, regardless of what's provided with the training + # command at the time of resumption (CLI `--*` flags still override). resume: bool = False # `seed` is used for training (eg: model initialization, dataset shuffling) # AND for the evaluation environments. @@ -100,8 +102,13 @@ class TrainPipelineConfig(HubMixin): prefetch_factor: int = 4 persistent_workers: bool = True steps: int = 100_000 - eval_freq: int = 20_000 + # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled). + env_eval_freq: int = 20_000 log_freq: int = 200 + # Compute eval loss on held-out episodes every N steps (0 = disabled). Requires eval_split > 0. + eval_steps: int = 0 + # Cap on total eval samples, split uniformly across tasks (0 = use all held-out data). + max_eval_samples: int = 0 tolerance_s: float = 1e-4 save_checkpoint: bool = True # Checkpoint is saved every `save_freq` training iterations and after the last training step. @@ -113,6 +120,13 @@ class TrainPipelineConfig(HubMixin): wandb: WandBConfig = field(default_factory=WandBConfig) peft: PeftConfig | None = None + # Where to run training (local default, or an HF Jobs flavor). See JobConfig. + job: JobConfig = field(default_factory=JobConfig) + # Push each saved checkpoint to the Hub (policy.repo_id) as it is written, not + # just the final model (useful to monitor progress mid-run). Optional; the + # final model is pushed regardless. Works the same locally and remotely. + save_checkpoint_to_hub: bool = False + # Sample weighting configuration (e.g., for RA-BC training) sample_weighting: SampleWeightingConfig | None = None @@ -132,10 +146,17 @@ class TrainPipelineConfig(HubMixin): return self.reward_model # type: ignore[return-value] return self.policy # type: ignore[return-value] - def validate(self) -> None: - # HACK: We parse again the cli args here to get the pretrained paths if there was some. - policy_path = parser.get_path_arg("policy") + def _resolve_pretrained_from_cli(self) -> None: + """Resolve the pretrained source passed on the CLI into a loaded config. + + The pretrained paths (`--policy.path`, `--reward_model.path`) and + `--config_path` are only recoverable by re-reading the CLI args: draccus + has already consumed them by the time `validate()` runs, so they are not + reflected on `self`. Exactly one source applies, in priority order: + reward-model path, policy path, then resume. + """ reward_model_path = parser.get_path_arg("reward_model") + policy_path = parser.get_path_arg("policy") if reward_model_path: cli_overrides = parser.get_cli_overrides("reward_model") @@ -144,31 +165,54 @@ class TrainPipelineConfig(HubMixin): ) self.reward_model.pretrained_path = str(Path(reward_model_path)) elif policy_path: - yaml_overrides = parser.get_yaml_overrides("policy") - cli_overrides = parser.get_cli_overrides("policy") or [] - self.policy = PreTrainedConfig.from_pretrained( - policy_path, cli_overrides=yaml_overrides + cli_overrides - ) + overrides = parser.get_yaml_overrides("policy") + (parser.get_cli_overrides("policy") or []) + self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=overrides) self.policy.pretrained_path = Path(policy_path) elif self.resume: - config_path = parser.parse_arg("config_path") - if not config_path: - raise ValueError( - f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}" - ) + self._resolve_resume_checkpoint() - if not Path(config_path).resolve().exists(): - raise NotADirectoryError( - f"{config_path=} is expected to be a local path. " - "Resuming from the hub is not supported for now." - ) + def _resolve_resume_checkpoint(self) -> None: + """Point the trainable config at the checkpoint named by `--config_path`. + `config_path` is either a local path (to a checkpoint's train_config.json or its + pretrained_model/ dir) or a Hub repo id. For a Hub repo, the latest checkpoint is downloaded + into a fresh local run dir and resumed from there. The download is skipped when dispatching to + an HF Job (`job.is_remote`): the pod performs it when it runs the resume locally, and + `submit_to_hf` resolves the source repo for the remote command. + """ + config_path = parser.parse_arg("config_path") + if not config_path: + raise ValueError( + f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}" + ) + + if Path(config_path).resolve().exists(): policy_dir = Path(config_path).parent - if self.policy is not None: - self.policy.pretrained_path = policy_dir - if self.reward_model is not None: - self.reward_model.pretrained_path = str(policy_dir) self.checkpoint_path = policy_dir.parent + elif self.job.is_remote: + return + else: + from lerobot.common.train_utils import resolve_resume_checkpoint + + # `self.output_dir` was loaded from the checkpoint's config and points at the original + # run's (now-absent) local dir. Resume into a fresh local dir instead, unless the user + # passed --output_dir explicitly. + cli_output_dir = parser.parse_arg("output_dir") + if cli_output_dir: + self.output_dir = Path(cli_output_dir) + else: + now = dt.datetime.now() + self.output_dir = Path("outputs/train") / f"{now:%Y-%m-%d}/{now:%H-%M-%S}_resume" + self.checkpoint_path = resolve_resume_checkpoint(config_path, self.output_dir) + policy_dir = self.checkpoint_path / PRETRAINED_MODEL_DIR + + if self.policy is not None: + self.policy.pretrained_path = policy_dir + if self.reward_model is not None: + self.reward_model.pretrained_path = str(policy_dir) + + def validate(self) -> None: + self._resolve_pretrained_from_cli() if self.policy is None and self.reward_model is None: raise ValueError( @@ -208,9 +252,22 @@ class TrainPipelineConfig(HubMixin): self.optimizer = active_cfg.get_optimizer_preset() self.scheduler = active_cfg.get_scheduler_preset() - if hasattr(active_cfg, "push_to_hub") and active_cfg.push_to_hub and not active_cfg.repo_id: + if self.eval_steps > 0 and self.dataset.eval_split == 0.0: + raise ValueError("eval_steps > 0 requires dataset.eval_split > 0.0 to hold out eval data.") + + # Remote runs auto-generate the repo_id in submit_to_hf (the policy may only be + # resolved here, from --policy.path), so don't demand it up front for them. + if ( + hasattr(active_cfg, "push_to_hub") + and active_cfg.push_to_hub + and not active_cfg.repo_id + and not self.job.is_remote + ): raise ValueError("'repo_id' argument missing. Please specify it to push the model to the hub.") + if self.save_checkpoint_to_hub and not (self.policy is not None and self.policy.repo_id): + raise ValueError("save_checkpoint_to_hub requires --policy.repo_id.") + @classmethod def __get_path_fields__(cls) -> list[str]: """Keys for draccus pretrained-path loading.""" @@ -247,22 +304,30 @@ class TrainPipelineConfig(HubMixin): elif Path(model_id).is_file(): config_file = model_id else: + dl_kwargs = { + "repo_id": model_id, + "revision": revision, + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "resume_download": resume_download, + "token": token, + "local_files_only": local_files_only, + } try: - config_file = hf_hub_download( - repo_id=model_id, - filename=TRAIN_CONFIG_NAME, - revision=revision, - cache_dir=cache_dir, - force_download=force_download, - proxies=proxies, - resume_download=resume_download, - token=token, - local_files_only=local_files_only, - ) + config_file = hf_hub_download(filename=TRAIN_CONFIG_NAME, **dl_kwargs) except HfHubHTTPError as e: - raise FileNotFoundError( - f"{TRAIN_CONFIG_NAME} not found on the HuggingFace Hub in {model_id}" - ) from e + # No root train_config.json: this is a repo of periodic checkpoints from an + # interrupted run. Fall back to the latest checkpoint's config so the run can be + # resumed straight from the repo with `--config_path=`. + latest = find_latest_hub_checkpoint(model_id, token=token, revision=revision) + if latest is None: + raise FileNotFoundError( + f"{TRAIN_CONFIG_NAME} not found on the HuggingFace Hub in {model_id}" + ) from e + config_file = hf_hub_download( + filename=f"{latest}/{PRETRAINED_MODEL_DIR}/{TRAIN_CONFIG_NAME}", **dl_kwargs + ) cli_args = kwargs.pop("cli_args", []) # Legacy RA-BC migration only applies to framework-saved checkpoints (always JSON). diff --git a/src/lerobot/configs/video.py b/src/lerobot/configs/video.py index bf2471453..4b956f30e 100644 --- a/src/lerobot/configs/video.py +++ b/src/lerobot/configs/video.py @@ -20,7 +20,9 @@ from __future__ import annotations import logging from dataclasses import dataclass, field -from typing import Any +from typing import Any, ClassVar, Self + +import numpy as np from lerobot.utils.import_utils import require_package @@ -36,11 +38,12 @@ HW_VIDEO_CODECS = [ "h264_vaapi", # Linux Intel/AMD "h264_qsv", # Intel Quick Sync ] -VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS}) +VALID_VIDEO_CODECS: frozenset[str] = frozenset( + {"h264", "hevc", "libsvtav1", "libaom-av1", "auto", *HW_VIDEO_CODECS} +) # Aliases for legacy video codec names. VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"} - LIBSVTAV1_DEFAULT_PRESET: int = 12 # Keys persisted under ``features[*]["info"]`` as ``video.`` (from :class:`VideoEncoderConfig`). @@ -52,40 +55,54 @@ VIDEO_ENCODER_INFO_KEYS: frozenset[str] = frozenset( f"video.{name}" for name in VIDEO_ENCODER_INFO_FIELD_NAMES ) +# Default depth quantization and encoding parameters. +DEPTH_QUANT_BITS: int = 12 +DEPTH_QMAX: int = (1 << DEPTH_QUANT_BITS) - 1 # 4095 + +DEFAULT_DEPTH_MIN: float = 0.01 +DEFAULT_DEPTH_MAX: float = 10.0 +DEFAULT_DEPTH_SHIFT: float = 3.5 +DEFAULT_DEPTH_USE_LOG: bool = True +DEFAULT_DEPTH_PIX_FMT: str = "gray12le" + +DEPTH_METER_UNIT: str = "m" +DEPTH_MILLIMETER_UNIT: str = "mm" +DEFAULT_DEPTH_UNIT: str = DEPTH_MILLIMETER_UNIT + + +def infer_depth_unit(dtype: np.dtype | type) -> str: + """Infer the physical unit of raw depth frames from their dtype. + + Floating-point frames are assumed to be in metres, integer frames in millimetres. + """ + return DEPTH_METER_UNIT if np.issubdtype(np.dtype(dtype), np.floating) else DEPTH_MILLIMETER_UNIT + + +# Depth-specific tuning fields persisted under ``features[*]["info"]`` as ``video.``. +DEPTH_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset({"depth_min", "depth_max", "shift", "use_log"}) + @dataclass class VideoEncoderConfig: - """Video encoder configuration. + """Video encoder configuration.""" - Attributes: - vcodec: Video encoder name. ``"auto"`` is resolved during - construction (HW encoder if available, else ``libsvtav1``). - pix_fmt: Pixel format (e.g. ``"yuv420p"``). - g: GOP size (keyframe interval). - crf: Quality level — mapped to the native quality parameter of the - codec (``crf`` for software, ``qp`` for NVENC/VAAPI, - ``q:v`` for VideoToolbox, ``global_quality`` for QSV). - preset: Speed/quality preset. Accepted type is per-codec. - fast_decode: Fast-decode tuning. For ``libsvtav1`` this is a level (0-2) - embedded in ``svtav1-params``. For ``h264`` and ``hevc`` non-zero values - set ``tune=fastdecode``. Ignored for other codecs. - video_backend: Python to be used for encoding. Only ``"pyav"`` - is currently supported. - extra_options: Free-form dictionary of additional video encoder options - (e.g. ``{"tune": "film", "profile:v": "high", "bf": 2}``). - """ - - vcodec: str = "libsvtav1" # TODO(CarolinePascal): rename to codec ? - pix_fmt: str = "yuv420p" - g: int | None = 2 - crf: int | float | None = 30 - preset: int | str | None = None - fast_decode: int = 0 + vcodec: str = "libsvtav1" # Video codec name. "auto" picks a hardware codec if available, else libsvtav1. + pix_fmt: str = "yuv420p" # Pixel format (e.g. yuv420p). + g: int | None = 2 # GOP size (keyframe interval). + crf: int | float | None = 30 # Quality level. Lower means better quality and larger files. + preset: int | str | None = None # Speed/quality preset. Accepted values are codec-specific. + fast_decode: int = 0 # Fast-decode tuning. Accepted values are codec-specific, 0 disables it. # TODO(CarolinePascal): add torchcodec support + find a way to unify the # two backends (encoding and decoding). - video_backend: str = "pyav" + video_backend: str = "pyav" # Encoding backend. Only "pyav" is currently supported. + # Extra codec options merged last, e.g. {"tune": "film"}. extra_options: dict[str, Any] = field(default_factory=dict) + # Source-data channel count this encoder is expected to handle. ``None`` + # disables the pix_fmt channel-count check; concrete subclasses set it + # (3 for RGB, 1 for depth, etc.). + _DEFAULT_CHANNELS: ClassVar[int | None] = None + def __post_init__(self) -> None: self.resolve_vcodec() # Empty-constructor ergonomics: ``VideoEncoderConfig()`` must "just work". @@ -94,9 +111,9 @@ class VideoEncoderConfig: self.validate() @classmethod - def from_video_info(cls, video_info: dict | None) -> VideoEncoderConfig: - """Reconstruct a :class:`VideoEncoderConfig` from a video feature's ``info`` block. - Missing or ``None`` values fall back to the class defaults. + def _kwargs_from_video_info(cls, video_info: dict | None) -> dict[str, Any]: + """Parse the ``video.*`` keys of a feature ``info`` block into + constructor kwargs. """ video_info = video_info or {} kwargs: dict[str, Any] = {} @@ -115,7 +132,15 @@ class VideoEncoderConfig: continue kwargs[field_name] = value - return cls(**kwargs) + return kwargs + + @classmethod + def from_video_info(cls, video_info: dict | None) -> Self: + """Reconstruct an encoder config from a video feature's ``info`` block. + + Missing or ``None`` values fall back to the class defaults. + """ + return cls(**cls._kwargs_from_video_info(video_info)) def detect_available_encoders(self, encoders: list[str] | str) -> list[str]: """Return the subset of available encoders based on the specified video backend. @@ -138,7 +163,9 @@ class VideoEncoderConfig: require_package("av", extra="dataset") from lerobot.datasets import check_video_encoder_parameters_pyav - check_video_encoder_parameters_pyav(self.vcodec, self.pix_fmt, self.get_codec_options()) + check_video_encoder_parameters_pyav( + self.vcodec, self.pix_fmt, self.get_codec_options(), channels=self._DEFAULT_CHANNELS + ) def resolve_vcodec(self) -> None: """Check ``vcodec`` and, when it is ``"auto"``, pick a concrete encoder. @@ -199,18 +226,24 @@ class VideoEncoderConfig: if encoder_threads is not None: svtav1_parts.append(f"lp={encoder_threads}") if svtav1_parts: - opts["svtav1-params"] = ":".join(svtav1_parts) + set_if("svtav1-params", ":".join(svtav1_parts)) elif self.vcodec in ("h264", "hevc"): set_if("crf", self.crf) set_if("preset", self.preset) if self.fast_decode: - opts["tune"] = "fastdecode" + set_if("tune", "fastdecode") set_if("threads", encoder_threads) + elif self.vcodec == "libaom-av1": + set_if("crf", self.crf) + set_if("preset", self.preset) + if encoder_threads is not None: + set_if("threads", encoder_threads) + set_if("row-mt", 1) elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"): if self.crf is not None: - opts["q:v"] = max(1, min(100, 100 - self.crf * 2)) + set_if("q:v", max(1, min(100, 100 - self.crf * 2))) elif self.vcodec in ("h264_nvenc", "hevc_nvenc"): - opts["rc"] = 0 + set_if("rc", 0) set_if("qp", self.crf) set_if("preset", self.preset) elif self.vcodec == "h264_vaapi": @@ -230,6 +263,79 @@ class VideoEncoderConfig: return opts -def camera_encoder_defaults() -> VideoEncoderConfig: - """Return a :class:`VideoEncoderConfig` with RGB-camera defaults.""" - return VideoEncoderConfig() +@dataclass +class RGBEncoderConfig(VideoEncoderConfig): + """Encoder configuration for RGB camera streams. + + Identical to :class:`VideoEncoderConfig` but declares the 3-channel + source-data layout so ``pix_fmt`` is validated against RGB inputs. + """ + + _DEFAULT_CHANNELS: ClassVar[int] = 3 + + +def rgb_encoder_defaults() -> RGBEncoderConfig: + """Return a :class:`RGBEncoderConfig` with RGB-camera defaults.""" + return RGBEncoderConfig() + + +@dataclass +class DepthEncoderConfig(VideoEncoderConfig): + """Encoder configuration for depth-map streams. + + Inherits the full :class:`VideoEncoderConfig` surface (codec, GOP, CRF, + preset, ``extra_options``…) and adds the parameters of the depth quantizer. + Defaults flip ``vcodec`` to ``"hevc"`` (Main 12 profile) and ``pix_fmt`` to + ``"gray12le"``. + """ + + vcodec: str = "hevc" # Video codec name. Defaults to HEVC Main 12 (a 12-bit-capable codec). + pix_fmt: str = "gray12le" # Pixel format. Defaults to 12-bit grayscale. + extra_options: dict[str, Any] = field(default_factory=lambda: {"x265-params": "lossless=1"}) + + depth_min: float = DEFAULT_DEPTH_MIN # Minimum depth in meters, mapped to the lowest quantum. + depth_max: float = DEFAULT_DEPTH_MAX # Maximum depth in meters, mapped to the highest quantum. + shift: float = DEFAULT_DEPTH_SHIFT # Pre-log offset in meters for numerical stability near zero. + use_log: bool = DEFAULT_DEPTH_USE_LOG # Use logarithmic quantization (True) or linear (False). + + _DEFAULT_CHANNELS: ClassVar[int] = 1 + + @classmethod + def _kwargs_from_video_info(cls, video_info: dict | None) -> dict[str, Any]: + """Layer the depth-specific tuning (``depth_min`` / ``depth_max`` / + ``shift`` / ``use_log``) on top of the base parser. Missing keys + fall back to the class defaults. + """ + kwargs = super()._kwargs_from_video_info(video_info) + video_info = video_info or {} + for name in DEPTH_ENCODER_INFO_FIELD_NAMES: + value = video_info.get(f"video.{name}") + if value is not None: + kwargs[name] = value + return kwargs + + +def depth_encoder_defaults() -> DepthEncoderConfig: + """Return a :class:`DepthEncoderConfig` with depth-camera defaults.""" + return DepthEncoderConfig() + + +def encoder_config_from_video_info(video_info: dict | None) -> VideoEncoderConfig: + """Build the appropriate encoder config from a feature's ``info`` block. + + Dispatches to :class:`DepthEncoderConfig` when the dict marks the feature + as a depth map and to :class:`RGBEncoderConfig` + otherwise. + + Args: + video_info: A feature's ``info`` dict as persisted in ``info.json``, + or ``None`` (treated as an empty dict). + + Returns: + A :class:`DepthEncoderConfig` for depth features, otherwise a + :class:`RGBEncoderConfig`. + """ + video_info = video_info or {} + is_depth = bool(video_info.get("is_depth_map") or video_info.get("video.is_depth_map")) + cls: type[VideoEncoderConfig] = DepthEncoderConfig if is_depth else RGBEncoderConfig + return cls.from_video_info(video_info) diff --git a/src/lerobot/datasets/__init__.py b/src/lerobot/datasets/__init__.py index bd12a7248..7715a115e 100644 --- a/src/lerobot/datasets/__init__.py +++ b/src/lerobot/datasets/__init__.py @@ -35,7 +35,7 @@ from .dataset_tools import ( remove_feature, split_dataset, ) -from .factory import make_dataset, resolve_delta_timestamps +from .factory import make_dataset, make_train_eval_datasets, resolve_delta_timestamps from .image_writer import safe_stop_image_writer from .io_utils import load_episodes, write_stats from .language import ( @@ -89,6 +89,7 @@ __all__ = [ "get_feature_stats", "load_episodes", "make_dataset", + "make_train_eval_datasets", "merge_datasets", "modify_features", "modify_tasks", diff --git a/src/lerobot/datasets/compute_stats.py b/src/lerobot/datasets/compute_stats.py index 09765c130..02ecd81a4 100644 --- a/src/lerobot/datasets/compute_stats.py +++ b/src/lerobot/datasets/compute_stats.py @@ -242,12 +242,12 @@ def sample_images(image_paths: list[str]) -> np.ndarray: images = None for i, idx in enumerate(sampled_indices): path = image_paths[idx] - # we load as uint8 to reduce memory usage + # we load RGB images as uint8 to reduce memory usage; depth keeps its native dtype img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True) img = auto_downsample_height_width(img) if images is None: - images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8) + images = np.empty((len(sampled_indices), *img.shape), dtype=img.dtype) images[i] = img @@ -506,8 +506,10 @@ def compute_episode_stats( Each statistics dictionary contains min, max, mean, std, count, and quantiles. Note: - Image statistics are normalized to [0,1] range and have shape (3,1,1) for - per-channel values when dtype is 'image' or 'video'. + For 'image'/'video' features, stats are computed per channel and kept with a + leading channel axis (e.g. shape (3, 1, 1) for RGB). RGB stats are divided by + 255 to land in [0, 1]; depth maps (features flagged with ``is_depth_map``) skip + this rescaling and remain in their stored units (stored in ``depth_unit``). """ if quantile_list is None: quantile_list = DEFAULT_QUANTILES @@ -531,8 +533,12 @@ def compute_episode_stats( ) if features[key]["dtype"] in ["image", "video"]: + normalization_factor = ( + 255.0 if not (features[key].get("info") or {}).get("is_depth_map", False) else 1.0 + ) ep_stats[key] = { - k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items() + k: v if k == "count" else np.squeeze(v / normalization_factor, axis=0) + for k, v in ep_stats[key].items() } return ep_stats @@ -552,8 +558,10 @@ def _validate_stat_value(value: np.ndarray, key: str, feature_key: str) -> None: if key == "count" and value.shape != (1,): raise ValueError(f"Shape of 'count' must be (1), but is {value.shape} instead.") - if "image" in feature_key and key != "count" and value.shape != (3, 1, 1): - raise ValueError(f"Shape of quantile '{key}' must be (3,1,1), but is {value.shape} instead.") + if "image" in feature_key and key != "count" and value.shape not in ((3, 1, 1), (1, 1, 1)): + raise ValueError( + f"Shape of quantile '{key}' must be (3,1,1) or (1,1,1) but is {value.shape} instead." + ) def _assert_type_and_shape(stats_list: list[dict[str, dict]]): diff --git a/src/lerobot/datasets/dataset_metadata.py b/src/lerobot/datasets/dataset_metadata.py index 39a1b6d2b..6e19d14fb 100644 --- a/src/lerobot/datasets/dataset_metadata.py +++ b/src/lerobot/datasets/dataset_metadata.py @@ -14,7 +14,9 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from collections.abc import Callable +import logging +from collections.abc import Callable, Iterable +from copy import deepcopy from pathlib import Path import numpy as np @@ -24,12 +26,13 @@ import pyarrow as pa import pyarrow.parquet as pq from huggingface_hub import snapshot_download -from lerobot.configs import VideoEncoderConfig +from lerobot.configs import DEPTH_METER_UNIT, VideoEncoderConfig from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE from lerobot.utils.feature_utils import _validate_feature_names from lerobot.utils.utils import flatten_dict from .compute_stats import aggregate_stats +from .depth_utils import MM_PER_METRE from .feature_utils import create_empty_dataset_info from .io_utils import ( get_file_size_in_mb, @@ -337,6 +340,54 @@ class LeRobotDatasetMetadata: """Keys to access visual modalities stored as videos.""" return [key for key, ft in self.features.items() if ft["dtype"] == "video"] + @property + def depth_keys(self) -> list[str]: + """Keys to access depth-map modalities stored as videos or images. + + A depth key is a feature whose ``info`` dict carries ``"is_depth_map": True`` + (or the legacy ``"video.is_depth_map"`` inside ``info`` or ``video_info``). + """ + + def _is_depth(ft: dict) -> bool: + info = ft.get("info") or {} + video_info = ft.get("video_info") or {} + return ( + info.get("is_depth_map", False) + or info.get("video.is_depth_map", False) + or video_info.get("video.is_depth_map", False) + ) + + return [key for key, ft in self.features.items() if _is_depth(ft)] + + def rescale_depth_stats(self, output_unit: str) -> None: + """Rescale depth feature stats in place from their recorded unit to ``output_unit``. + + Depth stats are stored in the unit the frames were recorded in + (``features[key]["info"]["depth_unit"]``), while frames are returned in + ``output_unit`` on read. This converts the unit-bearing stat entries so + stats match the frames consumers see. + """ + missing_unit_keys = [ + key for key in self.depth_keys if (self.features[key].get("info") or {}).get("depth_unit") is None + ] + if missing_unit_keys: + logging.warning( + f"Depth feature(s) {missing_unit_keys} have no recorded 'depth_unit' in their info. " + f"Depth maps and stats for these keys will be returned AS IS, with no unit conversion " + f"to the requested output unit {output_unit!r}. Re-record the dataset or set 'depth_unit' " + f"in the feature info (meta/info.json) to enable conversion." + ) + if self.stats is None: + return + for key in self.depth_keys: + stored_unit = (self.features[key].get("info") or {}).get("depth_unit") + if stored_unit is None or stored_unit == output_unit or key not in self.stats: + continue + factor = MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else 1.0 / MM_PER_METRE + self.stats[key] = { + stat: value if stat == "count" else value * factor for stat, value in self.stats[key].items() + } + @property def camera_keys(self) -> list[str]: """Keys to access visual modalities (regardless of their storage method).""" @@ -580,29 +631,48 @@ class LeRobotDatasetMetadata: def update_video_info( self, video_key: str | None = None, - camera_encoder: VideoEncoderConfig | None = None, + video_encoder: VideoEncoderConfig | None = None, + preserve_keys: Iterable[str] | None = None, ) -> None: - """Populate per-feature video info in ``info.json``. + """Populate or refresh per-feature video info in ``info.json``. Warning: this function writes info from first episode videos, implicitly assuming that all videos have been encoded the same way. Also, this means it assumes the first episode exists. + Always re-probes the videos and overwrites existing info for every recomputed + key. ``preserve_keys`` lists keys whose existing values must be kept (e.g. + data-intrinsic entries like ``is_depth_map`` and depth quantization params) + instead of being recomputed. + Args: video_key: If provided, only update this video key. Otherwise update all video keys in the dataset. - camera_encoder: Encoder configuration used to produce the + video_encoder: Encoder configuration used to produce the videos. When provided, its fields are recorded as ``video.`` entries alongside the stream-derived ``video.*`` entries (see :func:`get_video_info`). + preserve_keys: Keys whose existing values are kept instead of being + recomputed. ``None`` (default) recomputes every key. """ if video_key is not None and video_key not in self.video_keys: raise ValueError(f"Video key {video_key} not found in dataset") video_keys = [video_key] if video_key is not None else self.video_keys + preserve_set = set(preserve_keys or ()) for key in video_keys: - if not self.features[key].get("info", None): - video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0) - self.info.features[key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder) + existing = self.features[key].get("info") or {} + video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0) + new_info = get_video_info(video_path, video_encoder=video_encoder) + # Drop preserved keys so the existing values win on merge. + new_info = {k: v for k, v in new_info.items() if k not in preserve_set} + merged = {**existing, **new_info} + # Migrate the legacy depth marker to the canonical key. + if "video.is_depth_map" in merged: + logging.warning( + f"Migrating legacy 'video.is_depth_map' to 'is_depth_map' for feature {key!r}." + ) + merged.setdefault("is_depth_map", merged.pop("video.is_depth_map")) + self.info.features[key]["info"] = merged def update_chunk_settings( self, @@ -709,7 +779,7 @@ class LeRobotDatasetMetadata: obj.root.mkdir(parents=True, exist_ok=False) - features = {**features, **DEFAULT_FEATURES} + features = {**deepcopy(features), **DEFAULT_FEATURES} _validate_feature_names(features) obj.tasks = None diff --git a/src/lerobot/datasets/dataset_reader.py b/src/lerobot/datasets/dataset_reader.py index 59aaa40e5..f4e1f6a31 100644 --- a/src/lerobot/datasets/dataset_reader.py +++ b/src/lerobot/datasets/dataset_reader.py @@ -22,7 +22,14 @@ from pathlib import Path import datasets import torch +from lerobot.configs import ( + DEFAULT_DEPTH_UNIT, + DEPTH_METER_UNIT, + DepthEncoderConfig, +) + from .dataset_metadata import LeRobotDatasetMetadata +from .depth_utils import MM_PER_METRE, dequantize_depth from .feature_utils import ( check_delta_timestamps, get_delta_indices, @@ -51,6 +58,7 @@ class DatasetReader: delta_timestamps: dict[str, list[float]] | None, image_transforms: Callable | None, return_uint8: bool = False, + depth_output_unit: str = DEFAULT_DEPTH_UNIT, ): """Initialize the reader with metadata, filtering, and transform config. @@ -68,14 +76,21 @@ class DatasetReader: relative timestamp offsets for temporal context windows. image_transforms: Optional torchvision v2 transform applied to visual features. + return_uint8: If True, return RGB video frames as raw uint8 tensors + instead of normalized float32. + depth_output_unit: Physical unit depth maps are dequantized to + (``"m"`` or ``"mm"``). Defaults to ``"mm"``. """ self._meta = meta self.root = root self.episodes = episodes self._tolerance_s = tolerance_s self._video_backend = video_backend + if image_transforms is not None and not callable(image_transforms): + raise TypeError("image_transforms must be callable or None.") self._image_transforms = image_transforms self._return_uint8 = return_uint8 + self._depth_output_unit = depth_output_unit self.hf_dataset: datasets.Dataset | None = None self._absolute_to_relative_idx: dict[int, int] | None = None @@ -86,6 +101,28 @@ class DatasetReader: check_delta_timestamps(delta_timestamps, meta.fps, tolerance_s) self.delta_indices = get_delta_indices(delta_timestamps, meta.fps) + self._depth_encoder_configs: dict[str, DepthEncoderConfig] = { + vid_key: DepthEncoderConfig.from_video_info(self._meta.features[vid_key].get("info")) + for vid_key in self._meta.depth_keys + } + + # Get the input unit of each depth feature stored as raw images. + self._image_depth_units: dict[str, str | None] = { + key: (self._meta.features[key].get("info") or {}).get("depth_unit") + for key in self._meta.depth_keys + if key in self._meta.image_keys + } + + def set_image_transforms(self, image_transforms: Callable | None) -> None: + """Replace the transform applied to visual observations.""" + if image_transforms is not None and not callable(image_transforms): + raise TypeError("image_transforms must be callable or None.") + self._image_transforms = image_transforms + + def clear_image_transforms(self) -> None: + """Remove the transform applied to visual observations.""" + self._image_transforms = None + def try_load(self) -> bool: """Attempt to load from local cache. Returns True if data is sufficient.""" try: @@ -247,7 +284,18 @@ class DatasetReader: self._tolerance_s, self._video_backend, return_uint8=self._return_uint8, + is_depth=vid_key in self._meta.depth_keys, ) + if vid_key in self._meta.depth_keys: + depth_encoder = self._depth_encoder_configs[vid_key] + frames = dequantize_depth( + frames, + depth_min=depth_encoder.depth_min, + depth_max=depth_encoder.depth_max, + shift=depth_encoder.shift, + use_log=depth_encoder.use_log, + output_unit=self._depth_output_unit, + ) return vid_key, frames.squeeze(0) items = list(query_timestamps.items()) @@ -287,10 +335,18 @@ class DatasetReader: item = {**video_frames, **item} if self._image_transforms is not None: - image_keys = self._meta.camera_keys - for cam in image_keys: + for cam in self._meta.camera_keys: + if cam in self._meta.depth_keys: + continue item[cam] = self._image_transforms(item[cam]) + # Convert depth features to the output unit. + for key, stored_unit in self._image_depth_units.items(): + if key in item and stored_unit is not None and stored_unit != self._depth_output_unit: + item[key] = ( + item[key] * MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else item[key] / MM_PER_METRE + ) + # Add task as a string task_idx = item["task_index"].item() item["task"] = self._meta.tasks.iloc[task_idx].name diff --git a/src/lerobot/datasets/dataset_tools.py b/src/lerobot/datasets/dataset_tools.py index 91dc66af2..31e075d7c 100644 --- a/src/lerobot/datasets/dataset_tools.py +++ b/src/lerobot/datasets/dataset_tools.py @@ -27,6 +27,7 @@ import logging import shutil from collections.abc import Callable from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed +from copy import deepcopy from pathlib import Path import datasets @@ -36,7 +37,15 @@ import pyarrow.parquet as pq import torch from tqdm import tqdm -from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults +from lerobot.configs import ( + DepthEncoderConfig, + RGBEncoderConfig, + VideoEncoderConfig, + depth_encoder_defaults, + encoder_config_from_video_info, + rgb_encoder_defaults, +) +from lerobot.configs.video import DEPTH_ENCODER_INFO_FIELD_NAMES from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE from lerobot.utils.utils import flatten_dict @@ -47,6 +56,7 @@ from .compute_stats import ( compute_relative_action_stats, ) from .dataset_metadata import LeRobotDatasetMetadata +from .image_writer import write_image from .io_utils import ( get_parquet_file_size_in_mb, load_episodes, @@ -61,12 +71,13 @@ from .utils import ( DEFAULT_DATA_FILE_SIZE_IN_MB, DEFAULT_DATA_PATH, DEFAULT_EPISODES_PATH, + DEPTH_FILE_PATTERN, + IMAGE_FILE_PATTERN, VIDEO_DIR, update_chunk_file_indices, ) from .video_utils import ( encode_video_frames, - get_video_info, reencode_video, ) @@ -600,7 +611,7 @@ def _keep_episodes_from_video_with_av( output_path: Path, episodes_to_keep: list[tuple[int, int]], fps: float, - camera_encoder: VideoEncoderConfig, + video_encoder: VideoEncoderConfig, ) -> None: """Keep only specified episodes from a video file using PyAV. @@ -614,7 +625,7 @@ def _keep_episodes_from_video_with_av( Ranges are half-open intervals: [start_frame, end_frame), where start_frame is inclusive and end_frame is exclusive. fps: Frame rate of the video. - camera_encoder: Video encoder settings used to re-encode the kept frames. + video_encoder: Video encoder settings used to re-encode the kept frames. """ from fractions import Fraction @@ -639,13 +650,13 @@ def _keep_episodes_from_video_with_av( # Convert fps to Fraction for PyAV compatibility. fps_fraction = Fraction(fps).limit_denominator(1000) - codec_options = camera_encoder.get_codec_options(as_strings=True) - v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options) + codec_options = video_encoder.get_codec_options(as_strings=True) + v_out = out.add_stream(video_encoder.vcodec, rate=fps_fraction, options=codec_options) # PyAV type stubs don't distinguish video streams from audio/subtitle streams. v_out.width = v_in.codec_context.width v_out.height = v_in.codec_context.height - v_out.pix_fmt = camera_encoder.pix_fmt + v_out.pix_fmt = video_encoder.pix_fmt # Set time_base to match the frame rate for proper timestamp handling. v_out.time_base = Fraction(1, int(fps)) @@ -732,7 +743,7 @@ def _copy_and_reindex_videos( for video_key in src_dataset.meta.video_keys: logging.info(f"Processing videos for {video_key}") - camera_encoder = VideoEncoderConfig.from_video_info( + video_encoder = encoder_config_from_video_info( src_dataset.meta.info.features.get(video_key, {}).get("info") ) @@ -816,7 +827,7 @@ def _copy_and_reindex_videos( dst_video_path, episodes_to_keep_ranges, src_dataset.meta.fps, - camera_encoder, + video_encoder, ) cumulative_ts = 0.0 @@ -873,11 +884,11 @@ def _copy_and_reindex_episodes_metadata( episode_meta.update(video_metadata[new_idx]) # Extract episode statistics from parquet metadata. - # Note (maractingi): When pandas/pyarrow serializes numpy arrays with shape (3, 1, 1) to parquet, + # When pandas/pyarrow serializes numpy arrays with shape (C, 1, 1) to parquet, # they are being deserialized as nested object arrays like: # array([array([array([0.])]), array([array([0.])]), array([array([0.])])]) # This happens particularly with image/video statistics. We need to detect and flatten - # these nested structures back to proper (3, 1, 1) arrays so aggregate_stats can process them. + # these nested structures back to proper (C, 1, 1) arrays so aggregate_stats can process them. episode_stats = {} for key in src_episode_full: if key.startswith("stats/"): @@ -893,15 +904,16 @@ def _copy_and_reindex_episodes_metadata( if feature_name in src_dataset.meta.features: feature_dtype = src_dataset.meta.features[feature_name]["dtype"] if feature_dtype in ["image", "video"] and stat_name != "count": + # Stats are channel-first (C, 1, 1) if isinstance(value, np.ndarray) and value.dtype == object: flat_values = [] for item in value: while isinstance(item, np.ndarray): item = item.flatten()[0] flat_values.append(item) - value = np.array(flat_values, dtype=np.float64).reshape(3, 1, 1) - elif isinstance(value, np.ndarray) and value.shape == (3,): - value = value.reshape(3, 1, 1) + value = np.array(flat_values, dtype=np.float64).reshape(-1, 1, 1) + elif isinstance(value, np.ndarray) and value.ndim == 1: + value = value.reshape(-1, 1, 1) episode_stats[feature_name][stat_name] = value @@ -1101,7 +1113,9 @@ def _copy_episodes_metadata_and_stats( if dst_meta.video_keys and src_dataset.meta.video_keys: for key in dst_meta.video_keys: if key in src_dataset.meta.features: - dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {}) + dst_meta.info.features[key]["info"] = deepcopy( + src_dataset.meta.info.features[key].get("info", {}) + ) write_info(dst_meta.info, dst_meta.root) @@ -1150,15 +1164,15 @@ def _save_episode_images_for_video( # Get all items for this episode episode_dataset = imgs_dataset.select(range(from_idx, to_idx)) + is_depth = img_key in dataset.meta.depth_keys + frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN + # Define function to save a single image def save_single_image(i_item_tuple): i, item = i_item_tuple - img = item[img_key] - # Use frame-XXXXXX.png format to match encode_video_frames expectations - img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100) + write_image(item[img_key], imgs_dir / frame_pattern.format(frame_index=i)) return i - # Save images with proper naming convention for encode_video_frames (frame-XXXXXX.png) items = list(enumerate(episode_dataset)) with ThreadPoolExecutor(max_workers=num_workers) as executor: @@ -1190,13 +1204,14 @@ def _save_batch_episodes_images( hf_dataset = dataset.hf_dataset.with_format(None) imgs_dataset = hf_dataset.select_columns(img_key) + is_depth = img_key in dataset.meta.depth_keys + frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN + # Define function to save a single image with global frame index # Defined once outside the loop to avoid repeated closure creation def save_single_image(i_item_tuple, base_frame_idx, img_key_param): i, item = i_item_tuple - img = item[img_key_param] - # Use global frame index for naming - img.save(str(imgs_dir / f"frame-{base_frame_idx + i:06d}.png"), quality=100) + write_image(item[img_key_param], imgs_dir / frame_pattern.format(frame_index=base_frame_idx + i)) return i episode_durations = [] @@ -1287,7 +1302,7 @@ def _estimate_frame_size_via_calibration( episode_indices: list[int], temp_dir: Path, fps: int, - camera_encoder: VideoEncoderConfig, + video_encoder: VideoEncoderConfig, num_calibration_frames: int = 30, ) -> float: """Estimate MB per frame by encoding a small calibration sample. @@ -1301,7 +1316,7 @@ def _estimate_frame_size_via_calibration( episode_indices: List of episode indices being processed. temp_dir: Temporary directory for calibration files. fps: Frames per second for video encoding. - camera_encoder: Video encoder settings used for calibration encoding. + video_encoder: Video encoder settings used for calibration encoding. num_calibration_frames: Number of frames to use for calibration (default: 30). Returns: @@ -1326,10 +1341,11 @@ def _estimate_frame_size_via_calibration( hf_dataset = dataset.hf_dataset.with_format(None) sample_indices = range(from_idx, from_idx + num_frames) - # Save calibration frames + # Save calibration frames using the suffix/format the encoder expects. + is_depth = img_key in dataset.meta.depth_keys + frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN for i, idx in enumerate(sample_indices): - img = hf_dataset[idx][img_key] - img.save(str(calibration_dir / f"frame-{i:06d}.png"), quality=100) + write_image(hf_dataset[idx][img_key], calibration_dir / frame_pattern.format(frame_index=i)) # Encode calibration video calibration_video_path = calibration_dir / "calibration.mp4" @@ -1337,7 +1353,7 @@ def _estimate_frame_size_via_calibration( imgs_dir=calibration_dir, video_path=calibration_video_path, fps=fps, - camera_encoder=camera_encoder, + video_encoder=video_encoder, overwrite=True, ) @@ -1610,6 +1626,7 @@ def recompute_stats( raise ValueError(f"No parquet files found in {data_dir}") all_episode_stats = [] + # TODO: enable image and video stats re-computation numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]] for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"): @@ -1655,7 +1672,8 @@ def convert_image_to_video_dataset( dataset: LeRobotDataset, output_dir: Path | None = None, repo_id: str | None = None, - camera_encoder: VideoEncoderConfig | None = None, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, episode_indices: list[int] | None = None, num_workers: int = 4, max_episodes_per_batch: int | None = None, @@ -1667,21 +1685,32 @@ def convert_image_to_video_dataset( LeRobot dataset structure with videos stored in chunked MP4 files. Args: - dataset: The source LeRobot dataset with images - output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig. - repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig. - camera_encoder: Video encoder settings - (``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`). - episode_indices: List of episode indices to convert (None = all episodes) - num_workers: Number of threads for parallel processing (default: 4) - max_episodes_per_batch: Maximum episodes per video batch to avoid memory issues (None = no limit) - max_frames_per_batch: Maximum frames per video batch to avoid memory issues (None = no limit) + dataset: The source LeRobot dataset with images. + output_dir: Root directory where the converted dataset will be stored. When + ``None``, defaults to ``$HF_LEROBOT_HOME/repo_id``. Equivalent to + ``new_root`` in ``EditDatasetConfig``. + repo_id: Converted dataset identifier. Equivalent to ``new_repo_id`` in + ``EditDatasetConfig``. + rgb_encoder: Video encoder settings applied to RGB cameras. When ``None``, + :func:`~lerobot.configs.video.rgb_encoder_defaults` is used. + depth_encoder: Video encoder settings applied to depth-map cameras, including + the quantization parameters persisted to the dataset metadata. When + ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used. + episode_indices: Episode indices to convert. When ``None``, all episodes are + converted. + num_workers: Number of threads for parallel processing. + max_episodes_per_batch: Maximum episodes per video batch, to bound memory use. + ``None`` means no limit. + max_frames_per_batch: Maximum frames per video batch, to bound memory use. + ``None`` means no limit. Returns: - New LeRobotDataset with images encoded as videos + A new :class:`LeRobotDataset` with images encoded as videos. """ - if camera_encoder is None: - camera_encoder = camera_encoder_defaults() + if rgb_encoder is None: + rgb_encoder = rgb_encoder_defaults() + if depth_encoder is None: + depth_encoder = depth_encoder_defaults() # Check that it's an image dataset if len(dataset.meta.video_keys) > 0: @@ -1706,10 +1735,7 @@ def convert_image_to_video_dataset( logging.info( f"Converting {len(episode_indices)} episodes with {len(img_keys)} cameras from {dataset.repo_id}" ) - logging.info( - f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, " - f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}" - ) + logging.info(f"RGB video encoder: {rgb_encoder}, depth video encoder: {depth_encoder}") # Create new features dict, converting image features to video features new_features = {} @@ -1771,6 +1797,8 @@ def convert_image_to_video_dataset( episode_lengths = {ep_idx: dataset.meta.episodes["length"][ep_idx] for ep_idx in episode_indices} for img_key in tqdm(img_keys, desc="Processing cameras"): + target_encoder = depth_encoder if img_key in dataset.meta.depth_keys else rgb_encoder + # Estimate size per frame by encoding a small calibration sample # This provides accurate compression ratio for the specific codec parameters size_per_frame_mb = _estimate_frame_size_via_calibration( @@ -1779,7 +1807,7 @@ def convert_image_to_video_dataset( episode_indices=episode_indices, temp_dir=temp_dir, fps=fps, - camera_encoder=camera_encoder, + video_encoder=target_encoder, ) logging.info(f"Processing camera: {img_key}") @@ -1821,7 +1849,7 @@ def convert_image_to_video_dataset( imgs_dir=imgs_dir, video_path=video_path, fps=fps, - camera_encoder=camera_encoder, + video_encoder=target_encoder, overwrite=True, ) @@ -1860,16 +1888,11 @@ def convert_image_to_video_dataset( new_meta.info.total_tasks = dataset.meta.total_tasks new_meta.info.splits = {"train": f"0:{len(episode_indices)}"} - # Update video info for all image keys (now videos) - # We need to manually set video info since update_video_info() checks video_keys first + # Update video info for all image keys (now videos). They are registered as + # video features above, so update_video_info populates their (still-empty) info. for img_key in img_keys: - if not new_meta.features[img_key].get("info", None): - video_path = new_meta.root / new_meta.video_path.format( - video_key=img_key, chunk_index=0, file_index=0 - ) - new_meta.info.features[img_key]["info"] = get_video_info( - video_path, camera_encoder=camera_encoder - ) + target_encoder = depth_encoder if img_key in dataset.meta.depth_keys else rgb_encoder + new_meta.update_video_info(video_key=img_key, video_encoder=target_encoder) write_info(new_meta.info, new_meta.root) @@ -1896,11 +1919,11 @@ def convert_image_to_video_dataset( def _reencode_video_worker(args: tuple) -> Path: """Picklable worker for :func:`reencode_dataset`'s process pool.""" - video_path, camera_encoder, encoder_threads = args + video_path, video_encoder, encoder_threads = args reencode_video( input_video_path=video_path, output_video_path=video_path, - camera_encoder=camera_encoder, + video_encoder=video_encoder, encoder_threads=encoder_threads, overwrite=True, ) @@ -1909,7 +1932,8 @@ def _reencode_video_worker(args: tuple) -> Path: def reencode_dataset( dataset: LeRobotDataset, - camera_encoder: VideoEncoderConfig, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, encoder_threads: int | None = None, num_workers: int | None = None, ) -> LeRobotDataset: @@ -1920,8 +1944,11 @@ def reencode_dataset( Args: dataset: An existing :class:`LeRobotDataset` whose videos will be re-encoded. - camera_encoder: Target encoder configuration applied to every video - file. + rgb_encoder: Target encoder configuration applied to every RGB video + file. If ``None``, re-encoding is skipped for RGB videos. + depth_encoder: Target encoder configuration applied to every depth video + file. If ``None``, re-encoding is skipped for depth videos. + Quantization parameters will not override the ones in the current dataset. encoder_threads: Per-encoder thread count forwarded to :func:`reencode_video`. ``None`` lets the codec decide. num_workers: Number of parallel processes. ``None`` or ``0`` means @@ -1933,23 +1960,35 @@ def reencode_dataset( on disk. """ meta = dataset.meta - video_paths_list = [] + video_keys_encoders_dict = {} + video_keys_paths_dict = {} + + if rgb_encoder is None and depth_encoder is None: + raise ValueError("Either rgb_encoder or depth_encoder must be provided") # Only re-encode if the videos are not already encoded with the given video encoding parameters for video_key in meta.video_keys: current_info = meta.info.features[video_key].get("info", {}) - current_encoder = VideoEncoderConfig.from_video_info(current_info) - if current_encoder != camera_encoder: - video_paths_list.extend((meta.root / VIDEO_DIR / video_key).rglob("*.mp4")) + current_encoder = encoder_config_from_video_info(current_info) + target_encoder = depth_encoder if video_key in meta.depth_keys else rgb_encoder + if target_encoder is None: + logging.info(f"No encoder provided for {video_key} video. Skipping re-encoding.") + elif current_encoder != target_encoder: + video_keys_paths_dict[video_key] = list((meta.root / VIDEO_DIR / video_key).rglob("*.mp4")) + video_keys_encoders_dict[video_key] = target_encoder else: - logging.info(f"{video_key} videos are already encoded with {camera_encoder}. Nothing to do.") + logging.info(f"{video_key} videos are already encoded with {target_encoder}. Nothing to do.") - if len(video_paths_list) == 0: + if len(video_keys_paths_dict) == 0: logging.warning("Dataset has no videos to re-encode.") return dataset - logging.info(f"Re-encoding {len(video_paths_list)} video file(s) with {camera_encoder}") + logging.info(f"Re-encoding {sum(len(paths) for paths in video_keys_paths_dict.values())} video file(s).") - worker_args = [(vp, camera_encoder, encoder_threads) for vp in video_paths_list] + worker_args = [ + (path, encoder, encoder_threads) + for video_key, encoder in video_keys_encoders_dict.items() + for path in video_keys_paths_dict[video_key] + ] if num_workers and num_workers > 1: with ProcessPoolExecutor(max_workers=num_workers) as pool: futures = [pool.submit(_reencode_video_worker, args) for args in worker_args] @@ -1963,10 +2002,15 @@ def reencode_dataset( for args in tqdm(worker_args, desc="Re-encoding videos"): _reencode_video_worker(args) - # Refresh video info in metadata for every video key. - for vid_key in meta.video_keys: - video_path = meta.root / meta.get_video_file_path(0, vid_key) - meta.info.features[vid_key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder) + # Refresh video info in metadata for every re-encoded key. Re-encoding only + # changes codec/container params, so for depth videos we preserve ``is_depth_map`` + # and the depth quantization params (``video.depth_min`` / ``video.depth_max`` / + # ...), which describe the data rather than the codec and must survive a transcode. + # RGB videos pass an empty set: still a refresh, but nothing to preserve. + depth_preserve_keys = {"is_depth_map", *(f"video.{n}" for n in DEPTH_ENCODER_INFO_FIELD_NAMES)} + for video_key, encoder in video_keys_encoders_dict.items(): + preserve_keys = depth_preserve_keys if video_key in meta.depth_keys else set() + meta.update_video_info(video_key=video_key, video_encoder=encoder, preserve_keys=preserve_keys) write_info(meta.info, meta.root) logging.info("Dataset metadata updated.") diff --git a/src/lerobot/datasets/dataset_writer.py b/src/lerobot/datasets/dataset_writer.py index 633c00c1a..a6049312f 100644 --- a/src/lerobot/datasets/dataset_writer.py +++ b/src/lerobot/datasets/dataset_writer.py @@ -31,7 +31,14 @@ import PIL.Image import pyarrow.parquet as pq import torch -from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults +from lerobot.configs import ( + DepthEncoderConfig, + RGBEncoderConfig, + VideoEncoderConfig, + depth_encoder_defaults, + infer_depth_unit, + rgb_encoder_defaults, +) from .compute_stats import compute_episode_stats from .dataset_metadata import LeRobotDatasetMetadata @@ -48,6 +55,7 @@ from .io_utils import ( write_info, ) from .utils import ( + DEFAULT_DEPTH_PATH, DEFAULT_EPISODES_PATH, DEFAULT_IMAGE_PATH, update_chunk_file_indices, @@ -67,17 +75,22 @@ def _encode_video_worker( episode_index: int, root: Path, fps: int, - camera_encoder: VideoEncoderConfig | None = None, + video_encoder: VideoEncoderConfig | None = None, encoder_threads: int | None = None, ) -> Path: temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4" - fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0) + path_template = ( + DEFAULT_DEPTH_PATH + if video_encoder is not None and isinstance(video_encoder, DepthEncoderConfig) + else DEFAULT_IMAGE_PATH + ) + fpath = path_template.format(image_key=video_key, episode_index=episode_index, frame_index=0) img_dir = (root / fpath).parent encode_video_frames( img_dir, temp_path, fps, - camera_encoder=camera_encoder, + video_encoder=video_encoder, encoder_threads=encoder_threads, overwrite=True, ) @@ -96,7 +109,8 @@ class DatasetWriter: self, meta: LeRobotDatasetMetadata, root: Path, - camera_encoder: VideoEncoderConfig | None, + rgb_encoder: RGBEncoderConfig | None, + depth_encoder: DepthEncoderConfig | None, encoder_threads: int | None, batch_encoding_size: int, streaming_encoder: StreamingVideoEncoder | None = None, @@ -108,8 +122,11 @@ class DatasetWriter: meta: Dataset metadata instance (used for feature schema, chunk settings, and episode persistence). root: Local dataset root directory. - camera_encoder: Video encoder settings applied to all cameras. - ``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`. + rgb_encoder: Video encoder settings applied to RGB cameras. When + ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used. + depth_encoder: Video encoder settings applied to depth cameras, including + the quantization parameters. When ``None``, + :func:`~lerobot.configs.video.depth_encoder_defaults` is used. encoder_threads: Number of encoder threads (global). ``None`` lets the codec decide. batch_encoding_size: Number of episodes to accumulate before @@ -120,7 +137,8 @@ class DatasetWriter: """ self._meta = meta self._root = root - self._camera_encoder = camera_encoder or camera_encoder_defaults() + self._rgb_encoder = rgb_encoder or rgb_encoder_defaults() + self._depth_encoder = depth_encoder or depth_encoder_defaults() self._encoder_threads = encoder_threads self._batch_encoding_size = batch_encoding_size self._streaming_encoder = streaming_encoder @@ -145,7 +163,8 @@ class DatasetWriter: return ep_buffer def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path: - fpath = DEFAULT_IMAGE_PATH.format( + path_template = DEFAULT_DEPTH_PATH if image_key in self._meta.depth_keys else DEFAULT_IMAGE_PATH + fpath = path_template.format( image_key=image_key, episode_index=episode_index, frame_index=frame_index ) return self._root / fpath @@ -191,10 +210,20 @@ class DatasetWriter: self.episode_buffer["timestamp"].append(timestamp) self.episode_buffer["task"].append(frame.pop("task")) + # Record each depth feature's input unit once, inferred from the first frame's dtype. + if frame_index == 0: + for depth_key in self._meta.depth_keys: + if depth_key not in frame: + continue + info = self._meta.features[depth_key].setdefault("info", {}) + if info.get("depth_unit") is None: + info["depth_unit"] = infer_depth_unit(np.asarray(frame[depth_key]).dtype) + # Start streaming encoder on first frame of episode if frame_index == 0 and self._streaming_encoder is not None: self._streaming_encoder.start_episode( video_keys=list(self._meta.video_keys), + depth_video_keys=list(self._meta.depth_keys), temp_dir=self._root, ) @@ -282,10 +311,13 @@ class DatasetWriter: if use_streaming: streaming_results = self._streaming_encoder.finish_episode() for video_key in self._meta.video_keys: + normalization_factor = 255.0 if video_key not in self._meta.depth_keys else 1.0 temp_path, video_stats = streaming_results[video_key] if video_stats is not None: ep_stats[video_key] = { - k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0) + k: v + if k == "count" + else np.squeeze(v.reshape(1, -1, 1, 1) / normalization_factor, axis=0) for k, v in video_stats.items() } ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path)) @@ -300,7 +332,7 @@ class DatasetWriter: episode_index, self._root, self._meta.fps, - self._camera_encoder, + self._depth_encoder if video_key in self._meta.depth_keys else self._rgb_encoder, self._encoder_threads, ): video_key for video_key in self._meta.video_keys @@ -511,7 +543,12 @@ class DatasetWriter: # Update video info (only needed when first episode is encoded) if episode_index == 0: - self._meta.update_video_info(video_key, camera_encoder=self._camera_encoder) + self._meta.update_video_info( + video_key, + video_encoder=self._depth_encoder + if video_key in self._meta.depth_keys + else self._rgb_encoder, + ) write_info(self._meta.info, self._meta.root) metadata = { @@ -578,13 +615,14 @@ class DatasetWriter: self.image_writer.wait_until_done() def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path: - """Use ffmpeg to convert frames stored as png into mp4 videos.""" + """Use ffmpeg to convert frames stored as png/tiff into mp4 videos.""" + is_depth = video_key in self._meta.depth_keys return _encode_video_worker( video_key, episode_index, self._root, self._meta.fps, - self._camera_encoder, + self._depth_encoder if is_depth else self._rgb_encoder, self._encoder_threads, ) diff --git a/src/lerobot/datasets/depth_utils.py b/src/lerobot/datasets/depth_utils.py new file mode 100644 index 000000000..a4e187eb4 --- /dev/null +++ b/src/lerobot/datasets/depth_utils.py @@ -0,0 +1,265 @@ +#!/usr/bin/env python + +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Depth encoding/decoding helpers for :class:`DepthEncoderConfig`. +""" + +import math +from typing import Literal + +import av +import numpy as np +import torch +from numpy.typing import NDArray + +from lerobot.configs.video import ( + DEFAULT_DEPTH_MAX, + DEFAULT_DEPTH_MIN, + DEFAULT_DEPTH_PIX_FMT, + DEFAULT_DEPTH_SHIFT, + DEFAULT_DEPTH_USE_LOG, + DEPTH_METER_UNIT, + DEPTH_MILLIMETER_UNIT, + DEPTH_QMAX, + infer_depth_unit, +) + +from .image_writer import squeeze_single_channel +from .pyav_utils import write_u16_plane + +MM_PER_METRE = 1000.0 +_UINT16_MAX = 65535 + + +def _validate_log_quant_params(depth_min: float, shift: float) -> None: + """Ensure ``log(depth_min + shift)`` is finite.""" + if depth_min + shift <= 0: + raise ValueError( + f"depth_min + shift must be positive for logarithmic quantization, " + f"got depth_min={depth_min} + shift={shift} = {depth_min + shift}" + ) + + +def _depth_input_to_float32_and_unit( + depth: NDArray[np.integer] | NDArray[np.floating], + input_unit: Literal["auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT], +) -> tuple[NDArray[np.float32], Literal[DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT]]: + """Convert depth to float32 in the chosen unit, and return the resolved unit.""" + resolved_unit = infer_depth_unit(depth.dtype) if input_unit == "auto" else input_unit + return depth.astype(np.float32, order="K"), resolved_unit + + +def quantize_depth( + depth: NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor, + depth_min: float = DEFAULT_DEPTH_MIN, + depth_max: float = DEFAULT_DEPTH_MAX, + shift: float = DEFAULT_DEPTH_SHIFT, + use_log: bool = DEFAULT_DEPTH_USE_LOG, + pix_fmt: str = DEFAULT_DEPTH_PIX_FMT, + video_backend: str | None = "pyav", + input_unit: Literal["auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT] = "auto", +) -> NDArray[np.uint16] | av.VideoFrame: + """Quantize depth to 12-bit codes (``uint16``, values ``0…DEPTH_QMAX``). + + Depth maps are packed into 12-bit integer frames so they fit in standard + high-bit-depth pixel formats (e.g. ``yuv420p12le`` / ``gray12le``) + and can be encoded by widely supported video codecs (e.g. HEVC Main 12). + Logarithmic quantization is the default because it allocates more quanta + to near-range depth, which matches the (1/depth) error profile of typical + depth sensors. Math is ported from BEHAVIOR-1K's ``obs_utils.py``. + + **Input units**: + + - ``input_unit="auto"`` (default): infer from dtype (floating = m, non-floating = mm). + - ``input_unit="mm"``: interpret input values as millimetres. + - ``input_unit="m"``: interpret input values as metres. + + Quantization math runs in the **resolved input unit**. + + ``depth_min``, ``depth_max``, and ``shift`` are always in **metres**. + + Args: + depth: Depth map; ``torch.Tensor`` is moved to CPU for conversion. + depth_min: Depth (metres) at quantum ``0``. + depth_max: Depth (metres) at quantum :data:`DEPTH_QMAX`. + shift: Depth shift (metres); used in log mode. Must satisfy ``depth_min + shift > 0``. + use_log: If ``True`` (default), quantize in log space. + video_backend: Video backend to use for encoding. Defaults to "pyav". + input_unit: Input unit policy (``"auto"``, ``"mm"``, ``"m"``). + + Returns: + ``numpy.ndarray``, ``dtype=uint16``, same shape as ``depth``, values in + ``[0, DEPTH_QMAX]``. + + Raises: + ValueError: If ``input_unit`` is not ``"auto"``, ``"mm"``, or ``"m"``. + ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``. + """ + if input_unit not in ("auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT): + raise ValueError( + f"input_unit must be 'auto', '{DEPTH_METER_UNIT}', or '{DEPTH_MILLIMETER_UNIT}', got {input_unit!r}" + ) + + if isinstance(depth, torch.Tensor): + depth = depth.detach().cpu().numpy() + + # Squeeze single-channel dim: (H, W, 1) or (1, H, W) → (H, W) + depth = squeeze_single_channel(depth) + + depth_f, resolved_unit = _depth_input_to_float32_and_unit(depth, input_unit=input_unit) + + # Convert depth_min, depth_max, and shift to the resolved input unit. + depth_min_u = ( + np.float32(depth_min) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_min * MM_PER_METRE) + ) + depth_max_u = ( + np.float32(depth_max) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_max * MM_PER_METRE) + ) + shift_u = np.float32(shift) if resolved_unit == DEPTH_METER_UNIT else np.float32(shift * MM_PER_METRE) + + # Normalization and quantization is performed in the resolved input unit. + if use_log: + _validate_log_quant_params(depth_min, shift) + log_min = math.log(float(depth_min_u + shift_u)) + log_max = math.log(float(depth_max_u + shift_u)) + norm = (np.log(depth_f + shift_u) - log_min) / (log_max - log_min) + else: + norm = (depth_f - depth_min_u) / (depth_max_u - depth_min_u) + + quantized = np.rint(norm * DEPTH_QMAX).clip(0, DEPTH_QMAX).astype(np.uint16, copy=False) + + if video_backend == "pyav": + frame = av.VideoFrame.from_ndarray(quantized, format=pix_fmt) + write_u16_plane(frame.planes[0], quantized) + return frame + else: + return quantized + + +def dequantize_depth( + quantized: NDArray[np.uint16] | av.VideoFrame | torch.Tensor, + depth_min: float = DEFAULT_DEPTH_MIN, + depth_max: float = DEFAULT_DEPTH_MAX, + shift: float = DEFAULT_DEPTH_SHIFT, + use_log: bool = DEFAULT_DEPTH_USE_LOG, + pix_fmt: str = DEFAULT_DEPTH_PIX_FMT, + output_unit: Literal[DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT] = DEPTH_MILLIMETER_UNIT, + output_tensor: bool = True, + output_channel_last: bool = False, +) -> NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor: + """Inverse of :func:`quantize_depth`. + + Decoding inverts the same normalized code mapping as :func:`quantize_depth` + using ``depth_min`` / ``depth_max`` / ``shift`` (in metres), then returns + the requested output unit. Tuning arguments **must match** :func:`quantize_depth`. + + Accepted input layouts : + + - ``(H, W, 1)`` or ``(H, W)`` — single frame with channel-last. + - ``(..., 1, H, W)`` — batched frames with channel-first. + - ``(..., H, W, 1)`` — batched frames with channel-last. + Output layout is determined by ``output_channel_last``. + + Args: + quantized: 12-bit codes in ``[0, DEPTH_QMAX]``. ``np.ndarray``, + ``av.VideoFrame``, or ``torch.Tensor`` (any integer or float dtype). + depth_min, depth_max, shift, use_log: Same as :func:`quantize_depth` (metres). + pix_fmt: Pixel format used to extract the plane from an ``av.VideoFrame``. + output_unit: ``"mm"`` returns ``uint16`` millimetres (rint, clip + ``[0, 65535]``) when returning a numpy array, or ``float32`` mm when + ``output_tensor=True``. ``"m"`` returns ``float32`` metres in + ``[depth_min, depth_max]``. + output_tensor: If True, return a ``torch.Tensor`` instead of a numpy array. + + Returns: + Depth map in the requested unit and dtype. + + Raises: + ValueError: If ``output_unit`` is not ``"m"`` or ``"mm"``. + ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``. + """ + if output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT): + raise ValueError( + f"output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {output_unit!r}" + ) + if use_log: + _validate_log_quant_params(depth_min, shift) + + if isinstance(quantized, av.VideoFrame): + quantized = quantized.to_ndarray(format=pix_fmt) + + # Compute the scale and offset first. + depth_min_m = float(depth_min) + depth_max_m = float(depth_max) + shift_m = float(shift) + if use_log: + log_min = math.log(depth_min_m + shift_m) + log_max = math.log(depth_max_m + shift_m) + scale = (log_max - log_min) / DEPTH_QMAX + offset = log_min + else: + scale = (depth_max_m - depth_min_m) / DEPTH_QMAX + offset = depth_min_m + + # ── Torch path: stay on the input device, single fp32 allocation. ──────── + if isinstance(quantized, torch.Tensor): + if quantized.ndim >= 3: + # Drop the single-channel dimension so the math runs on (..., H, W). + quantized = quantized.squeeze(-3) if quantized.shape[-3] == 1 else quantized.squeeze(-1) + + # Single allocation we own; everything else is in-place. + buf = quantized.to(dtype=torch.float32, copy=True) + buf.mul_(scale).add_(offset) + if use_log: + buf.exp_().sub_(shift_m) + buf.clamp_(depth_min_m, depth_max_m) + buf.unsqueeze_(-1) if output_channel_last else buf.unsqueeze_(-3) + + if output_unit == DEPTH_METER_UNIT: + return buf if output_tensor else buf.cpu().numpy() + + # mm path: round + clamp in float32, skipping the uint16 round-trip + # when returning a tensor (torch.uint16 is poorly supported). + buf.mul_(MM_PER_METRE).round_().clamp_(0.0, _UINT16_MAX) + if output_tensor: + return buf + return buf.cpu().numpy().astype(np.uint16, copy=False) + + # ── NumPy path: single fp32 allocation, ``out=`` for in-place math. ───── + arr = np.asarray(quantized) + if arr.ndim >= 3: + # Drop the single-channel dimension so the math runs on (..., H, W). + arr = np.squeeze(arr, axis=-3) if arr.shape[-3] == 1 else np.squeeze(arr, axis=-1) + + buf = np.empty(arr.shape, dtype=np.float32) + np.multiply(arr, scale, out=buf) + np.add(buf, offset, out=buf) + if use_log: + np.exp(buf, out=buf) + np.subtract(buf, shift_m, out=buf) + np.clip(buf, depth_min_m, depth_max_m, out=buf) + buf = np.expand_dims(buf, axis=-1) if output_channel_last else np.expand_dims(buf, axis=-3) + + if output_unit == DEPTH_METER_UNIT: + return torch.from_numpy(buf) if output_tensor else buf + + np.multiply(buf, MM_PER_METRE, out=buf) + np.rint(buf, out=buf) + np.clip(buf, 0.0, _UINT16_MAX, out=buf) + if output_tensor: + # torch.uint16 support is very limited; return float32 millimetres. + return torch.from_numpy(buf) + return buf.astype(np.uint16, copy=False) diff --git a/src/lerobot/datasets/factory.py b/src/lerobot/datasets/factory.py index cbbe83dc8..da7b4365a 100644 --- a/src/lerobot/datasets/factory.py +++ b/src/lerobot/datasets/factory.py @@ -14,6 +14,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import logging +import math from pprint import pformat import torch @@ -96,6 +97,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas revision=cfg.dataset.revision, video_backend=cfg.dataset.video_backend, return_uint8=True, + depth_output_unit=cfg.dataset.depth_output_unit, tolerance_s=cfg.tolerance_s, ) else: @@ -126,7 +128,87 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas if cfg.dataset.use_imagenet_stats: for key in dataset.meta.camera_keys: + if key in dataset.meta.depth_keys: + continue # Exclude depth keys from ImageNet stats for stats_type, stats in IMAGENET_STATS.items(): dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32) return dataset + + +def make_train_eval_datasets( + cfg: TrainPipelineConfig, +) -> tuple[LeRobotDataset | MultiLeRobotDataset, LeRobotDataset | None]: + """Create train and optional eval datasets by splitting episodes based on eval_split. + + The last ceil(n_episodes * eval_split) episodes per task are held out for evaluation. + If eval_split == 0.0, returns (full_dataset, None). + """ + full_dataset = make_dataset(cfg) + + if cfg.dataset.eval_split == 0.0: + return full_dataset, None + + base_episodes = ( + full_dataset.episodes if full_dataset.episodes is not None else list(range(full_dataset.num_episodes)) + ) + + episode_tasks = full_dataset.meta.episodes["tasks"] + task_to_episodes: dict[str, list[int]] = {} + for ep_idx in base_episodes: + task_key = episode_tasks[ep_idx][0] if episode_tasks[ep_idx] else "" + task_to_episodes.setdefault(task_key, []).append(ep_idx) + + train_episodes, eval_episodes = [], [] + for eps in task_to_episodes.values(): + n_eval = math.ceil(len(eps) * cfg.dataset.eval_split) + train_episodes.extend(eps[: len(eps) - n_eval]) + eval_episodes.extend(eps[len(eps) - n_eval :]) + + if not train_episodes: + raise ValueError( + f"eval_split={cfg.dataset.eval_split} leaves 0 training episodes from {len(base_episodes)} total." + ) + + logging.info( + f"Train/eval split: {len(train_episodes)} train, {len(eval_episodes)} eval " + f"(eval_split={cfg.dataset.eval_split}, {len(task_to_episodes)} tasks)" + ) + + delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, full_dataset.meta) + + train_image_transforms = ( + ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None + ) + + train_dataset = LeRobotDataset( + cfg.dataset.repo_id, + root=cfg.dataset.root, + episodes=train_episodes, + delta_timestamps=delta_timestamps, + image_transforms=train_image_transforms, + revision=cfg.dataset.revision, + video_backend=cfg.dataset.video_backend, + return_uint8=True, + tolerance_s=cfg.tolerance_s, + ) + + eval_dataset = LeRobotDataset( + cfg.dataset.repo_id, + root=cfg.dataset.root, + episodes=eval_episodes, + delta_timestamps=delta_timestamps, + image_transforms=None, + revision=cfg.dataset.revision, + video_backend=cfg.dataset.video_backend, + return_uint8=True, + tolerance_s=cfg.tolerance_s, + ) + + if cfg.dataset.use_imagenet_stats: + for ds in (train_dataset, eval_dataset): + for key in ds.meta.camera_keys: + for stats_type, stats in IMAGENET_STATS.items(): + ds.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32) + + return train_dataset, eval_dataset diff --git a/src/lerobot/datasets/feature_utils.py b/src/lerobot/datasets/feature_utils.py index 56264408f..343b2fdcc 100644 --- a/src/lerobot/datasets/feature_utils.py +++ b/src/lerobot/datasets/feature_utils.py @@ -336,7 +336,7 @@ def validate_feature_image_or_video( Args: name (str): The name of the feature. - expected_shape (list[str]): The expected shape (C, H, W). + expected_shape (list[str]): The expected shape, e.g. (C, H, W) or (H, W, C). value: The image data to validate. Returns: diff --git a/src/lerobot/datasets/image_writer.py b/src/lerobot/datasets/image_writer.py index 8fb5804a5..41790b46a 100644 --- a/src/lerobot/datasets/image_writer.py +++ b/src/lerobot/datasets/image_writer.py @@ -41,11 +41,51 @@ def safe_stop_image_writer(func): return wrapper -def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image: - # TODO(aliberts): handle 1 channel and 4 for depth images - if image_array.ndim != 3: - raise ValueError(f"The array has {image_array.ndim} dimensions, but 3 is expected for an image.") +def squeeze_single_channel(array: np.ndarray) -> np.ndarray: + """Drop a leading or trailing singleton channel dim: ``(1, H, W)`` / ``(H, W, 1)`` -> ``(H, W)``. + Unlike ``array.squeeze()``, this only removes the channel axis, never an ``H`` or ``W`` of size 1. + """ + if array.ndim == 3: + if array.shape[0] == 1: + return array[0] + if array.shape[-1] == 1: + return array[..., 0] + return array + + +def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image: + """Convert a NumPy array to a PIL Image, preserving precision for grayscale. + + Behaviour by shape: + + - ``(H, W)`` or ``(1, H, W)`` / ``(H, W, 1)``: single-channel grayscale. + The native dtype is preserved using the matching PIL mode + (``I;16`` / ``F``). This is the path used for raw depth maps (no rescaling, clamping, or downcasting) + - ``(3, H, W)`` / ``(H, W, 3)``: RGB. Channels-first inputs are transposed + to channels-last. Float inputs in ``[0, 1]`` are scaled to ``uint8`` + (existing behaviour, gated by ``range_check``). + + Other shapes / channel counts raise ``NotImplementedError`` or + ``ValueError``. + """ + # TODO(CarolinePascal): 4 dimensions RGB-D images + if image_array.ndim not in (2, 3): + raise ValueError(f"The array has {image_array.ndim} dimensions, but 2 or 3 is expected for an image.") + + # Squeeze 3D single-channel inputs to 2D so depth maps work whether the + # caller emits (H, W), (1, H, W), or (H, W, 1). + image_array = squeeze_single_channel(image_array) + + if image_array.ndim == 2: + if image_array.dtype not in [np.uint16, np.float32]: + raise ValueError( + f"Unsupported single-channel image dtype: {image_array.dtype}. " + f"Supported dtypes: {sorted(str(d) for d in [np.uint16, np.float32])}." + ) + return PIL.Image.fromarray(np.ascontiguousarray(image_array)) + + # 3D path: must be RGB (3 channels), channels-first or channels-last. if image_array.shape[0] == 3: # Transpose from pytorch convention (C, H, W) to (H, W, C) image_array = image_array.transpose(1, 2, 0) @@ -71,13 +111,29 @@ def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) return PIL.Image.fromarray(image_array) +def save_kwargs_for_path(fpath: Path, compress_level: int) -> dict: + """Pick the right format-specific kwargs for :meth:`PIL.Image.Image.save`. + + PNG uses ``compress_level`` (0-9, zlib). TIFF uses ``compression`` (raw) for lossless raw depth maps. + """ + suffix = Path(fpath).suffix.lower() + if suffix == ".png": + return {"compress_level": compress_level} + if suffix in (".tif", ".tiff"): + return {"compression": "raw"} + else: + raise ValueError(f"Unsupported image file extension: {suffix}") + + def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1): """ Saves a NumPy array or PIL Image to a file. This function handles both NumPy arrays and PIL Image objects, converting the former to a PIL Image before saving. It includes error handling for - the save operation. + the save operation. The output format is inferred from the *fpath* + extension: ``.png`` → PNG with ``compress_level``, ``.tiff`` / ``.tif`` + → lossless raw depth maps (TIFF). Args: image (np.ndarray | PIL.Image.Image): The image data to save. @@ -101,7 +157,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level img = image else: raise TypeError(f"Unsupported image type: {type(image)}") - img.save(fpath, compress_level=compress_level) + img.save(fpath, **save_kwargs_for_path(fpath, compress_level)) except Exception as e: logger.error("Error writing image %s: %s", fpath, e) diff --git a/src/lerobot/datasets/io_utils.py b/src/lerobot/datasets/io_utils.py index b6344942c..868a114f5 100644 --- a/src/lerobot/datasets/io_utils.py +++ b/src/lerobot/datasets/io_utils.py @@ -154,7 +154,7 @@ def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]: Returns: dict: The statistics dictionary with values cast to numpy arrays. """ - stats = {key: np.array(value) for key, value in flatten_dict(stats).items()} + stats = {key: np.atleast_1d(np.array(value)) for key, value in flatten_dict(stats).items()} return unflatten_dict(stats) @@ -226,28 +226,50 @@ def load_image_as_numpy( Args: fpath (str | Path): Path to the image file. dtype (np.dtype): The desired data type of the output array. If floating, - pixels are scaled to [0, 1]. + pixels are scaled to [0, 1]. Only used for RGB images. channel_first (bool): If True, converts the image to (C, H, W) format. Otherwise, it remains in (H, W, C) format. Returns: np.ndarray: The image as a numpy array. """ - img = PILImage.open(fpath).convert("RGB") - img_array = np.array(img, dtype=dtype) + is_depth = fpath.endswith(".tiff") or fpath.endswith(".tif") + if is_depth: + # Preserve the native depth dtype (uint16 -> "I;16", float32 -> "F"). + img = PILImage.open(fpath) + img_array = np.array(img) + else: + img = PILImage.open(fpath).convert("RGB") + img_array = np.array(img, dtype=dtype) + if np.issubdtype(dtype, np.floating): + img_array /= 255.0 if channel_first: # (H, W, C) -> (C, H, W) - img_array = np.transpose(img_array, (2, 0, 1)) - if np.issubdtype(dtype, np.floating): - img_array /= 255.0 + img_array = img_array[np.newaxis, ...] if img_array.ndim == 2 else np.transpose(img_array, (2, 0, 1)) return img_array +# PIL modes for 16-bit unsigned depth maps. +UINT16_PIL_MODES = {"I;16", "I;16B", "I;16L"} + + +def pil_to_chw_tensor(img: PILImage.Image) -> torch.Tensor: + """Convert a PIL image to a channel-first tensor. + + ``uint16`` depth maps become ``float32 (1, H, W)`` in native units (``ToTensor`` + would overflow them to ``int16``); all other modes use the standard ``ToTensor`` path. + """ + if img.mode in UINT16_PIL_MODES: + return torch.from_numpy(np.array(img, dtype=np.float32))[None, ...] + return transforms.ToTensor()(img) + + def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]: """Convert a batch from a Hugging Face dataset to torch tensors. This transform function converts items from Hugging Face dataset format (pyarrow) - to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8) - to a torch image representation (C, H, W, float32) in the range [0, 1]. Other + to torch tensors. RGB images are converted from PIL objects (H, W, C, uint8) + to a torch image representation (C, H, W, float32) in the range [0, 1]. Depth + maps are returned as float32 (1, H, W) in their native units. Other types are converted to torch.tensor. Args: @@ -262,8 +284,7 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to continue first_item = items_dict[key][0] if isinstance(first_item, PILImage.Image): - to_tensor = transforms.ToTensor() - items_dict[key] = [to_tensor(img) for img in items_dict[key]] + items_dict[key] = [pil_to_chw_tensor(img) for img in items_dict[key]] elif first_item is None or isinstance(first_item, dict): pass else: @@ -329,7 +350,11 @@ def item_to_torch(item: dict) -> dict: """ skip_keys = {"task", *LANGUAGE_COLUMNS} for key, val in item.items(): - if isinstance(val, (np.ndarray | list)) and key not in skip_keys: + if key in skip_keys: + continue + if isinstance(val, PILImage.Image): + item[key] = pil_to_chw_tensor(val) + elif isinstance(val, (np.ndarray | list)): # Convert numpy arrays and lists to torch tensors item[key] = torch.tensor(val) return item diff --git a/src/lerobot/datasets/lerobot_dataset.py b/src/lerobot/datasets/lerobot_dataset.py index d0dcf087d..aba86efe3 100644 --- a/src/lerobot/datasets/lerobot_dataset.py +++ b/src/lerobot/datasets/lerobot_dataset.py @@ -24,7 +24,7 @@ import torch.utils from huggingface_hub import HfApi, snapshot_download from huggingface_hub.errors import RevisionNotFoundError -from lerobot.configs import VideoEncoderConfig +from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig, RGBEncoderConfig from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata @@ -58,8 +58,10 @@ class LeRobotDataset(torch.utils.data.Dataset): download_videos: bool = True, video_backend: str | None = None, return_uint8: bool = False, + depth_output_unit: str = DEFAULT_DEPTH_UNIT, batch_encoding_size: int = 1, - camera_encoder: VideoEncoderConfig | None = None, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, encoder_threads: int | None = None, streaming_encoding: bool = False, encoder_queue_maxsize: int = 30, @@ -183,8 +185,11 @@ class LeRobotDataset(torch.utils.data.Dataset): You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision. batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos. Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1. - camera_encoder (VideoEncoderConfig | None, optional): Video encoder settings for cameras - (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` + rgb_encoder (RGBEncoderConfig | None, optional): Video encoder settings for cameras + (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` + is used by the writer. + depth_encoder (DepthEncoderConfig | None, optional): Video encoder settings for depth cameras + (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used by the writer. encoder_threads (int | None, optional): Number of encoder threads (global). ``None`` lets the codec decide. @@ -201,13 +206,12 @@ class LeRobotDataset(torch.utils.data.Dataset): super().__init__() self.repo_id = repo_id self._requested_root = Path(root) if root else None - self.reader = None - self.set_image_transforms(image_transforms) self.delta_timestamps = delta_timestamps self.tolerance_s = tolerance_s self.revision = revision if revision else CODEBASE_VERSION self._video_backend = video_backend if video_backend else get_safe_default_video_backend() self._return_uint8 = return_uint8 + self._depth_output_unit = depth_output_unit self._batch_encoding_size = batch_encoding_size self._encoder_threads = encoder_threads @@ -220,6 +224,7 @@ class LeRobotDataset(torch.utils.data.Dataset): ) self.root = self.meta.root self.revision = self.meta.revision + self.meta.rescale_depth_stats(self._depth_output_unit) if episodes is not None and any( episode >= self.meta.total_episodes or episode < 0 for episode in episodes @@ -248,7 +253,9 @@ class LeRobotDataset(torch.utils.data.Dataset): delta_timestamps=delta_timestamps, image_transforms=image_transforms, return_uint8=self._return_uint8, + depth_output_unit=self._depth_output_unit, ) + self.image_transforms = image_transforms # Load actual data if force_cache_sync or not self.reader.try_load(): @@ -272,14 +279,16 @@ class LeRobotDataset(torch.utils.data.Dataset): if streaming_encoding and len(self.meta.video_keys) > 0: streaming_enc = self._build_streaming_encoder( self.meta.fps, - camera_encoder, + rgb_encoder, + depth_encoder, encoder_queue_maxsize, encoder_threads, ) self.writer = DatasetWriter( meta=self.meta, root=self.root, - camera_encoder=camera_encoder, + rgb_encoder=rgb_encoder, + depth_encoder=depth_encoder, encoder_threads=encoder_threads, batch_encoding_size=batch_encoding_size, streaming_encoder=streaming_enc, @@ -315,19 +324,22 @@ class LeRobotDataset(torch.utils.data.Dataset): delta_timestamps=self.delta_timestamps, image_transforms=self.image_transforms, return_uint8=self._return_uint8, + depth_output_unit=self._depth_output_unit, ) return self.reader @staticmethod def _build_streaming_encoder( fps: int, - camera_encoder: VideoEncoderConfig | None, + rgb_encoder: RGBEncoderConfig | None, + depth_encoder: DepthEncoderConfig | None, encoder_queue_maxsize: int, encoder_threads: int | None, ) -> StreamingVideoEncoder: return StreamingVideoEncoder( fps=fps, - camera_encoder=camera_encoder, + rgb_encoder=rgb_encoder, + depth_encoder=depth_encoder, queue_maxsize=encoder_queue_maxsize, encoder_threads=encoder_threads, ) @@ -339,6 +351,11 @@ class LeRobotDataset(torch.utils.data.Dataset): """Frames per second used during data collection.""" return self.meta.fps + @property + def depth_output_unit(self) -> str: + """Physical unit (``"m"`` or ``"mm"``) depth maps and statistics are returned in on read.""" + return self._depth_output_unit + @property def num_frames(self) -> int: """Number of frames in selected episodes.""" @@ -370,6 +387,18 @@ class LeRobotDataset(torch.utils.data.Dataset): self.reader.load_and_activate() return self.reader.hf_dataset + @property + def absolute_to_relative_idx(self) -> dict[int, int] | None: + """Mapping from absolute frame indices to HF dataset row positions. + + Non-None only for episode-filtered datasets where absolute indices + (from metadata) differ from row positions in the loaded HF dataset. + """ + reader = self._ensure_reader() + if reader.hf_dataset is None: + reader.load_and_activate() + return reader._absolute_to_relative_idx + # ── Writer-delegated methods ────────────────────────────────────── def add_frame(self, frame: dict) -> None: @@ -505,15 +534,14 @@ class LeRobotDataset(torch.utils.data.Dataset): def set_image_transforms(self, image_transforms: Callable | None) -> None: """Replace the transform applied to visual observations.""" - if image_transforms is not None and not callable(image_transforms): - raise TypeError("image_transforms must be callable or None.") + self._ensure_reader().set_image_transforms(image_transforms) self.image_transforms = image_transforms - if self.reader is not None: - self.reader._image_transforms = image_transforms def clear_image_transforms(self) -> None: """Remove the transform applied to visual observations.""" - self.set_image_transforms(None) + if self.reader is not None: + self.reader.set_image_transforms(None) + self.image_transforms = None # ── Hub methods (stay on facade) ────────────────────────────────── @@ -645,7 +673,8 @@ class LeRobotDataset(torch.utils.data.Dataset): image_writer_threads: int = 0, video_backend: str | None = None, batch_encoding_size: int = 1, - camera_encoder: VideoEncoderConfig | None = None, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, metadata_buffer_size: int = 10, streaming_encoding: bool = False, encoder_queue_maxsize: int = 30, @@ -676,8 +705,10 @@ class LeRobotDataset(torch.utils.data.Dataset): video_backend: Video decoding backend (used when reading back). batch_encoding_size: Number of episodes to accumulate before batch-encoding videos. ``1`` means encode immediately. - camera_encoder: Video encoder settings for cameras (codec, quality, etc.). - When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used. + rgb_encoder: Video encoder settings for cameras (codec, quality, etc.). + When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used. + depth_encoder: Video encoder settings for depth cameras (codec, quality, etc.). + When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used. encoder_threads: Number of encoder threads (global). ``None`` lets the codec decide. metadata_buffer_size: Number of episode metadata records to buffer @@ -712,6 +743,7 @@ class LeRobotDataset(torch.utils.data.Dataset): obj.episodes = None obj._video_backend = video_backend if video_backend is not None else get_safe_default_video_backend() obj._return_uint8 = False + obj._depth_output_unit = DEFAULT_DEPTH_UNIT obj._batch_encoding_size = batch_encoding_size obj._encoder_threads = encoder_threads @@ -721,12 +753,13 @@ class LeRobotDataset(torch.utils.data.Dataset): streaming_enc = None if streaming_encoding and len(obj.meta.video_keys) > 0: streaming_enc = cls._build_streaming_encoder( - fps, camera_encoder, encoder_queue_maxsize, encoder_threads + fps, rgb_encoder, depth_encoder, encoder_queue_maxsize, encoder_threads ) obj.writer = DatasetWriter( meta=obj.meta, root=obj.root, - camera_encoder=camera_encoder, + rgb_encoder=rgb_encoder, + depth_encoder=depth_encoder, encoder_threads=encoder_threads, batch_encoding_size=batch_encoding_size, streaming_encoder=streaming_enc, @@ -749,7 +782,8 @@ class LeRobotDataset(torch.utils.data.Dataset): force_cache_sync: bool = False, video_backend: str | None = None, batch_encoding_size: int = 1, - camera_encoder: VideoEncoderConfig | None = None, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, encoder_threads: int | None = None, image_writer_processes: int = 0, image_writer_threads: int = 0, @@ -777,8 +811,10 @@ class LeRobotDataset(torch.utils.data.Dataset): video_backend: Video decoding backend for reading back data. batch_encoding_size: Number of episodes to accumulate before batch-encoding videos. - camera_encoder: Video encoder settings for cameras (codec, quality, etc.). - When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used. + rgb_encoder: Video encoder settings for cameras (codec, quality, etc.). + When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used. + depth_encoder: Video encoder settings for depth cameras (codec, quality, etc.). + When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used. encoder_threads: Number of encoder threads (global). ``None`` lets the codec decide. image_writer_processes: Subprocesses for async image writing. @@ -806,6 +842,7 @@ class LeRobotDataset(torch.utils.data.Dataset): obj.episodes = None obj._video_backend = video_backend if video_backend else get_safe_default_video_backend() obj._return_uint8 = False + obj._depth_output_unit = DEFAULT_DEPTH_UNIT obj._batch_encoding_size = batch_encoding_size if obj._requested_root is not None: @@ -825,12 +862,13 @@ class LeRobotDataset(torch.utils.data.Dataset): streaming_enc = None if streaming_encoding and len(obj.meta.video_keys) > 0: streaming_enc = cls._build_streaming_encoder( - obj.meta.fps, camera_encoder, encoder_queue_maxsize, encoder_threads + obj.meta.fps, rgb_encoder, depth_encoder, encoder_queue_maxsize, encoder_threads ) obj.writer = DatasetWriter( meta=obj.meta, root=obj.root, - camera_encoder=camera_encoder, + rgb_encoder=rgb_encoder, + depth_encoder=depth_encoder, encoder_threads=encoder_threads, batch_encoding_size=batch_encoding_size, streaming_encoder=streaming_enc, diff --git a/src/lerobot/datasets/pyav_utils.py b/src/lerobot/datasets/pyav_utils.py index d291f8b40..7b7d1e5de 100644 --- a/src/lerobot/datasets/pyav_utils.py +++ b/src/lerobot/datasets/pyav_utils.py @@ -24,6 +24,7 @@ import logging from typing import Any import av +import numpy as np logger = logging.getLogger(__name__) @@ -31,6 +32,34 @@ FFMPEG_NUMERIC_OPTION_TYPES = ("INT", "INT64", "UINT64", "FLOAT", "DOUBLE") FFMPEG_INTEGER_OPTION_TYPES = ("INT", "INT64", "UINT64") +def write_u16_plane(plane: av.video.plane.VideoPlane, src: np.ndarray, fill_value: int | None = None) -> None: + """Copy a 2D ``uint16`` image into the plane's memory buffer, row by row. + + For speed, each row is padded to a wider size than ``width``, so the true row width in + memory is ``plane.line_size`` (bytes), not ``width``. Copying as one straight stream + would skew the image, so we write only the first ``width`` columns of each row and + leave the padding untouched. + + Args: + plane: Destination 16-bit plane. + src: Source image, shape ``(height, width)``, dtype ``uint16``. + fill_value: If given, every pixel (padding included) is set to this first, so the + padding holds clean data instead of garbage. + """ + height, width = src.shape + stride_u16 = plane.line_size // np.dtype(np.uint16).itemsize + dst = np.frombuffer(plane, dtype=np.uint16).reshape(height, stride_u16) + if fill_value is not None: + dst.fill(fill_value) + dst[:, :width] = src + + +@functools.cache +def get_pix_fmt_channels(pix_fmt: str) -> int: + """Return the number of components (channels) for *pix_fmt*.""" + return len(av.VideoFormat(pix_fmt).components) + + @functools.cache def get_codec(vcodec: str) -> av.codec.Codec | None: """PyAV write-mode ``Codec`` for *vcodec*, or ``None`` if unavailable.""" @@ -92,7 +121,7 @@ def _check_option_value(vcodec: str, label: str, value: Any, opt: av.option.Opti f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option." ) from e elif isinstance(value, (float, int)): - num_val = value + num_val = float(value) else: raise ValueError( f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option." @@ -142,6 +171,16 @@ def _check_pixel_format(vcodec: str, pix_fmt: str) -> None: ) +def _check_pix_fmt_channels(pix_fmt: str, channels: int) -> None: + """Ensure *pix_fmt* can carry at least *channels* components.""" + pix_fmt_channels = get_pix_fmt_channels(pix_fmt) + if pix_fmt_channels < channels: + raise ValueError( + f"pix_fmt={pix_fmt!r} carries only {pix_fmt_channels} component(s) " + f"but the source data has {channels} channel(s)." + ) + + def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None: """Validate merged encoder options (typed) against the codec's published AVOptions.""" supported_options = _get_codec_options_by_name(vcodec) @@ -156,12 +195,18 @@ def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None: _check_option_value(vcodec, key, value, supported_options[key]) -def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options: dict[str, Any]) -> None: +def check_video_encoder_parameters_pyav( + vcodec: str, + pix_fmt: str, + codec_options: dict[str, Any], + channels: int | None = None, +) -> None: """Verify *config* is compatible with the bundled FFmpeg build. Checks pixel format, abstract tuning-field compatibility, and each merged encoder option from :meth:`~lerobot.configs.video.VideoEncoderConfig.get_codec_options` against PyAV (including numeric ``extra_options`` present in that dict). + When given, additionally verify that *pix_fmt* carries as many components as the source data channels. No-op when ``config.vcodec`` isn't in the local FFmpeg build. Raises: @@ -171,4 +216,6 @@ def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options if not options: raise ValueError(f"Codec {vcodec!r} is not available in the bundled FFmpeg build") _check_pixel_format(vcodec, pix_fmt) + if channels is not None: + _check_pix_fmt_channels(pix_fmt, channels) _check_codec_options(vcodec, codec_options) diff --git a/src/lerobot/datasets/sampler.py b/src/lerobot/datasets/sampler.py index af85dff9b..aee6ce46d 100644 --- a/src/lerobot/datasets/sampler.py +++ b/src/lerobot/datasets/sampler.py @@ -53,6 +53,7 @@ class EpisodeAwareSampler: drop_n_last_frames: int = 0, shuffle: bool = False, seed: int = 0, + absolute_to_relative_idx: dict[int, int] | None = None, ): """ Args: @@ -107,6 +108,7 @@ class EpisodeAwareSampler: self.seed = seed self._epoch = 0 self._start_index = 0 + self._absolute_to_relative = absolute_to_relative_idx @property def indices(self) -> list[int]: @@ -132,7 +134,10 @@ class EpisodeAwareSampler: def _frame_index(self, position: int) -> int: episode = int(np.searchsorted(self._cum_lengths, position, side="right")) position_in_episode = position - (int(self._cum_lengths[episode - 1]) if episode > 0 else 0) - return int(self._starts[episode]) + position_in_episode + absolute_idx = int(self._starts[episode]) + position_in_episode + if self._absolute_to_relative is not None: + return self._absolute_to_relative[absolute_idx] + return absolute_idx def __iter__(self) -> Iterator[int]: # Advance epoch state eagerly, not on first consumption of the generator. diff --git a/src/lerobot/datasets/streaming_dataset.py b/src/lerobot/datasets/streaming_dataset.py index 3c1e4a73c..14d4a52a4 100644 --- a/src/lerobot/datasets/streaming_dataset.py +++ b/src/lerobot/datasets/streaming_dataset.py @@ -22,9 +22,11 @@ import numpy as np import torch from datasets import load_dataset +from lerobot.configs import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DepthEncoderConfig from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata +from .depth_utils import MM_PER_METRE, dequantize_depth from .feature_utils import get_delta_indices from .io_utils import item_to_torch from .utils import ( @@ -35,6 +37,7 @@ from .utils import ( ) from .video_utils import ( VideoDecoderCache, + decode_video_frames, decode_video_frames_torchcodec, ) @@ -252,6 +255,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): rng: np.random.Generator | None = None, shuffle: bool = True, return_uint8: bool = False, + depth_output_unit: str = DEFAULT_DEPTH_UNIT, ): """Initialize a StreamingLeRobotDataset. @@ -272,6 +276,8 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): seed (int, optional): Reproducibility random seed. rng (np.random.Generator | None, optional): Random number generator. shuffle (bool, optional): Whether to shuffle the dataset across exhaustions. Defaults to True. + depth_output_unit (str, optional): Physical unit depth maps are dequantized to ("m" or "mm"). + Defaults to "mm". """ super().__init__() self.repo_id = repo_id @@ -290,6 +296,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): self.streaming = streaming self.buffer_size = buffer_size self._return_uint8 = return_uint8 + self._depth_output_unit = depth_output_unit # We cache the video decoders to avoid re-initializing them at each frame (avoiding a ~10x slowdown) self.video_decoder_cache = None @@ -303,9 +310,22 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): ) self.root = self.meta.root self.revision = self.meta.revision + self.meta.rescale_depth_stats(self._depth_output_unit) # Check version check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION) + self._depth_encoder_configs: dict[str, DepthEncoderConfig] = { + vid_key: DepthEncoderConfig.from_video_info(self.meta.features[vid_key].get("info")) + for vid_key in self.meta.depth_keys + } + + # Input unit of each depth feature stored as raw images (dequantized separately from videos). + self._image_depth_units: dict[str, str | None] = { + key: (self.meta.features[key].get("info") or {}).get("depth_unit") + for key in self.meta.depth_keys + if key in self.meta.image_keys + } + self.delta_timestamps = None self.delta_indices = None @@ -336,6 +356,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): def fps(self): return self.meta.fps + @property + def depth_output_unit(self) -> str: + """Physical unit (``"m"`` or ``"mm"``) depth maps are returned in on read.""" + return self._depth_output_unit + @staticmethod def _iter_random_indices( rng: np.random.Generator, buffer_size: int, random_batch_size=100 @@ -518,6 +543,15 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): for update in updates: result.update(update) + # Convert raw-image depth features to the output unit (video depth is already converted). + for key, stored_unit in self._image_depth_units.items(): + if key in result and stored_unit is not None and stored_unit != self._depth_output_unit: + result[key] = ( + result[key] * MM_PER_METRE + if stored_unit == DEPTH_METER_UNIT + else result[key] / MM_PER_METRE + ) + result["task"] = self.meta.tasks.iloc[item["task_index"]].name yield result @@ -554,13 +588,34 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset): for video_key, query_ts in query_timestamps.items(): root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, video_key)}" - frames = decode_video_frames_torchcodec( - video_path, - query_ts, - self.tolerance_s, - decoder_cache=self.video_decoder_cache, - return_uint8=self._return_uint8, - ) + if video_key in self.meta.depth_keys: + # Depth maps are 12-bit quantized and only decodable via pyav; dequantize back + # to physical units to match the non-streaming reader. + frames = decode_video_frames( + video_path, + query_ts, + self.tolerance_s, + backend="pyav", + return_uint8=False, + is_depth=True, + ) + depth_encoder = self._depth_encoder_configs[video_key] + frames = dequantize_depth( + frames, + depth_min=depth_encoder.depth_min, + depth_max=depth_encoder.depth_max, + shift=depth_encoder.shift, + use_log=depth_encoder.use_log, + output_unit=self._depth_output_unit, + ) + else: + frames = decode_video_frames_torchcodec( + video_path, + query_ts, + self.tolerance_s, + decoder_cache=self.video_decoder_cache, + return_uint8=self._return_uint8, + ) item[video_key] = frames.squeeze(0) if len(query_ts) == 1 else frames diff --git a/src/lerobot/datasets/utils.py b/src/lerobot/datasets/utils.py index de91978ea..d30761515 100644 --- a/src/lerobot/datasets/utils.py +++ b/src/lerobot/datasets/utils.py @@ -87,11 +87,14 @@ DATA_DIR = "data" VIDEO_DIR = "videos" CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}" +IMAGE_FILE_PATTERN = "frame-{frame_index:06d}.png" +DEPTH_FILE_PATTERN = "frame-{frame_index:06d}.tiff" DEFAULT_TASKS_PATH = "meta/tasks.parquet" DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet" DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet" DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4" -DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png" +DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/" + IMAGE_FILE_PATTERN +DEFAULT_DEPTH_PATH = "images/{image_key}/episode-{episode_index:06d}/" + DEPTH_FILE_PATTERN LEGACY_EPISODES_PATH = "meta/episodes.jsonl" LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl" diff --git a/src/lerobot/datasets/video_utils.py b/src/lerobot/datasets/video_utils.py index ca90fba45..ef3005dd8 100644 --- a/src/lerobot/datasets/video_utils.py +++ b/src/lerobot/datasets/video_utils.py @@ -39,11 +39,17 @@ from datasets.features.features import register_feature from PIL import Image from lerobot.configs import ( + DepthEncoderConfig, + RGBEncoderConfig, VideoEncoderConfig, - camera_encoder_defaults, + depth_encoder_defaults, + rgb_encoder_defaults, ) from lerobot.utils.import_utils import get_safe_default_video_backend +from .depth_utils import quantize_depth +from .pyav_utils import get_pix_fmt_channels + logger = logging.getLogger(__name__) @@ -53,6 +59,7 @@ def decode_video_frames( tolerance_s: float, backend: str | None = None, return_uint8: bool = False, + is_depth: bool = False, ) -> torch.Tensor: """ Decodes video frames using the specified backend. @@ -64,23 +71,35 @@ def decode_video_frames( backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav". The legacy value "video_reader" is accepted for one release as an alias for "pyav" and will be removed in a future version. - return_uint8 (bool): If True, return raw uint8 frames without float32 normalization. + return_uint8 (bool): For RGB videos, if True return raw uint8 frames without float32 normalization. This reduces memory for DataLoader IPC; normalization can be done on GPU afterward. + is_depth (bool): Set to True if the video is a depth map (1 channel, uint12). Returns: - torch.Tensor: Decoded frames (float32 in [0,1] by default, or uint8 if return_uint8=True). + torch.Tensor: Decoded frames (RGB: float32 in [0,1] by default, or uint8 if return_uint8=True, Depth: uint12). Currently supports torchcodec on cpu and pyav. """ + if backend != "pyav" and is_depth: + logger.debug("Decoding depth maps is only supported with the 'pyav' backend, falling back to pyav.") + # We do not actually return uint8 here, but we avoid the 255 normalization step. + return decode_video_frames_pyav( + video_path, timestamps, tolerance_s, return_uint8=False, is_depth=True + ) + if backend is None: backend = get_safe_default_video_backend() if backend == "torchcodec": return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8) elif backend == "pyav": - return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8) + return decode_video_frames_pyav( + video_path, timestamps, tolerance_s, return_uint8=return_uint8, is_depth=is_depth + ) elif backend == "video_reader": logger.warning("backend='video_reader' is deprecated and now aliases to 'pyav'.") - return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8) + return decode_video_frames_pyav( + video_path, timestamps, tolerance_s, return_uint8=return_uint8, is_depth=is_depth + ) else: raise ValueError(f"Unsupported video backend: {backend}") @@ -91,6 +110,7 @@ def decode_video_frames_pyav( tolerance_s: float, log_loaded_timestamps: bool = False, return_uint8: bool = False, + is_depth: bool = False, ) -> torch.Tensor: """Loads frames associated to the requested timestamps of a video using PyAV. @@ -109,8 +129,9 @@ def decode_video_frames_pyav( tolerance_s: Allowed deviation in seconds between a queried timestamp and the closest decoded frame. log_loaded_timestamps: When True, log every decoded frame's timestamp at INFO level. - return_uint8: When True, return raw uint8 frames (C, H, W). Otherwise, return float32 in - [0, 1] range. + return_uint8: For RGB videos, if True return raw uint8 frames (C, H, W). + Otherwise, return float32 in [0, 1] range. + is_depth: Set to True if the video is a depth map (1 channel, uint12). Returns: torch.Tensor of shape (len(timestamps), C, H, W). @@ -132,7 +153,13 @@ def decode_video_frames_pyav( # https://pyav.basswood-io.com/docs/stable/api/container.html#av.container.InputContainer.seek with av.open(video_path) as container: stream = container.streams.video[0] - container.seek(int(first_ts * av.time_base), backward=True) + # Seek to the nearest keyframe at or before `first_ts` with a 1 frame margin + container.seek( + round(first_ts / stream.time_base) - 1, + backward=True, + any_frame=False, + stream=stream, + ) for frame in container.decode(stream): if frame.pts is None: @@ -140,9 +167,13 @@ def decode_video_frames_pyav( current_ts = float(frame.pts * stream.time_base) if log_loaded_timestamps: logger.info(f"frame loaded at timestamp={current_ts:.4f}") - # Convert to CHW uint8 to match torchcodec's output layout. - arr = frame.to_ndarray(format="rgb24") # H, W, 3 - loaded_frames.append(torch.from_numpy(arr).permute(2, 0, 1).contiguous()) + if is_depth: + arr = frame.to_ndarray(format="gray12le") # (H, W) uint12 + loaded_frames.append(torch.from_numpy(arr).unsqueeze(0).contiguous()) + else: + arr = frame.to_ndarray(format="rgb24") # (H, W, 3) + # Convert to CHW uint8 to match torchcodec's output layout. + loaded_frames.append(torch.from_numpy(arr).permute(2, 0, 1).contiguous()) loaded_ts.append(current_ts) if current_ts >= last_ts: break @@ -185,7 +216,7 @@ def decode_video_frames_pyav( f"number of queried timestamps ({len(timestamps)})" ) - if return_uint8: + if return_uint8 or is_depth: return closest_frames # convert to the pytorch format which is float32 in [0,1] range (and channel first) @@ -406,17 +437,38 @@ def encode_video_frames( imgs_dir: Path | str, video_path: Path | str, fps: int, - camera_encoder: VideoEncoderConfig | None = None, + video_encoder: VideoEncoderConfig | None = None, encoder_threads: int | None = None, *, log_level: int | None = av.logging.WARNING, overwrite: bool = False, ) -> None: - """More info on ffmpeg arguments tuning on `benchmark/video/README.md`""" - if camera_encoder is None: - camera_encoder = camera_encoder_defaults() - vcodec = camera_encoder.vcodec - pix_fmt = camera_encoder.pix_fmt + """Encode a directory of image frames into an MP4 video. + + When ``video_encoder`` is a :class:`~lerobot.configs.video.DepthEncoderConfig`, + frames are read from ``.tiff`` files and quantized to 12-bit depth codes using the + encoder's ``depth_min`` / ``depth_max`` / ``shift`` / ``use_log``; otherwise ``.png`` + RGB frames are encoded directly. + + Args: + imgs_dir: Directory containing the frames to encode, named ``frame-000000`` + onwards (``.png`` for RGB, ``.tiff`` for depth). + video_path: Output path for the encoded ``.mp4`` file. + fps: Frame rate of the output video. + video_encoder: Encoder settings (codec, pixel format, quality, ...). When + ``None``, :func:`rgb_encoder_defaults` is used. Pass a + :class:`~lerobot.configs.video.DepthEncoderConfig` to encode depth frames. + encoder_threads: Per-encoder thread count forwarded to the codec. ``None`` + lets the codec decide. + log_level: libav log level to set while encoding, or ``None`` to leave the + current logging configuration unchanged. + overwrite: When ``False`` and ``video_path`` already exists, skip encoding and + log a warning. When ``True``, re-encode and replace the existing file. + """ + if video_encoder is None: + video_encoder = rgb_encoder_defaults() + vcodec = video_encoder.vcodec + pix_fmt = video_encoder.pix_fmt video_path = Path(video_path) imgs_dir = Path(imgs_dir) @@ -428,17 +480,19 @@ def encode_video_frames( video_path.parent.mkdir(parents=True, exist_ok=True) # Get input frames - template = "frame-" + ("[0-9]" * 6) + ".png" + is_depth = isinstance(video_encoder, DepthEncoderConfig) + suffix = ".png" if not is_depth else ".tiff" + template = "frame-" + ("[0-9]" * 6) + suffix input_list = sorted( glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("-")[-1].split(".")[0]) ) if len(input_list) == 0: - raise FileNotFoundError(f"No images found in {imgs_dir}.") + raise FileNotFoundError(f"No images with suffix {suffix} found in {imgs_dir}.") with Image.open(input_list[0]) as dummy_image: width, height = dummy_image.size - video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True) + video_options = video_encoder.get_codec_options(encoder_threads, as_strings=True) # Set logging level if log_level is not None: @@ -455,8 +509,19 @@ def encode_video_frames( # Loop through input frames and encode them for input_data in input_list: with Image.open(input_data) as input_image: - input_image = input_image.convert("RGB") - input_frame = av.VideoFrame.from_image(input_image) + if is_depth: + input_frame = quantize_depth( + np.array(input_image), + depth_min=video_encoder.depth_min, + depth_max=video_encoder.depth_max, + shift=video_encoder.shift, + use_log=video_encoder.use_log, + pix_fmt=video_encoder.pix_fmt, + video_backend="pyav", + ) + else: + input_image = input_image.convert("RGB") + input_frame = av.VideoFrame.from_image(input_image) packet = output_stream.encode(input_frame) if packet: output.mux(packet) @@ -477,7 +542,7 @@ def encode_video_frames( def reencode_video( input_video_path: Path | str, output_video_path: Path | str, - camera_encoder: VideoEncoderConfig | None = None, + video_encoder: VideoEncoderConfig | None = None, encoder_threads: int | None = None, log_level: int | None = av.logging.WARNING, overwrite: bool = False, @@ -489,7 +554,7 @@ def reencode_video( Args: input_video_path: Existing video file to read. output_video_path: Path for the re-encoded file. - camera_encoder: Encoder configuration. Defaults to :func:`camera_encoder_defaults`. + video_encoder: Encoder configuration. Defaults to :func:`rgb_encoder_defaults`. encoder_threads: Optional thread count forwarded to :meth:`VideoEncoderConfig.get_codec_options`. log_level: libav log level while encoding, or ``None`` to leave logging unchanged. Defaults to WARNING. overwrite: When ``False`` and ``output_video_path`` already exists, skip and log a warning. @@ -497,7 +562,7 @@ def reencode_video( end_time_s: When set, trim the output to end at this timestamp (seconds, exclusive). """ - camera_encoder = camera_encoder or camera_encoder_defaults() + video_encoder = video_encoder or rgb_encoder_defaults() if (start_time_s is not None and start_time_s < 0) or (end_time_s is not None and end_time_s < 0): raise ValueError(f"Trim times must be non-negative, got start={start_time_s}, end={end_time_s}.") @@ -512,9 +577,9 @@ def reencode_video( output_video_path.parent.mkdir(parents=True, exist_ok=True) - video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True) - vcodec = camera_encoder.vcodec - pix_fmt = camera_encoder.pix_fmt + video_options = video_encoder.get_codec_options(encoder_threads, as_strings=True) + vcodec = video_encoder.vcodec + pix_fmt = video_encoder.pix_fmt with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file: tmp_output_video_path = tmp_named_file.name @@ -696,22 +761,21 @@ class _CameraEncoderThread(threading.Thread): self, video_path: Path, fps: int, - vcodec: str, - pix_fmt: str, - codec_options: dict[str, str], + video_encoder: VideoEncoderConfig, frame_queue: queue.Queue, result_queue: queue.Queue, stop_event: threading.Event, + encoder_threads: int | None = None, ): super().__init__(daemon=True) self.video_path = video_path self.fps = fps - self.vcodec = vcodec - self.pix_fmt = pix_fmt - self.codec_options = codec_options + self.video_encoder = video_encoder + self.is_depth = isinstance(video_encoder, DepthEncoderConfig) self.frame_queue = frame_queue self.result_queue = result_queue self.stop_event = stop_event + self.encoder_threads = encoder_threads def run(self) -> None: from .compute_stats import RunningQuantileStats, auto_downsample_height_width @@ -736,12 +800,12 @@ class _CameraEncoderThread(threading.Thread): # Sentinel: flush and close break - # Ensure HWC uint8 numpy array + # Ensure HWC (RGB or depth) uint8 (RGB only) numpy array if isinstance(frame_data, np.ndarray): - if frame_data.ndim == 3 and frame_data.shape[0] == 3: + if frame_data.ndim == 3 and frame_data.shape[0] in (1, 3): # CHW -> HWC frame_data = frame_data.transpose(1, 2, 0) - if frame_data.dtype != np.uint8: + if not self.is_depth and frame_data.dtype != np.uint8: frame_data = (frame_data * 255).astype(np.uint8) # Open container on first frame (to get width/height) @@ -749,15 +813,29 @@ class _CameraEncoderThread(threading.Thread): height, width = frame_data.shape[:2] Path(self.video_path).parent.mkdir(parents=True, exist_ok=True) container = av.open(str(self.video_path), "w") - output_stream = container.add_stream(self.vcodec, self.fps, options=self.codec_options) - output_stream.pix_fmt = self.pix_fmt + output_stream = container.add_stream( + self.video_encoder.vcodec, + self.fps, + options=self.video_encoder.get_codec_options(self.encoder_threads, as_strings=True), + ) + output_stream.pix_fmt = self.video_encoder.pix_fmt output_stream.width = width output_stream.height = height output_stream.time_base = Fraction(1, self.fps) # Encode frame with explicit timestamps - pil_img = Image.fromarray(frame_data) - video_frame = av.VideoFrame.from_image(pil_img) + if not self.is_depth: + pil_img = Image.fromarray(frame_data) + video_frame = av.VideoFrame.from_image(pil_img) + else: + video_frame = quantize_depth( + frame_data, + depth_min=self.video_encoder.depth_min, + depth_max=self.video_encoder.depth_max, + shift=self.video_encoder.shift, + use_log=self.video_encoder.use_log, + video_backend=self.video_encoder.video_backend, + ) video_frame.pts = frame_count video_frame.time_base = Fraction(1, self.fps) packet = output_stream.encode(video_frame) @@ -815,22 +893,27 @@ class StreamingVideoEncoder: def __init__( self, fps: int, - camera_encoder: VideoEncoderConfig | None = None, + rgb_encoder: RGBEncoderConfig | None = None, + depth_encoder: DepthEncoderConfig | None = None, queue_maxsize: int = 30, encoder_threads: int | None = None, ): """ Args: fps: Frames per second for the output videos. - camera_encoder: Video encoder settings applied to all cameras. - When ``None``, :func:`camera_encoder_defaults` is used. - encoder_threads: Number of encoder threads (global setting). - ``None`` lets the codec decide. + rgb_encoder: Video encoder settings applied to all RGB cameras. + When ``None``, :func:`rgb_encoder_defaults` is used. + depth_encoder: Video encoder settings applied to all depth cameras, + including the depth quantization parameters. When ``None``, + :func:`depth_encoder_defaults` is used. queue_maxsize: Max frames to buffer per camera before back-pressure drops frames. + encoder_threads: Number of encoder threads (global setting). + ``None`` lets the codec decide. """ self.fps = fps - self._camera_encoder = camera_encoder or camera_encoder_defaults() + self._rgb_encoder = rgb_encoder or rgb_encoder_defaults() + self._depth_encoder = depth_encoder or depth_encoder_defaults() self._encoder_threads = encoder_threads self.queue_maxsize = queue_maxsize @@ -843,18 +926,25 @@ class StreamingVideoEncoder: self._episode_active = False self._closed = False - def start_episode(self, video_keys: list[str], temp_dir: Path) -> None: + def start_episode( + self, video_keys: list[str], temp_dir: Path, depth_video_keys: list[str] | None = None + ) -> None: """Start encoder threads for a new episode. Args: video_keys: List of video feature keys (e.g. ["observation.images.laptop"]) temp_dir: Base directory for temporary MP4 files + depth_video_keys: List of video or image feature keys that carry depth maps (e.g. + ["observation.images.laptop_depth"]). Defaults to ``[]`` (no depth keys). """ if self._episode_active: self.cancel_episode() self._dropped_frames.clear() + if depth_video_keys is None: + depth_video_keys = [] + for video_key in video_keys: frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize) result_queue: queue.Queue = queue.Queue(maxsize=1) @@ -863,17 +953,15 @@ class StreamingVideoEncoder: temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir)) video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4" - vcodec = self._camera_encoder.vcodec - codec_options = self._camera_encoder.get_codec_options(self._encoder_threads, as_strings=True) + encoder = self._depth_encoder if video_key in depth_video_keys else self._rgb_encoder encoder_thread = _CameraEncoderThread( video_path=video_path, fps=self.fps, - vcodec=vcodec, - pix_fmt=self._camera_encoder.pix_fmt, - codec_options=codec_options, + video_encoder=encoder, frame_queue=frame_queue, result_queue=result_queue, stop_event=stop_event, + encoder_threads=self._encoder_threads, ) encoder_thread.start() @@ -1080,15 +1168,23 @@ def get_audio_info(video_path: Path | str) -> dict: def get_video_info( video_path: Path | str, - camera_encoder: VideoEncoderConfig | None = None, + video_encoder: VideoEncoderConfig | None = None, ) -> dict: """Build the ``video.*`` / ``audio.*`` info dict persisted in ``info.json``. Args: video_path: Path to the encoded video file to probe. - camera_encoder: If provided, record the exact encoder settings used to encode this + video_encoder: If provided, record the exact encoder settings used to encode this video. Stream-derived values take precedence — encoder fields are only written for keys - not already populated from the video file itself. + not already populated from the video file itself. When a + :class:`~lerobot.configs.video.DepthEncoderConfig` is passed, the depth + quantization parameters (``depth_min`` / ``depth_max`` / ``shift`` / + ``use_log``) are recorded so frames can be dequantized on read. + + Returns: + The ``video.*`` / ``audio.*`` info dict, including ``is_depth_map`` which is + ``True`` only when ``video_encoder`` is a + :class:`~lerobot.configs.video.DepthEncoderConfig`. """ logging.getLogger("libav").setLevel(av.logging.WARNING) @@ -1106,13 +1202,10 @@ def get_video_info( video_info["video.width"] = video_stream.width video_info["video.codec"] = video_stream.codec.canonical_name video_info["video.pix_fmt"] = video_stream.pix_fmt - video_info["video.is_depth_map"] = False # Calculate fps from r_frame_rate video_info["video.fps"] = int(video_stream.base_rate) - - pixel_channels = get_video_pixel_channels(video_stream.pix_fmt) - video_info["video.channels"] = pixel_channels + video_info["video.channels"] = get_pix_fmt_channels(video_stream.pix_fmt) # Reset logging level av.logging.restore_default_callback() @@ -1121,27 +1214,18 @@ def get_video_info( video_info.update(**get_audio_info(video_path)) # Add additional encoder configuration if provided - if camera_encoder is not None: - for field_name, field_value in asdict(camera_encoder).items(): + if video_encoder is not None: + for field_name, field_value in asdict(video_encoder).items(): # vcodec is already populated from the video stream if field_name == "vcodec": continue video_info.setdefault(f"video.{field_name}", field_value) + video_info["is_depth_map"] = isinstance(video_encoder, DepthEncoderConfig) + return video_info -def get_video_pixel_channels(pix_fmt: str) -> int: - if "gray" in pix_fmt or "depth" in pix_fmt or "monochrome" in pix_fmt: - return 1 - elif "rgba" in pix_fmt or "yuva" in pix_fmt: - return 4 - elif "rgb" in pix_fmt or "yuv" in pix_fmt: - return 3 - else: - raise ValueError("Unknown format") - - def get_video_duration_in_s(video_path: Path | str) -> float: """ Get the duration of a video file in seconds using PyAV. @@ -1202,10 +1286,13 @@ class VideoEncodingManager: img_dir = self.dataset.root / "images" if img_dir.exists(): png_files = list(img_dir.rglob("*.png")) - if len(png_files) == 0: + tiff_files = list(img_dir.rglob("*.tiff")) + if len(png_files) == 0 and len(tiff_files) == 0: shutil.rmtree(img_dir) logger.debug("Cleaned up empty images directory") else: - logger.debug(f"Images directory is not empty, containing {len(png_files)} PNG files") + logger.debug( + f"Images directory is not empty, containing {len(png_files)} PNG and {len(tiff_files)} TIFF files" + ) return False # Don't suppress the original exception diff --git a/src/lerobot/envs/configs.py b/src/lerobot/envs/configs.py index 84c40472f..3624357e2 100644 --- a/src/lerobot/envs/configs.py +++ b/src/lerobot/envs/configs.py @@ -757,7 +757,7 @@ class RoboTwinEnvConfig(EnvConfig): task: str = "beat_block_hammer" # single task or comma-separated list fps: int = 25 - episode_length: int = 300 + episode_length: int = 1200 obs_type: str = "pixels_agent_pos" render_mode: str = "rgb_array" # Available cameras from RoboTwin's aloha-agilex embodiment: head_camera @@ -768,6 +768,9 @@ class RoboTwinEnvConfig(EnvConfig): # must equal what SAPIEN actually renders. observation_height: int = 240 observation_width: int = 320 + # "joint": 14-d joint-space control. "ee": 16-d end-effector-pose deltas executed via CuRobo IK + # (for world-model policies like LingBot-VA that predict per-arm xyz+quaternion+gripper poses). + action_mode: str = "joint" features: dict[str, PolicyFeature] = field( default_factory=lambda: { ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)), @@ -784,6 +787,8 @@ class RoboTwinEnvConfig(EnvConfig): ) def __post_init__(self): + if self.action_mode == "ee": + self.features[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(16,)) cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()] for cam in cam_list: self.features[f"pixels/{cam}"] = PolicyFeature( @@ -826,6 +831,7 @@ class RoboTwinEnvConfig(EnvConfig): observation_height=self.observation_height, observation_width=self.observation_width, episode_length=self.episode_length, + action_mode=self.action_mode, ) diff --git a/src/lerobot/envs/robotwin.py b/src/lerobot/envs/robotwin.py index 823f14fa0..5b03f337b 100644 --- a/src/lerobot/envs/robotwin.py +++ b/src/lerobot/envs/robotwin.py @@ -17,6 +17,7 @@ from __future__ import annotations import importlib import logging +import os from collections import defaultdict from collections.abc import Callable, Sequence from functools import partial @@ -28,9 +29,17 @@ import torch from gymnasium import spaces from lerobot.types import RobotObservation +from lerobot.utils.import_utils import _scipy_available from .utils import _LazyAsyncVectorEnv +# scipy is only used for end-effector-pose composition (``--env.action_mode=ee``); guard it so this +# module (and its base-env unit tests, which mock the RoboTwin runtime) imports without scipy installed. +if _scipy_available: + from scipy.spatial.transform import Rotation +else: + Rotation = None + logger = logging.getLogger(__name__) # Camera names as used by RoboTwin 2.0. The wrapper appends "_rgb" when looking @@ -41,10 +50,124 @@ ROBOTWIN_CAMERA_NAMES: tuple[str, ...] = ( "right_camera", ) -ACTION_DIM = 14 # 7 DOF × 2 arms +ACTION_DIM = 14 # 7 DOF × 2 arms (joint-space control mode) +# End-effector-pose control mode: per arm [x, y, z, qx, qy, qz, qw, gripper] = 8, dual-arm = 16. +# Used by world-model policies (e.g. LingBot-VA) that predict eef-pose deltas executed via CuRobo IK. +EEF_ACTION_DIM = 16 ACTION_LOW = -1.0 ACTION_HIGH = 1.0 -DEFAULT_EPISODE_LENGTH = 300 +DEFAULT_EPISODE_LENGTH = 1200 +OFFICIAL_INSTRUCTION_ENV = "LEROBOT_ROBOTWIN_OFFICIAL_INSTRUCTION" +OFFICIAL_INSTRUCTION_TYPE_ENV = "LEROBOT_ROBOTWIN_INSTRUCTION_TYPE" +OFFICIAL_INSTRUCTION_MAX_ENV = "LEROBOT_ROBOTWIN_INSTRUCTION_MAX" + + +def _compose_eef_pose(new_pose: np.ndarray, init_pose: np.ndarray) -> np.ndarray: + """Compose a single-arm predicted delta pose onto the initial pose. + + ``new_pose`` / ``init_pose`` are 8-vectors ``[x, y, z, qx, qy, qz, qw, gripper]``. Translation + is added, rotation is composed (``init_R * new_R``), and the gripper is taken from the + prediction. Mirrors ``add_eef_pose`` in the upstream LingBot-VA RoboTwin client. + """ + new_r = Rotation.from_quat(new_pose[3:7]) + init_r = Rotation.from_quat(init_pose[3:7]) + out_rot = (init_r * new_r).as_quat() + out_trans = new_pose[:3] + init_pose[:3] + return np.concatenate([out_trans, out_rot, new_pose[7:8]]) + + +def _add_init_eef_pose(delta_pose: np.ndarray, init_pose: np.ndarray) -> np.ndarray: + """Compose a dual-arm (16-d) predicted delta pose onto the initial eef pose, normalizing quats.""" + left = _compose_eef_pose(delta_pose[:8], init_pose[:8]) + right = _compose_eef_pose(delta_pose[8:], init_pose[8:]) + out = np.concatenate([left, right]) + # Normalize the two quaternions (indices 3:7 and 11:15) as the upstream client does. + out[3:7] = out[3:7] / (np.linalg.norm(out[3:7]) + 1e-8) + out[11:15] = out[11:15] / (np.linalg.norm(out[11:15]) + 1e-8) + return out + + +def _env_flag(name: str, default: bool = False) -> bool: + raw = os.environ.get(name) + if raw is None: + return default + return raw.strip().lower() in {"1", "true", "yes", "on"} + + +def _arm_for_block(block: Any) -> str: + return "left" if float(block.get_pose().p[0]) < 0 else "right" + + +def _robotwin_blocks_episode_info(task_name: str, env: Any) -> dict[str, str] | None: + """Infer the episode-info dict used by RoboTwin's official instruction generator for block ranking.""" + if task_name == "blocks_ranking_rgb": + return { + "{A}": "red block", + "{B}": "green block", + "{C}": "blue block", + "{a}": _arm_for_block(env.block1), + "{b}": _arm_for_block(env.block2), + "{c}": _arm_for_block(env.block3), + } + if task_name == "blocks_ranking_size": + return { + "{A}": "large block", + "{B}": "medium block", + "{C}": "small block", + "{a}": _arm_for_block(env.block1), + "{b}": _arm_for_block(env.block2), + "{c}": _arm_for_block(env.block3), + } + return None + + +def _generate_robotwin_official_instruction(task_name: str, env: Any) -> str: + """Generate language with RoboTwin's official task templates, matching its eval client.""" + fallback = task_name.replace("_", " ") + episode_info = _robotwin_blocks_episode_info(task_name, env) + if episode_info is None: + logger.warning( + "Official RoboTwin instruction is not implemented for task=%s; using %r.", task_name, fallback + ) + return fallback + + try: + # Part of the robotwin simulator repo, this is being pulled by the docker image running robotwin + # see https://github.com/RoboTwin-Platform/RoboTwin/tree/main/description + # Used to generate the official instructions + from description.utils.generate_episode_instructions import generate_episode_descriptions + except Exception: + logger.warning( + "Failed to import RoboTwin official instruction generator; using %r.", fallback, exc_info=True + ) + return fallback + + instruction_type = os.environ.get(OFFICIAL_INSTRUCTION_TYPE_ENV, "seen") + try: + max_descriptions = int(os.environ.get(OFFICIAL_INSTRUCTION_MAX_ENV, "1000000")) + except ValueError: + max_descriptions = 1000000 + + results = generate_episode_descriptions(task_name, [episode_info], max_descriptions=max_descriptions) + if not results: + logger.warning( + "RoboTwin generated no official instructions for task=%s; using %r.", task_name, fallback + ) + return fallback + + options = results[0].get(instruction_type) or results[0].get("seen") or results[0].get("unseen") + if not options: + logger.warning( + "RoboTwin generated no %s official instructions for task=%s; using %r.", + instruction_type, + task_name, + fallback, + ) + return fallback + + return str(np.random.choice(options)) + + # D435 dims from task_config/_camera_config.yml (what demo_clean.yml selects). DEFAULT_CAMERA_H = 240 DEFAULT_CAMERA_W = 320 @@ -234,6 +357,7 @@ class RoboTwinEnv(gym.Env): observation_width: int | None = None, episode_length: int = DEFAULT_EPISODE_LENGTH, render_mode: str = "rgb_array", + action_mode: str = "joint", ): super().__init__() self.task_name = task_name @@ -241,6 +365,13 @@ class RoboTwinEnv(gym.Env): self.task_description = task_name.replace("_", " ") self.episode_index = episode_index self._reset_stride = n_envs + # "joint": 14-d joint-space actions via take_action(action). "ee": 16-d end-effector-pose + # deltas (added onto the episode's initial eef pose) executed via take_action(.., "ee") + IK. + if action_mode not in ("joint", "ee"): + raise ValueError(f"action_mode must be 'joint' or 'ee'; got {action_mode!r}") + self.action_mode = action_mode + self._action_dim = EEF_ACTION_DIM if action_mode == "ee" else ACTION_DIM + self._init_eef_pose: np.ndarray | None = None self.camera_names = list(camera_names) # Default to D435 dims (the camera type baked into task_config/demo_clean.yml). # The YAML-driven lookup is deferred to reset() so construction doesn't @@ -271,7 +402,7 @@ class RoboTwinEnv(gym.Env): } ) self.action_space = spaces.Box( - low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32 + low=ACTION_LOW, high=ACTION_HIGH, shape=(self._action_dim,), dtype=np.float32 ) def _ensure_env(self) -> None: @@ -317,6 +448,18 @@ class RoboTwinEnv(gym.Env): return {"pixels": images, "agent_pos": joint_state} + def _read_eef_pose(self) -> np.ndarray: + """Read the current 16-d dual-arm eef pose [left(xyz+quat)+grip, right(xyz+quat)+grip].""" + assert self._env is not None, "_read_eef_pose called before _ensure_env()" + ep = self._env.get_obs()["endpose"] + pose = ( + list(ep["left_endpose"]) + + [ep["left_gripper"]] + + list(ep["right_endpose"]) + + [ep["right_gripper"]] + ) + return np.asarray(pose, dtype=np.float64) + def reset(self, seed: int | None = None, **kwargs) -> tuple[RobotObservation, dict]: self._ensure_env() super().reset(seed=seed) @@ -330,16 +473,32 @@ class RoboTwinEnv(gym.Env): self.episode_index += self._reset_stride self._step_count = 0 + use_official_instruction = self.task_name in {"blocks_ranking_rgb", "blocks_ranking_size"} + if _env_flag(OFFICIAL_INSTRUCTION_ENV, default=use_official_instruction): + self.task_description = _generate_robotwin_official_instruction(self.task_name, self._env) + if hasattr(self._env, "set_instruction"): + self._env.set_instruction(instruction=self.task_description) + logger.info("RoboTwin official instruction | task=%s | %s", self.task_name, self.task_description) + else: + self.task_description = self.task_name.replace("_", " ") + + # In eef mode the policy predicts pose deltas relative to the initial eef pose. + if self.action_mode == "ee": + self._init_eef_pose = self._read_eef_pose() + obs = self._get_obs() return obs, {"is_success": False, "task": self.task_name} def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]: assert self._env is not None, "step() called before reset()" - if action.ndim != 1 or action.shape[0] != ACTION_DIM: - raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}") + if action.ndim != 1 or action.shape[0] != self._action_dim: + raise ValueError(f"Expected 1-D action of shape ({self._action_dim},), got {action.shape}") with torch.enable_grad(): - if hasattr(self._env, "take_action"): + if self.action_mode == "ee": + ee_action = _add_init_eef_pose(np.asarray(action, dtype=np.float64), self._init_eef_pose) + self._env.take_action(ee_action, action_type="ee") + elif hasattr(self._env, "take_action"): self._env.take_action(action) else: self._env.step(action) @@ -398,6 +557,7 @@ def _make_env_fns( observation_height: int, observation_width: int, episode_length: int, + action_mode: str = "joint", ) -> list[Callable[[], RoboTwinEnv]]: """Return n_envs factory callables for a single task.""" @@ -410,6 +570,7 @@ def _make_env_fns( observation_height=observation_height, observation_width=observation_width, episode_length=episode_length, + action_mode=action_mode, ) return [partial(_make_one, i) for i in range(n_envs)] @@ -423,6 +584,7 @@ def create_robotwin_envs( observation_height: int = DEFAULT_CAMERA_H, observation_width: int = DEFAULT_CAMERA_W, episode_length: int = DEFAULT_EPISODE_LENGTH, + action_mode: str = "joint", ) -> dict[str, dict[int, Any]]: """Create vectorized RoboTwin 2.0 environments. @@ -473,6 +635,7 @@ def create_robotwin_envs( observation_height=observation_height, observation_width=observation_width, episode_length=episode_length, + action_mode=action_mode, ) if is_async: lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata) diff --git a/src/lerobot/envs/utils.py b/src/lerobot/envs/utils.py index 6e6f352e9..8b9c4f94b 100644 --- a/src/lerobot/envs/utils.py +++ b/src/lerobot/envs/utils.py @@ -126,6 +126,26 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten if "camera_obs" in observations: return_observations[f"{OBS_STR}.camera_obs"] = observations["camera_obs"] + # Pass through any remaining ndarray/tensor keys not already handled above, + # so env plugins can expose extra observation keys via get_env_processors(). + _handled = {"pixels", "environment_state", "agent_pos", "robot_state", "policy", "camera_obs"} + for key, value in observations.items(): + if key in _handled: + continue + target = f"{OBS_STR}.{key}" + if target in return_observations: + continue + if isinstance(value, np.ndarray): + val = torch.from_numpy(value).float() + if val.dim() == 1: + val = val.unsqueeze(0) + return_observations[target] = val + elif isinstance(value, Tensor): + val = value.float() + if val.dim() == 1: + val = val.unsqueeze(0) + return_observations[target] = val + return return_observations diff --git a/src/lerobot/jobs/__init__.py b/src/lerobot/jobs/__init__.py new file mode 100644 index 000000000..674b98b85 --- /dev/null +++ b/src/lerobot/jobs/__init__.py @@ -0,0 +1,23 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from lerobot.utils.import_utils import require_package + +# LeRobotDataset (imported at module top in dataset.py) pulls in heavy dataset deps; +# guard the optional dependency here so importing this package fails loudly if it's missing. +require_package("datasets", extra="dataset") + +from .hf import submit_to_hf + +__all__ = ["submit_to_hf"] diff --git a/src/lerobot/jobs/dataset.py b/src/lerobot/jobs/dataset.py new file mode 100644 index 000000000..497f8445e --- /dev/null +++ b/src/lerobot/jobs/dataset.py @@ -0,0 +1,53 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Make a training dataset reachable from an HF Job pod. + +The pod can't see the host's ~/.cache/huggingface/lerobot, so the dataset has to +live on the Hub: the pod downloads it by repo_id at train time (the forwarded +HF_TOKEN covers private datasets). A dataset already on the Hub is used as-is; a +local-only dataset is pushed to a PRIVATE repo first (never public). +""" + +from __future__ import annotations + +from typing import TYPE_CHECKING + +from lerobot.datasets import LeRobotDataset +from lerobot.utils.constants import HF_LEROBOT_HOME + +if TYPE_CHECKING: + from huggingface_hub import HfApi + + +def ensure_dataset_available(repo_id: str, *, api: HfApi, tags: list[str] | None = None) -> None: + """Ensure repo_id resolves on the Hub, pushing a local-only dataset privately first. + + `tags` are attached to the dataset only when we push it (an already-on-Hub + dataset is left untouched). Raises RuntimeError if the dataset is neither on + the Hub nor in the local cache. + """ + if api.repo_exists(repo_id, repo_type="dataset"): + return + + local_present = (HF_LEROBOT_HOME / repo_id / "meta" / "info.json").is_file() + if not local_present: + raise RuntimeError( + f"Dataset '{repo_id}' is not in the local cache ({HF_LEROBOT_HOME}) and could not be " + f"reached on the Hub — it may not exist, or be private and inaccessible with your " + f"token. Record or download it first, or run `hf auth login`." + ) + + print(f"[dataset] '{repo_id}' is local-only; pushing to a PRIVATE Hub repo...") + LeRobotDataset(repo_id).push_to_hub(private=True, tags=tags) + print(f"[dataset] '{repo_id}' uploaded (private). The job will download it by repo_id.") diff --git a/src/lerobot/jobs/hf.py b/src/lerobot/jobs/hf.py new file mode 100644 index 000000000..645666412 --- /dev/null +++ b/src/lerobot/jobs/hf.py @@ -0,0 +1,425 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Run a lerobot training on HF Jobs (HuggingFace GPUs). + +Ported and simplified from lelab's runners/hf_cloud.py: no UI log queue, no +registry — just submit and stream to stdout. +""" + +from __future__ import annotations + +import copy +import datetime as dt +import json +import netrc +import os +import re +import signal +import sys +import tempfile +import threading +from pathlib import Path +from typing import TYPE_CHECKING + +import httpx +from huggingface_hub import ( + HfApi, + create_repo, + fetch_job_logs, + get_token, + inspect_job, + run_job, + upload_file, +) + +from lerobot.common.train_utils import push_checkpoint_to_hub +from lerobot.configs import parser + +from .dataset import ensure_dataset_available + +if TYPE_CHECKING: + from lerobot.configs.train import TrainPipelineConfig + +_SLUG_RE = re.compile(r"[^a-zA-Z0-9._-]+") + +_TERMINAL_STAGES = {"COMPLETED", "CANCELED", "ERROR", "DELETED"} + +# huggingface_hub 1.x runs on httpx: transient HTTP/transport failures surface as +# httpx.HTTPError and socket-level errors as OSError. Catching only these keeps real +# bugs (TypeError, AttributeError, ...) from being silently retried or counted as +# job failures. +_TRANSIENT_NET_ERRORS = (OSError, httpx.HTTPError) + +# Always attached to remote jobs and pushed datasets so LeRobot-originated work +# is identifiable on the Hub; callers (e.g. LeLab) add their own via --job.tags. +LEROBOT_TAG = "lerobot" + + +def resolve_job_tags(extra: list[str] | None) -> list[str]: + """Return the tag list for a run: the lerobot tag plus any extras, deduped, order-stable.""" + tags = [LEROBOT_TAG, *(extra or [])] + seen: set[str] = set() + return [t for t in tags if not (t in seen or seen.add(t))] + + +def resolve_wandb_api_key() -> str | None: + """Host's wandb key for forwarding to the job: $WANDB_API_KEY, else ~/.netrc.""" + key = os.environ.get("WANDB_API_KEY") + if key: + return key + try: + rc = netrc.netrc() + except (FileNotFoundError, netrc.NetrcParseError, OSError): + return None + auth = rc.authenticators("api.wandb.ai") + if auth is None: + return None + _login, _account, password = auth + return password or None + + +def build_repo_id(username: str, job_name: str, now: dt.datetime) -> str: + """Generate the model repo id for a remote run: /_.""" + slug = _SLUG_RE.sub("-", job_name).strip("-") or "train" + stamp = now.strftime("%Y-%m-%d_%H-%M-%S") + return f"{username}/{slug}_{stamp}" + + +def build_remote_config_file(cfg, repo_id: str, dest: Path, tags: list[str] | None = None) -> Path: + """Write a train_config.json for the pod, with remote overrides applied. + + The pod runs `lerobot-train --config_path=` and downloads the dataset + by repo_id into its own cache. Client-only fields are stripped so the config + is accepted by the trainer image: `job` (pure client orchestration) is always + removed, and `save_checkpoint_to_hub` is removed unless explicitly enabled — + older lerobot images reject unknown keys, so the default keeps the config + compatible with the released `lerobot-gpu` image. `tags` are merged into + policy.tags so the trained model the pod pushes carries them too. + """ + remote = copy.deepcopy(cfg) + remote.policy.push_to_hub = True + remote.policy.repo_id = repo_id + # Don't pin the client's resolved device (e.g. "mps"); let the pod auto-detect its GPU. + remote.policy.device = None + # Drop any host-local dataset root; the pod resolves the dataset by repo_id. + remote.dataset.root = None + if tags: + existing = list(remote.policy.tags or []) + remote.policy.tags = existing + [t for t in tags if t not in existing] + + # Encode to the canonical, pod-parseable dict, then drop the keys the released + # trainer image doesn't know about. + data = remote.to_dict() + data.pop("job", None) + if not remote.save_checkpoint_to_hub: + data.pop("save_checkpoint_to_hub", None) + + dest.parent.mkdir(parents=True, exist_ok=True) + dest.write_text(json.dumps(data, indent=4)) + return dest + + +def _stage_config_on_hub(cfg, repo_id: str, token: str, tags: list[str] | None = None) -> str: + """Upload train_config.json to the model repo and return the repo_id for --config_path.""" + create_repo(repo_id, repo_type="model", private=True, exist_ok=True, token=token) + with tempfile.TemporaryDirectory() as tmp: + config_path = build_remote_config_file(cfg, repo_id, Path(tmp) / "train_config.json", tags=tags) + upload_file( + path_or_fileobj=config_path, + path_in_repo="train_config.json", + repo_id=repo_id, + repo_type="model", + token=token, + ) + return repo_id + + +def _tail_logs( + job_id: str, + done: threading.Event, + success_marker: str | None = None, + success_event: threading.Event | None = None, +) -> None: + """Stream job logs to stdout, reconnecting on dropped streams until done is set. + + Each reconnect re-fetches the full buffered log, so we track how many lines + were already printed and skip them — otherwise a fast-failing job's traceback + gets reprinted on every reconnect. + + When `success_marker` appears in a line, set `success_event` and `done` so the + caller can finish as soon as the trained model lands on the Hub, rather than + waiting out the platform's post-run finalization (which can add ~30s). + """ + printed = 0 + while not done.is_set(): + try: + seen = 0 + for line in fetch_job_logs(job_id=job_id, follow=True): + seen += 1 + if seen <= printed: + continue # already shown on a previous connection + printed = seen + # fetch_job_logs yields SSE data without trailing newlines, so add one + # per entry — otherwise all log lines concatenate onto a single line. + print(line.rstrip("\n"), flush=True) + if success_marker and success_event is not None and success_marker in line: + success_event.set() + done.set() + return + if done.is_set(): + return + # Stream closed cleanly. Wait a moment so the status poller can mark + # the job terminal before we reconnect (avoids re-tailing the buffer). + if done.wait(3): + return + except _TRANSIENT_NET_ERRORS: + if done.wait(2): + return + + +def _poll_until_done( + job_id: str, + done: threading.Event, + poll_interval: float = 5.0, + status_holder: dict | None = None, + max_failures: int = 6, +) -> str | None: + """Poll inspect_job until a terminal stage or until `done` is set. + + Returns the terminal stage string, or None if `done` was set first (detach) + or after `max_failures` consecutive inspect_job errors. When a terminal stage + is reached and `status_holder` is given, records `status_holder["message"]` + (the platform's status message, e.g. "Job timeout"). + """ + failures = 0 + while not done.is_set(): + try: + info = inspect_job(job_id=job_id) + failures = 0 + # `stage` is an enum in some huggingface_hub versions and a plain str in others. + stage = getattr(info.status.stage, "value", info.status.stage) + if stage in _TERMINAL_STAGES: + if status_holder is not None: + status_holder["message"] = getattr(info.status, "message", None) + done.set() + return stage + except _TRANSIENT_NET_ERRORS: + failures += 1 + if failures >= max_failures: + done.set() + return None + done.wait(poll_interval) + return None + + +def _pod_forwarded_args( + argv: list[str], drop_names: tuple[str, ...] = (), drop_prefixes: tuple[str, ...] = () +) -> list[str]: + """User CLI overrides to replay on the pod, minus flags the submitter sets itself. + + Handles both `--name=value` and `--name value` forms. Forwarding the user's overrides (e.g. + `--steps`, `--save_checkpoint_to_hub`) makes a remote resume behave like the same local command. + """ + out: list[str] = [] + skip_next = False + for i, tok in enumerate(argv): + if skip_next: + skip_next = False + continue + name = tok.split("=", 1)[0] + if name in drop_names or any(name.startswith(p) for p in drop_prefixes): + if "=" not in tok and i + 1 < len(argv) and not argv[i + 1].startswith("--"): + skip_next = True # also drop the space-separated value + continue + out.append(tok) + return out + + +def _build_resume_job(cfg: TrainPipelineConfig, username: str) -> tuple[str, list[str]]: + """Resolve the model repo and pod command to resume a run on a job. + + A Hub `config_path` is resumed from directly: its checkpoint config already targets that repo, + so new checkpoints continue the lineage there. A local `config_path` has its checkpoint uploaded + to a new PRIVATE repo first, and the resumed run is forced to push back to it. The pod command + always carries `--job.target=local` so the checkpoint's saved `job.target` can't make the pod + re-dispatch itself. + """ + config_path = parser.parse_arg("config_path") + forwarded = _pod_forwarded_args( + sys.argv[1:], + drop_names=("--config_path", "--policy.repo_id", "--policy.push_to_hub", "--dataset.root"), + drop_prefixes=("--job.",), + ) + + if Path(config_path).exists(): + # Local checkpoint: stage it on the Hub so the pod can resume from it, and push back there. + # Resolve so a `last` symlink uploads under its real step name (digit), which the pod's + # latest-checkpoint lookup keys on. + checkpoint_dir = Path(cfg.checkpoint_path).resolve() + source_repo = build_repo_id(username, cfg.job_name or "train", dt.datetime.now(dt.UTC)) + push_checkpoint_to_hub(checkpoint_dir, source_repo, private=True) + extra = [f"--policy.repo_id={source_repo}", "--policy.push_to_hub=true"] + else: + source_repo = config_path + extra = [] + + command = [ + "lerobot-train", + *forwarded, + f"--config_path={source_repo}", + "--job.target=local", + *extra, + ] + return source_repo, command + + +def submit_to_hf(cfg: TrainPipelineConfig) -> None: + """Submit a training job to HF Jobs infrastructure. + + Validates cfg, resolves credentials, ensures the dataset is on the Hub, then either stages a + sanitized config (fresh run) or resumes from a checkpoint repo, submits the job, and tails logs + until completion or detaches immediately. Ctrl-C detaches without cancelling the remote job. + """ + token = get_token() + if not token: + raise RuntimeError("Not logged in to Hugging Face. Run `hf auth login` first.") + + api = HfApi(token=token) + user_info = api.whoami(token=token) + username = user_info["name"] + + now = dt.datetime.now(dt.UTC) + fresh_repo_id: str | None = None + if not cfg.resume: + # Resolve the model repo and mark it for push BEFORE validate(): validate() requires repo_id + # to be set whenever push_to_hub is True. (A resume reuses the checkpoint's repo instead.) + if cfg.policy is not None: + base_name = cfg.job_name or cfg.policy.type + fresh_repo_id = cfg.policy.repo_id or build_repo_id(username, base_name, now) + cfg.policy.repo_id = fresh_repo_id + cfg.policy.push_to_hub = True + else: + # Path-based policy is resolved inside validate(); fall back to a generic slug. + fresh_repo_id = build_repo_id(username, cfg.job_name or "train", now) + + cfg.validate() + + if cfg.is_reward_model_training: + raise ValueError( + "Remote training via --job.target only supports policy training, not reward models. " + "Run reward-model training locally." + ) + + secrets: dict[str, str] = {"HF_TOKEN": token} + if cfg.wandb.enable: + wandb_key = resolve_wandb_api_key() + if wandb_key is None: + raise ValueError( + "wandb is enabled but no WANDB_API_KEY found. " + "Set it via `export WANDB_API_KEY=...` or add it to ~/.netrc." + ) + secrets["WANDB_API_KEY"] = wandb_key + + tags = resolve_job_tags(cfg.job.tags) + # The dataset must be reachable from the pod for both fresh and resumed runs; a local-only + # dataset is pushed PRIVATE here. Hoisted before the resume/fresh branch since it applies to both. + ensure_dataset_available(cfg.dataset.repo_id, api=api, tags=tags) + + if cfg.resume: + repo_id, command = _build_resume_job(cfg, username) + else: + config_repo_id = _stage_config_on_hub(cfg, fresh_repo_id, token, tags=tags) + repo_id = fresh_repo_id + command = ["lerobot-train", f"--config_path={config_repo_id}"] + + print(f"Submitting job to HF Jobs (flavor={cfg.job.target}, image={cfg.job.image}) ...") + job_info = run_job( + image=cfg.job.image, + command=command, + flavor=cfg.job.target, + secrets=secrets, + timeout=cfg.job.timeout, + # HF Jobs labels are key/value; expose each tag as a queryable label. + labels=dict.fromkeys(tags, "true"), + ) + job_id = job_info.id + job_url = getattr(job_info, "url", None) + print(f"Job submitted: {job_id}") + if job_url: + print(f" Job page: {job_url}") + print(f" Model repo: https://huggingface.co/{repo_id}") + print(f" Monitor: hf jobs logs {job_id}") + print(f" Cancel: hf jobs cancel {job_id}") + + if cfg.job.detach: + return + + done = threading.Event() + detached = threading.Event() + pushed_ok = threading.Event() + stage_holder: dict[str, str | None] = {} + + def _poll() -> None: + stage_holder["stage"] = _poll_until_done(job_id, done, status_holder=stage_holder) + + poll_thread = threading.Thread(target=_poll, daemon=True) + poll_thread.start() + # Finish as soon as the model is pushed, rather than waiting out the platform's + # post-run finalization before the job stage flips to COMPLETED. This matches the + # exact log line emitted by PreTrainedPolicy.push_model_to_hub — the two must stay + # in sync. If it ever stops matching we just fall back to stage-based completion + # (~30s slower), so the contract is an optimization, not a correctness requirement. + success_marker = f"Model pushed to https://huggingface.co/{repo_id}" + log_thread = threading.Thread( + target=_tail_logs, args=(job_id, done, success_marker, pushed_ok), daemon=True + ) + log_thread.start() + + def _detach(sig, frame): + detached.set() + done.set() + print("\nDetached. Job is still running.") + print(f" Monitor: hf jobs logs {job_id}") + print(f" Cancel: hf jobs cancel {job_id}") + + # signal.signal only works on the main thread; when called from a worker thread + # (e.g. an orchestration framework) skip the Ctrl-C-detaches-instead-of-cancels + # handler rather than crashing with ValueError. + install_sigint = threading.current_thread() is threading.main_thread() + original_sigint = signal.getsignal(signal.SIGINT) if install_sigint else None + if install_sigint: + signal.signal(signal.SIGINT, _detach) + try: + # Timeout-based join so SIGINT is delivered to the main thread promptly. + while poll_thread.is_alive(): + poll_thread.join(timeout=0.5) + log_thread.join(timeout=5) + finally: + if install_sigint: + signal.signal(signal.SIGINT, original_sigint) + + if detached.is_set(): + return + + if pushed_ok.is_set(): + print(f"\nTraining complete — model pushed to https://huggingface.co/{repo_id}") + return + + stage = stage_holder.get("stage") + if stage != "COMPLETED": + message = stage_holder.get("message") + detail = f" ({message})" if message else "" + raise RuntimeError( + f"Job {job_id} ended with stage={stage}{detail}. Check logs: hf jobs logs {job_id}" + ) diff --git a/src/lerobot/optim/__init__.py b/src/lerobot/optim/__init__.py index 46676027b..2d564c25f 100644 --- a/src/lerobot/optim/__init__.py +++ b/src/lerobot/optim/__init__.py @@ -20,6 +20,7 @@ from .optimizers import ( SGDConfig as SGDConfig, XVLAAdamWConfig as XVLAAdamWConfig, load_optimizer_state, + load_optimizer_state_dict, save_optimizer_state, ) from .schedulers import ( @@ -50,6 +51,7 @@ __all__ = [ "VQBeTSchedulerConfig", # State management "load_optimizer_state", + "load_optimizer_state_dict", "load_scheduler_state", "save_optimizer_state", "save_scheduler_state", diff --git a/src/lerobot/optim/optimizers.py b/src/lerobot/optim/optimizers.py index 0bdd7a37e..0a462e1aa 100644 --- a/src/lerobot/optim/optimizers.py +++ b/src/lerobot/optim/optimizers.py @@ -27,7 +27,7 @@ from lerobot.utils.constants import ( OPTIMIZER_PARAM_GROUPS, OPTIMIZER_STATE, ) -from lerobot.utils.io_utils import deserialize_json_into_object, write_json +from lerobot.utils.io_utils import deserialize_json_into_object, load_json, write_json from lerobot.utils.utils import flatten_dict, unflatten_dict # Type alias for parameters accepted by optimizer build() methods. @@ -281,28 +281,37 @@ class MultiAdamConfig(OptimizerConfig): def save_optimizer_state( - optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path + optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], + save_dir: Path, + optim_state_dict: dict | None = None, ) -> None: """Save optimizer state to disk. Args: optimizer: Either a single optimizer or a dictionary of optimizers. save_dir: Directory to save the optimizer state. + optim_state_dict: Pre-gathered optimizer state dict (for FSDP, where the sharded state must + be gathered across ranks first). If provided, it is saved directly instead of calling + ``optimizer.state_dict()``. Only supported for a single optimizer. Defaults to None. """ if isinstance(optimizer, dict): # Handle dictionary of optimizers + if optim_state_dict is not None: + raise ValueError("optim_state_dict is not supported for a dict of optimizers") for name, opt in optimizer.items(): optimizer_dir = save_dir / name optimizer_dir.mkdir(exist_ok=True, parents=True) _save_single_optimizer_state(opt, optimizer_dir) else: # Handle single optimizer - _save_single_optimizer_state(optimizer, save_dir) + _save_single_optimizer_state(optimizer, save_dir, optim_state_dict=optim_state_dict) -def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None: +def _save_single_optimizer_state( + optimizer: torch.optim.Optimizer, save_dir: Path, optim_state_dict: dict | None = None +) -> None: """Save a single optimizer's state to disk.""" - state = optimizer.state_dict() + state = dict(optim_state_dict) if optim_state_dict is not None else optimizer.state_dict() param_groups = state.pop("param_groups") flat_state = flatten_dict(state) save_file(flat_state, save_dir / OPTIMIZER_STATE) @@ -356,3 +365,19 @@ def _load_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Pat optimizer.load_state_dict(loaded_state_dict) return optimizer + + +def load_optimizer_state_dict(save_dir: Path) -> dict: + """Read a saved optimizer state dict (safetensors + json) back into a plain dict. + + Unlike `load_optimizer_state`, this does not load into an optimizer and preserves the original + ``state`` keys verbatim (e.g. FSDP parameter FQNs, which are not integer-castable). It is used by + the FSDP resume path, where the full state must be resharded via `FSDP.optim_state_dict_to_load` + before being loaded into the (sharded) optimizer. + """ + flat_state = load_file(save_dir / OPTIMIZER_STATE) + state = unflatten_dict(flat_state) + return { + "state": state.get("state", {}), + "param_groups": load_json(save_dir / OPTIMIZER_PARAM_GROUPS), + } diff --git a/src/lerobot/optim/schedulers.py b/src/lerobot/optim/schedulers.py index 250650089..2a80f74fb 100644 --- a/src/lerobot/optim/schedulers.py +++ b/src/lerobot/optim/schedulers.py @@ -83,6 +83,50 @@ class VQBeTSchedulerConfig(LRSchedulerConfig): return LambdaLR(optimizer, lr_lambda, -1) +@LRSchedulerConfig.register_subclass("constant_with_warmup") +@dataclass +class ConstantWithWarmupSchedulerConfig(LRSchedulerConfig): + """Linear warmup followed by a constant learning rate. + + Mirrors the ``warmup_constant_lambda`` used by LingBot-VA (upstream ``wan_va/train.py``): + the LR ramps linearly from 0 to the peak over ``num_warmup_steps`` steps, then stays flat. + """ + + num_warmup_steps: int = 1000 + + def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR: + warmup_steps = self.num_warmup_steps or 0 + + def lr_lambda(current_step): + if current_step < warmup_steps: + return float(current_step) / float(max(1, warmup_steps)) + return 1.0 + + return LambdaLR(optimizer, lr_lambda, -1) + + +@LRSchedulerConfig.register_subclass("cosine_annealing_with_warmup") +@dataclass +class CosineAnnealingWithWarmupSchedulerConfig(LRSchedulerConfig): + """Linear warmup followed by cosine annealing from the peak LR to zero. + + Used by EVO1; the annealing phase always spans the remaining training steps. + """ + + num_warmup_steps: int + + def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR: + def lr_lambda(current_step: int) -> float: + if current_step < self.num_warmup_steps: + return current_step / max(1, self.num_warmup_steps) + progress = (current_step - self.num_warmup_steps) / max( + 1, num_training_steps - self.num_warmup_steps + ) + return max(0.0, 0.5 * (1.0 + math.cos(math.pi * progress))) + + return LambdaLR(optimizer, lr_lambda, -1) + + @LRSchedulerConfig.register_subclass("cosine_decay_with_warmup") @dataclass class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig): diff --git a/src/lerobot/policies/__init__.py b/src/lerobot/policies/__init__.py index 68d23c9ca..7f0bed2e0 100644 --- a/src/lerobot/policies/__init__.py +++ b/src/lerobot/policies/__init__.py @@ -17,9 +17,12 @@ from lerobot.utils.action_interpolator import ActionInterpolator as ActionInterp from .act.configuration_act import ACTConfig as ACTConfig from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig from .eo1.configuration_eo1 import EO1Config as EO1Config +from .evo1.configuration_evo1 import Evo1Config as Evo1Config from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors +from .fastwam.configuration_fastwam import FastWAMConfig as FastWAMConfig from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig as GaussianActorConfig from .groot.configuration_groot import GrootConfig as GrootConfig +from .lingbot_va.configuration_lingbot_va import LingBotVAConfig as LingBotVAConfig from .molmoact2.configuration_molmoact2 import MolmoAct2Config as MolmoAct2Config from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig from .pi0.configuration_pi0 import PI0Config as PI0Config @@ -42,8 +45,11 @@ __all__ = [ "ACTConfig", "DiffusionConfig", "EO1Config", + "FastWAMConfig", "GaussianActorConfig", + "Evo1Config", "GrootConfig", + "LingBotVAConfig", "MolmoAct2Config", "MultiTaskDiTConfig", "PI0Config", diff --git a/src/lerobot/policies/act/modeling_act.py b/src/lerobot/policies/act/modeling_act.py index 5651fbfb1..1432b68a5 100644 --- a/src/lerobot/policies/act/modeling_act.py +++ b/src/lerobot/policies/act/modeling_act.py @@ -148,7 +148,7 @@ class ACTPolicy(PreTrainedPolicy): l1_loss = (abs_err * valid_mask).sum() / num_valid.clamp_min(1) loss_dict = {"l1_loss": l1_loss.item()} - if self.config.use_vae: + if self.config.use_vae and log_sigma_x2_hat is not None: # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for # each dimension independently, we sum over the latent dimension to get the total # KL-divergence per batch element, then take the mean over the batch. diff --git a/src/lerobot/policies/diffusion/modeling_diffusion.py b/src/lerobot/policies/diffusion/modeling_diffusion.py index 9fbe1f703..8758a7e29 100644 --- a/src/lerobot/policies/diffusion/modeling_diffusion.py +++ b/src/lerobot/policies/diffusion/modeling_diffusion.py @@ -101,11 +101,23 @@ class DiffusionPolicy(PreTrainedPolicy): @torch.no_grad() def predict_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor: - """Predict a chunk of actions given environment observations.""" - # stack n latest observations from the queue - batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues} - actions = self.diffusion.generate_actions(batch, noise=noise) + """Predict a chunk of actions given environment observations. + Supports two modes: + - Online (queues populated via select_action): stacks observations from internal queues. + - Offline (empty queues, e.g. dataloader batch): uses the batch directly. + """ + queues_populated = any(len(q) > 0 for q in self._queues.values()) + if queues_populated: + batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues} + else: + batch = dict(batch) + if self.config.image_features: + for key in self.config.image_features: + if batch[key].ndim == 4: + batch[key] = batch[key].unsqueeze(1) + batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4) + actions = self.diffusion.generate_actions(batch, noise=noise) return actions @torch.no_grad() diff --git a/src/lerobot/policies/evo1/README.md b/src/lerobot/policies/evo1/README.md new file mode 120000 index 000000000..6c4284fb9 --- /dev/null +++ b/src/lerobot/policies/evo1/README.md @@ -0,0 +1 @@ +../../../../docs/source/policy_evo1_README.md \ No newline at end of file diff --git a/src/lerobot/policies/evo1/__init__.py b/src/lerobot/policies/evo1/__init__.py new file mode 100644 index 000000000..581b2b824 --- /dev/null +++ b/src/lerobot/policies/evo1/__init__.py @@ -0,0 +1,19 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from .configuration_evo1 import Evo1Config +from .modeling_evo1 import Evo1Policy +from .processor_evo1 import make_evo1_pre_post_processors + +__all__ = ["Evo1Config", "Evo1Policy", "make_evo1_pre_post_processors"] diff --git a/src/lerobot/policies/evo1/configuration_evo1.py b/src/lerobot/policies/evo1/configuration_evo1.py new file mode 100644 index 000000000..534e84f75 --- /dev/null +++ b/src/lerobot/policies/evo1/configuration_evo1.py @@ -0,0 +1,252 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +from dataclasses import dataclass, field + +from lerobot.configs.policies import PreTrainedConfig +from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature +from lerobot.optim.optimizers import AdamWConfig +from lerobot.optim.schedulers import CosineAnnealingWithWarmupSchedulerConfig +from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE + +from ..rtc.configuration_rtc import RTCConfig + +logger = logging.getLogger(__name__) + + +@PreTrainedConfig.register_subclass("evo1") +@dataclass +class Evo1Config(PreTrainedConfig): + training_stage: str = "stage1" + # When True and the policy runs on CUDA, EVO1 wraps its own forward passes (training and + # inference) in a bfloat16 autocast block, so its numerics do not depend on the dtype of any + # outer autocast context opened by lerobot-train/lerobot-eval. + use_amp: bool = True + + n_obs_steps: int = 1 + chunk_size: int = 50 + n_action_steps: int = 50 + + max_state_dim: int = 24 + max_action_dim: int = 24 + max_views: int = 3 + image_resolution: tuple[int, int] = (448, 448) + empty_cameras: int = 0 + postprocess_action_dim: int | None = None + binarize_gripper: bool = False + gripper_index: int = 6 + gripper_threshold: float = 0.5 + gripper_below_threshold_value: float = 1.0 + gripper_above_threshold_value: float = -1.0 + + normalization_mapping: dict[str, NormalizationMode] = field( + default_factory=lambda: { + "VISUAL": NormalizationMode.IDENTITY, + "STATE": NormalizationMode.MIN_MAX, + "ACTION": NormalizationMode.MIN_MAX, + } + ) + + vlm_model_name: str = "OpenGVLab/InternVL3-1B-hf" + vlm_num_layers: int | None = 14 + vlm_dtype: str = "bfloat16" + # Max token length for tokenizing the (image placeholders + instruction) prompt. Prompts longer + # than this are right-truncated, so raise it for tasks with long language instructions or many views. + max_text_length: int = 1024 + use_flash_attn: bool = True + action_head: str = "flowmatching" + embed_dim: int = 896 + hidden_dim: int = 1024 + state_hidden_dim: int = 1024 + num_heads: int = 8 + num_layers: int = 8 + dropout: float = 0.0 + num_inference_timesteps: int = 32 + num_categories: int = 1 + # When True, the action head is conditioned on a single pooled VL token (the last non-padding + # token of the causal decoder) instead of the full fused token sequence. + return_cls_only: bool = False + enable_gradient_checkpointing: bool = True + gradient_checkpointing_use_reentrant: bool = False + + finetune_vlm: bool | None = None + finetune_language_model: bool | None = None + finetune_vision_model: bool | None = None + finetune_action_head: bool | None = None + # Reapply stage defaults after loading checkpoint configs so stage2 cannot + # accidentally inherit the frozen VLM flags stored by a stage1 checkpoint. + apply_training_stage_defaults: bool = True + + task_field: str = "task" + embodiment_id_field: str | None = None + default_embodiment_id: int = 0 + + # Real-Time Chunking guidance for asynchronous inference (lerobot-rollout --inference.type=rtc + # sets this and calls init_rtc_processor()); None disables RTC. + rtc_config: RTCConfig | None = None + + optimizer_lr: float = 1e-5 + optimizer_betas: tuple[float, float] = (0.9, 0.999) + optimizer_eps: float = 1e-8 + optimizer_weight_decay: float = 1e-5 + optimizer_grad_clip_norm: float = 1.0 + + scheduler_warmup_steps: int = 300 + + def __post_init__(self): + super().__post_init__() + if self.training_stage not in {"stage1", "stage2"}: + raise ValueError( + f"Unsupported EVO1 training_stage '{self.training_stage}', expected 'stage1' or 'stage2'" + ) + + if self.apply_training_stage_defaults: + stage_defaults = { + "stage1": { + "finetune_vlm": False, + "finetune_language_model": False, + "finetune_vision_model": False, + "finetune_action_head": True, + }, + "stage2": { + "finetune_vlm": True, + "finetune_language_model": True, + "finetune_vision_model": True, + "finetune_action_head": True, + }, + }[self.training_stage] + for flag_name, default_value in stage_defaults.items(): + current_value = getattr(self, flag_name) + if current_value is not None and current_value != default_value: + logger.warning( + "EVO1 %s=%s is overridden by training_stage=%s default %s. " + "Set apply_training_stage_defaults=false to keep explicit finetuning flags.", + flag_name, + current_value, + self.training_stage, + default_value, + ) + setattr(self, flag_name, default_value) + elif self.training_stage == "stage1": + if self.finetune_vlm is None: + self.finetune_vlm = False + if self.finetune_language_model is None: + self.finetune_language_model = False + if self.finetune_vision_model is None: + self.finetune_vision_model = False + if self.finetune_action_head is None: + self.finetune_action_head = True + elif self.training_stage == "stage2": + has_explicit_branch_flags = any( + flag is not None for flag in (self.finetune_language_model, self.finetune_vision_model) + ) + if not has_explicit_branch_flags: + # An explicit finetune_vlm decides both branches; otherwise stage2 defaults to a + # full-VLM finetune. + vlm_finetune = self.finetune_vlm if self.finetune_vlm is not None else True + self.finetune_vlm = vlm_finetune + self.finetune_language_model = vlm_finetune + self.finetune_vision_model = vlm_finetune + elif self.finetune_vlm is None: + self.finetune_vlm = bool(self.finetune_language_model or self.finetune_vision_model) + if self.finetune_action_head is None: + self.finetune_action_head = True + + if self.finetune_vlm is None: + self.finetune_vlm = False + if self.finetune_language_model is None: + self.finetune_language_model = False + if self.finetune_vision_model is None: + self.finetune_vision_model = False + if self.finetune_action_head is None: + self.finetune_action_head = False + + branch_vlm = self.finetune_language_model or self.finetune_vision_model + if self.finetune_vlm != branch_vlm: + raise ValueError( + "Inconsistent EVO1 finetune config: " + f"finetune_vlm={self.finetune_vlm} but " + f"(finetune_language_model or finetune_vision_model)={branch_vlm}. " + "When branch-level flags are used, finetune_vlm must match their effective union." + ) + + if self.n_action_steps > self.chunk_size: + raise ValueError( + f"n_action_steps ({self.n_action_steps}) must be <= chunk_size ({self.chunk_size})" + ) + if len(self.image_resolution) != 2 or self.image_resolution[0] != self.image_resolution[1]: + raise ValueError( + "EVO1 currently expects a square image_resolution because InternVL3 preprocessing " + f"uses a scalar image_size, got {self.image_resolution}." + ) + if not 0 <= self.default_embodiment_id < self.num_categories: + raise ValueError( + f"default_embodiment_id ({self.default_embodiment_id}) must be in " + f"[0, num_categories={self.num_categories})" + ) + + def validate_features(self) -> None: + if self.input_features is None: + self.input_features = {} + if self.output_features is None: + self.output_features = {} + + for i in range(self.empty_cameras): + key = OBS_IMAGES + f".empty_camera_{i}" + if key not in self.input_features: + self.input_features[key] = PolicyFeature( + type=FeatureType.VISUAL, + shape=(3, *self.image_resolution), + ) + + if OBS_STATE not in self.input_features: + self.input_features[OBS_STATE] = PolicyFeature( + type=FeatureType.STATE, + shape=(self.max_state_dim,), + ) + + if ACTION not in self.output_features: + self.output_features[ACTION] = PolicyFeature( + type=FeatureType.ACTION, + shape=(self.max_action_dim,), + ) + + def get_optimizer_preset(self) -> AdamWConfig: + return AdamWConfig( + lr=self.optimizer_lr, + betas=self.optimizer_betas, + eps=self.optimizer_eps, + weight_decay=self.optimizer_weight_decay, + grad_clip_norm=self.optimizer_grad_clip_norm, + ) + + def get_scheduler_preset(self): + return CosineAnnealingWithWarmupSchedulerConfig( + num_warmup_steps=self.scheduler_warmup_steps, + ) + + @property + def observation_delta_indices(self) -> list[int]: + return [0] + + @property + def action_delta_indices(self) -> list[int]: + return list(range(self.chunk_size)) + + @property + def reward_delta_indices(self) -> None: + return None diff --git a/src/lerobot/policies/evo1/evo1_model.py b/src/lerobot/policies/evo1/evo1_model.py new file mode 100644 index 000000000..129071fda --- /dev/null +++ b/src/lerobot/policies/evo1/evo1_model.py @@ -0,0 +1,210 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import torch +import torch.nn as nn + +from .configuration_evo1 import Evo1Config +from .flow_matching import FlowmatchingActionHead +from .internvl3_embedder import InternVL3Embedder + + +class Evo1Model(nn.Module): + def __init__(self, config: Evo1Config, vlm_hub_kwargs: dict | None = None): + super().__init__() + self.config = config + self._device = config.device + self.return_cls_only = config.return_cls_only + # Set by Evo1Policy.init_rtc_processor() when config.rtc_config is provided. + self.rtc_processor = None + + # Gradient checkpointing only pays off when the VLM is actually being trained; keep it off + # whenever every VLM branch is frozen so the frozen forward stays cheap. + tracks_vlm_gradients = bool( + config.finetune_vlm or config.finetune_language_model or config.finetune_vision_model + ) + enable_gradient_checkpointing = config.enable_gradient_checkpointing and tracks_vlm_gradients + + self.embedder = InternVL3Embedder( + model_name=config.vlm_model_name, + image_size=int(config.image_resolution[0]), + device=self._device, + num_language_layers=config.vlm_num_layers, + model_dtype=config.vlm_dtype, + use_flash_attn=config.use_flash_attn, + max_text_length=config.max_text_length, + enable_gradient_checkpointing=enable_gradient_checkpointing, + gradient_checkpointing_use_reentrant=config.gradient_checkpointing_use_reentrant, + hub_kwargs=vlm_hub_kwargs, + ) + + action_head_type = config.action_head.lower() + if action_head_type != "flowmatching": + raise NotImplementedError(f"Unknown action_head: {action_head_type}") + + horizon = config.chunk_size + per_action_dim = config.max_action_dim + action_dim = horizon * per_action_dim + + self.horizon = horizon + self.per_action_dim = per_action_dim + self.action_head = FlowmatchingActionHead( + embed_dim=config.embed_dim, + hidden_dim=config.hidden_dim, + action_dim=action_dim, + horizon=horizon, + per_action_dim=per_action_dim, + num_heads=config.num_heads, + num_layers=config.num_layers, + dropout=config.dropout, + num_inference_timesteps=config.num_inference_timesteps, + num_categories=config.num_categories, + state_dim=config.max_state_dim, + state_hidden_dim=config.state_hidden_dim, + ).to(self._device) + + def get_vl_embeddings( + self, + images: list[torch.Tensor], + image_mask: torch.Tensor, + prompt: str | list[str] | None = None, + return_cls_only: bool | None = None, + ) -> tuple[torch.Tensor, torch.Tensor | None]: + """Fused VL embeddings from per-camera image batches. + + Args: + images: list of per-camera tensors, each shaped ``(B, C, H, W)`` with values in ``[0, 1]``. + image_mask: bool tensor ``(B, max_views)`` marking present views. + + Returns: + ``(embeddings, valid_mask)``: the fused tokens and the bool mask of attendable context + positions (None when a single pooled token is returned). + """ + if return_cls_only is None: + return_cls_only = self.return_cls_only + if not images: + raise ValueError("EVO1 expects at least one image per sample.") + + batch_size = images[0].shape[0] + if prompt is None: + prompts = [""] * batch_size + elif isinstance(prompt, str): + prompts = [prompt] * batch_size + else: + prompts = [str(p) for p in prompt] + if len(prompts) != batch_size: + raise ValueError( + f"Prompt batch size {len(prompts)} does not match image batch size {batch_size}" + ) + + if image_mask.dim() == 1: + image_mask = image_mask.unsqueeze(0) + if image_mask.shape[0] != batch_size: + raise ValueError( + f"image_mask batch size {image_mask.shape[0]} does not match image batch size {batch_size}" + ) + + return self.embedder.get_fused_image_text_embedding_batched( + camera_images=images, + image_masks=image_mask, + text_prompts=prompts, + return_cls_only=return_cls_only, + ) + + def predict_action( + self, + fused_tokens: torch.Tensor, + state: torch.Tensor, + actions_gt: torch.Tensor | None = None, + action_mask: torch.Tensor | None = None, + embodiment_ids: torch.Tensor | None = None, + context_mask: torch.Tensor | None = None, + inference_delay: int | None = None, + prev_chunk_left_over: torch.Tensor | None = None, + execution_horizon: int | None = None, + ): + if actions_gt is None: + return self.action_head.get_action( + fused_tokens, + state=state, + action_mask=action_mask, + embodiment_id=embodiment_ids, + context_mask=context_mask, + inference_delay=inference_delay, + prev_chunk_left_over=prev_chunk_left_over, + execution_horizon=execution_horizon, + rtc_processor=self.rtc_processor, + ) + return self.action_head( + fused_tokens, + state=state, + actions_gt=actions_gt, + action_mask=action_mask, + embodiment_id=embodiment_ids, + context_mask=context_mask, + ) + + def forward( + self, + fused_tokens: torch.Tensor, + state: torch.Tensor | None = None, + actions_gt: torch.Tensor | None = None, + action_mask: torch.Tensor | None = None, + embodiment_ids: torch.Tensor | None = None, + context_mask: torch.Tensor | None = None, + inference_delay: int | None = None, + prev_chunk_left_over: torch.Tensor | None = None, + execution_horizon: int | None = None, + ): + return self.predict_action( + fused_tokens, + state, + actions_gt, + action_mask, + embodiment_ids, + context_mask, + inference_delay, + prev_chunk_left_over, + execution_horizon, + ) + + def _set_module_trainable(self, module: nn.Module, trainable: bool): + for param in module.parameters(): + param.requires_grad = trainable + + def _vlm_submodule(self, name: str) -> nn.Module: + module = getattr(self.embedder.model, name, None) + if not isinstance(module, nn.Module): + raise AttributeError( + f"InternVL model {type(self.embedder.model).__name__} has no '{name}' submodule; " + "the native HF InternVL layout (language_model / vision_tower / " + "multi_modal_projector) is required to apply the EVO1 finetune flags." + ) + return module + + def set_finetune_flags(self): + # __post_init__ resolves every finetune flag to a concrete boolean, so branch-level flags + # are authoritative here. Freeze everything first, then re-enable the requested branches. + self._set_module_trainable(self.embedder, False) + self._set_module_trainable( + self._vlm_submodule("language_model"), bool(self.config.finetune_language_model) + ) + finetune_vision = bool(self.config.finetune_vision_model) + self._set_module_trainable(self._vlm_submodule("vision_tower"), finetune_vision) + self._set_module_trainable(self._vlm_submodule("multi_modal_projector"), finetune_vision) + + if not self.config.finetune_action_head: + self._set_module_trainable(self.action_head, False) diff --git a/src/lerobot/policies/evo1/flow_matching.py b/src/lerobot/policies/evo1/flow_matching.py new file mode 100644 index 000000000..207d47039 --- /dev/null +++ b/src/lerobot/policies/evo1/flow_matching.py @@ -0,0 +1,483 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +import math + +import torch +import torch.nn as nn + +logger = logging.getLogger(__name__) + + +class SinusoidalPositionalEncoding(nn.Module): + def __init__(self, dim: int, max_len: int = 1000): + super().__init__() + pe = torch.zeros(max_len, dim) + position = torch.arange(0, max_len).unsqueeze(1) + div_term = torch.exp(torch.arange(0, dim, 2) * -(math.log(10000.0) / dim)) + pe[:, 0::2] = torch.sin(position * div_term) + pe[:, 1::2] = torch.cos(position * div_term) + pe = pe.unsqueeze(0) + self.register_buffer("pe", pe) + + def forward(self, seq_len: int): + if seq_len > self.pe.size(1): + self._extend_pe(seq_len) + return self.pe[:, :seq_len, :] + + def _extend_pe(self, new_max_len): + old_max_len, dim = self.pe.size(1), self.pe.size(2) + if new_max_len <= old_max_len: + return + extra_positions = torch.arange(old_max_len, new_max_len, dtype=torch.float).unsqueeze(1) + div_term = torch.exp(torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim)) + extra_pe = torch.zeros(new_max_len - old_max_len, dim) + extra_pe[:, 0::2] = torch.sin(extra_positions * div_term) + extra_pe[:, 1::2] = torch.cos(extra_positions * div_term) + extra_pe = extra_pe.unsqueeze(0) + new_pe = torch.cat([self.pe, extra_pe.to(self.pe.device)], dim=1) + self.pe = new_pe + + +class CategorySpecificLinear(nn.Module): + def __init__(self, in_dim: int, out_dim: int, num_categories: int = 1): + super().__init__() + self.num_categories = num_categories + if num_categories <= 1: + self.linear = nn.Linear(in_dim, out_dim) + else: + self.weight = nn.Parameter(torch.empty(num_categories, in_dim, out_dim)) + self.bias = nn.Parameter(torch.zeros(num_categories, out_dim)) + # Initialize each per-category (in_dim, out_dim) matrix separately: xavier on the full + # 3D tensor would compute fan_in = in_dim * out_dim and badly under-scale the weights. + for category in range(num_categories): + nn.init.xavier_uniform_(self.weight[category]) + + def forward(self, x: torch.Tensor, category_id: torch.LongTensor): + if self.num_categories <= 1: + if x.dtype != self.linear.weight.dtype: + x = x.to(dtype=self.linear.weight.dtype) + return self.linear(x) + + if x.dtype != self.weight.dtype: + x = x.to(dtype=self.weight.dtype) + + orig_shape = x.shape + x_flat = x.reshape(-1, orig_shape[-1]) + if category_id.dim() == 0: + cid = category_id.item() + out = x_flat @ self.weight[cid] + self.bias[cid] + else: + category_id = category_id.reshape(-1) + if category_id.numel() != x_flat.size(0): + raise ValueError( + f"category_id length {category_id.numel()} does not match flattened batch {x_flat.size(0)}" + ) + weight_selected = self.weight[category_id] + bias_selected = self.bias[category_id] + out = torch.bmm(x_flat.unsqueeze(1), weight_selected).squeeze(1) + bias_selected + out_shape = orig_shape[:-1] + (out.shape[-1],) + return out.view(out_shape) + + +class CategorySpecificMLP(nn.Module): + def __init__(self, input_dim: int, hidden_dim: int, output_dim: int, num_categories: int = 1): + super().__init__() + self.fc1 = CategorySpecificLinear(input_dim, hidden_dim, num_categories) + self.fc2 = CategorySpecificLinear(hidden_dim, output_dim, num_categories) + self.activation = nn.ReLU(inplace=True) + + def forward(self, x: torch.Tensor, category_id: torch.LongTensor): + out = self.activation(self.fc1(x, category_id)) + out = self.fc2(out, category_id) + return out + + +class MultiEmbodimentActionEncoder(nn.Module): + def __init__( + self, action_dim: int, embed_dim: int, hidden_dim: int, horizon: int, num_categories: int = 1 + ): + super().__init__() + self.horizon = horizon + self.embed_dim = embed_dim + self.num_categories = num_categories + + self.W1 = CategorySpecificLinear(action_dim, hidden_dim, num_categories) + self.W2 = CategorySpecificLinear(hidden_dim, hidden_dim, num_categories) + self.W3 = CategorySpecificLinear(hidden_dim, embed_dim, num_categories) + + self.pos_encoding = SinusoidalPositionalEncoding(hidden_dim, max_len=horizon) + self.activation = nn.ReLU(inplace=True) + + def forward(self, action_seq: torch.Tensor, category_id: torch.LongTensor): + batch_size, horizon, action_dim = action_seq.shape + if self.horizon != horizon: + raise ValueError( + f"Action sequence length must match horizon: got {horizon}, expected {self.horizon}." + ) + + x = action_seq.reshape(batch_size * horizon, action_dim) + if category_id.dim() == 0: + cat_ids = category_id.expand(horizon * batch_size) + else: + cat_ids = category_id.unsqueeze(1).expand(batch_size, horizon).reshape(batch_size * horizon) + + out = self.activation(self.W1(x, cat_ids)) + pos_enc = self.pos_encoding(horizon).to(device=out.device, dtype=out.dtype) + out = out.view(batch_size, horizon, -1) + pos_enc + out = out.view(batch_size * horizon, -1) + out = self.activation(self.W2(out, cat_ids)) + out = self.W3(out, cat_ids) + return out.view(batch_size, horizon, self.embed_dim) + + +class BasicTransformerBlock(nn.Module): + def __init__(self, embed_dim: int, num_heads: int, hidden_dim: int, dropout: float = 0.0): + super().__init__() + self.attn = nn.MultiheadAttention(embed_dim, num_heads, dropout=dropout, batch_first=True) + self.norm1 = nn.LayerNorm(embed_dim) + self.norm2 = nn.LayerNorm(embed_dim) + self.ff = nn.Sequential(nn.Linear(embed_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, embed_dim)) + + def forward( + self, + action_tokens: torch.Tensor, + context_tokens: torch.Tensor, + time_emb: torch.Tensor, + context_key_padding_mask: torch.Tensor | None = None, + ): + x = self.norm1(action_tokens) + attn_out, _ = self.attn(x, context_tokens, context_tokens, key_padding_mask=context_key_padding_mask) + x = action_tokens + attn_out + x2 = self.norm2(x) + if time_emb is not None: + x2 = x2 + time_emb.unsqueeze(1) + ff_out = self.ff(x2) + return x + ff_out + + +class FlowmatchingActionHead(nn.Module): + def __init__( + self, + embed_dim: int = 896, + hidden_dim: int = 1024, + action_dim: int = 16 * 7, + horizon: int = 16, + per_action_dim: int = 7, + num_heads: int = 8, + num_layers: int = 8, + dropout: float = 0.0, + num_inference_timesteps: int = 20, + num_categories: int = 1, + state_dim: int | None = None, + state_hidden_dim: int | None = None, + ): + super().__init__() + + logger.info("FlowmatchingActionHead num_inference_timesteps=%s", num_inference_timesteps) + self.embed_dim = embed_dim + self.horizon = horizon + self.per_action_dim = per_action_dim + self.action_dim = action_dim + self.num_inference_timesteps = num_inference_timesteps + self.num_categories = num_categories + + self.time_pos_enc = SinusoidalPositionalEncoding(embed_dim, max_len=1000) + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + embed_dim=embed_dim, + num_heads=num_heads, + hidden_dim=embed_dim * 4, + dropout=dropout, + ) + for _ in range(num_layers) + ] + ) + self.norm_out = nn.LayerNorm(embed_dim) + self.seq_pool_proj = nn.Linear(self.horizon * self.embed_dim, self.embed_dim) + self.mlp_head = CategorySpecificMLP( + input_dim=embed_dim, + hidden_dim=hidden_dim, + output_dim=action_dim, + num_categories=num_categories, + ) + + self.state_encoder = None + if state_dim is not None: + state_hidden = state_hidden_dim if state_hidden_dim is not None else embed_dim + self.state_encoder = CategorySpecificMLP( + input_dim=state_dim, + hidden_dim=state_hidden, + output_dim=embed_dim, + num_categories=num_categories, + ) + + if horizon > 1: + self.action_encoder = MultiEmbodimentActionEncoder( + action_dim=self.per_action_dim, + embed_dim=embed_dim, + hidden_dim=embed_dim, + horizon=horizon, + num_categories=num_categories, + ) + self.single_action_proj = None + else: + self.action_encoder = None + self.single_action_proj = nn.Linear(self.per_action_dim, self.embed_dim) + + def _project_actions(self, action_seq: torch.Tensor, embodiment_id: torch.LongTensor) -> torch.Tensor: + if self.horizon > 1 and self.action_encoder is not None: + return self.action_encoder(action_seq, embodiment_id) + if self.single_action_proj is None: + raise RuntimeError("single_action_proj is not initialized for horizon <= 1.") + return self.single_action_proj(action_seq) + + def _expand_action_mask( + self, + action_mask: torch.Tensor, + batch_size: int, + per_action_dim: int, + device: torch.device, + dtype: torch.dtype, + ) -> torch.Tensor: + if action_mask is None: + raise ValueError("action_mask must be provided for flow matching inference.") + + if action_mask.dim() == 2: + expected_last_dim = self.horizon * per_action_dim + if action_mask.shape == (batch_size, expected_last_dim): + expanded_mask = action_mask.reshape(batch_size, self.horizon, per_action_dim) + elif action_mask.shape == (batch_size, per_action_dim): + expanded_mask = action_mask.unsqueeze(1).expand(batch_size, self.horizon, per_action_dim) + else: + raise ValueError( + f"Expected action_mask shape {(batch_size, expected_last_dim)} or " + f"{(batch_size, per_action_dim)}, got {tuple(action_mask.shape)}" + ) + elif action_mask.dim() == 3: + expected_shape = (batch_size, self.horizon, per_action_dim) + if tuple(action_mask.shape) != expected_shape: + raise ValueError( + f"Expected action_mask shape {expected_shape}, got {tuple(action_mask.shape)}" + ) + expanded_mask = action_mask + else: + raise ValueError(f"Unsupported action_mask rank: {action_mask.dim()}") + + return expanded_mask.to(device=device, dtype=dtype) + + def _prepare_context( + self, + fused_tokens: torch.Tensor, + state: torch.Tensor | None, + embodiment_id: torch.LongTensor | None, + context_mask: torch.Tensor | None, + ) -> tuple[torch.Tensor, torch.Tensor | None, torch.LongTensor]: + """Normalize the VL context and embodiment ids shared by training and inference. + + Returns the context tokens ``(B, S, E)``, a key_padding_mask for + ``nn.MultiheadAttention`` (True = ignore) or None, and the resolved embodiment ids. + """ + batch_size = fused_tokens.size(0) + device = fused_tokens.device + if embodiment_id is None: + embodiment_id = torch.zeros(batch_size, dtype=torch.long, device=device) + elif self.num_categories > 1 and ( + int(embodiment_id.min()) < 0 or int(embodiment_id.max()) >= self.num_categories + ): + raise ValueError( + f"embodiment ids must be in [0, num_categories={self.num_categories}), " + f"got range [{int(embodiment_id.min())}, {int(embodiment_id.max())}]" + ) + + context_tokens = fused_tokens + if context_tokens.dim() == 2: + # A single pooled VL token (return_cls_only): give it a sequence dim of 1. + context_tokens = context_tokens.unsqueeze(1) + context_mask = None + if state is not None and self.state_encoder is not None: + state_emb = self.state_encoder(state, embodiment_id).unsqueeze(1) + context_tokens = torch.cat([context_tokens, state_emb], dim=1) + if context_mask is not None: + state_valid = torch.ones(batch_size, 1, dtype=torch.bool, device=context_mask.device) + context_mask = torch.cat([context_mask.to(torch.bool), state_valid], dim=1) + + key_padding_mask = None if context_mask is None else ~context_mask.to(torch.bool) + return context_tokens, key_padding_mask, embodiment_id + + def forward( + self, + fused_tokens: torch.Tensor, + state: torch.Tensor = None, + actions_gt: torch.Tensor = None, + embodiment_id: torch.LongTensor = None, + action_mask: torch.Tensor = None, + context_mask: torch.Tensor = None, + ): + if actions_gt is None: + return self.get_action( + fused_tokens, + state=state, + embodiment_id=embodiment_id, + action_mask=action_mask, + context_mask=context_mask, + ) + + batch_size = fused_tokens.size(0) + device = fused_tokens.device + context_tokens, key_padding_mask, embodiment_id = self._prepare_context( + fused_tokens, state, embodiment_id, context_mask + ) + + t = ( + torch.distributions.Beta(2, 2) + .sample((batch_size,)) + .clamp(0.02, 0.98) + .to(device) + .to(dtype=self.dtype) + ) + time_index = (t * 999).long().clamp_(0, 999) + time_emb = self.time_pos_enc(1000)[:, time_index, :].squeeze(0).to(dtype=context_tokens.dtype) + + actions_gt_seq = actions_gt + noise = torch.rand_like(actions_gt) * 2 - 1 + if action_mask is not None: + action_mask = action_mask.to(dtype=noise.dtype, device=noise.device) + if action_mask.shape != noise.shape: + raise ValueError(f"action_mask shape {action_mask.shape} != noise shape {noise.shape}") + actions_gt_seq = actions_gt_seq * action_mask + noise = noise * action_mask + + if self.horizon > 1: + noise_seq = noise.view(batch_size, self.horizon, self.per_action_dim) + else: + noise_seq = noise if noise.dim() == 3 else noise.unsqueeze(1) + t_broadcast = t.view(batch_size, 1, 1) + action_intermediate_seq = (1 - t_broadcast) * noise_seq + t_broadcast * actions_gt_seq + + action_tokens = self._project_actions(action_intermediate_seq, embodiment_id) + target_dtype = self.dtype + action_tokens = action_tokens.to(dtype=target_dtype) + context_tokens = context_tokens.to(dtype=target_dtype) + time_emb = time_emb.to(dtype=target_dtype) + + x = action_tokens + for block in self.transformer_blocks: + x = block(x, context_tokens, time_emb, key_padding_mask) + x = self.norm_out(x) + + if self.horizon > 1: + x_flat = x.reshape(batch_size, -1) + x_pooled = self.seq_pool_proj(x_flat) + else: + x_pooled = x.squeeze(1) + + pred_velocity = self.mlp_head(x_pooled, embodiment_id) + return pred_velocity, noise + + def get_action( + self, + fused_tokens: torch.Tensor, + state: torch.Tensor = None, + embodiment_id: torch.LongTensor = None, + action_mask: torch.Tensor = None, + context_mask: torch.Tensor = None, + inference_delay: int | None = None, + prev_chunk_left_over: torch.Tensor | None = None, + execution_horizon: int | None = None, + rtc_processor=None, + ): + batch_size = fused_tokens.size(0) + device = fused_tokens.device + context_tokens, key_padding_mask, embodiment_id = self._prepare_context( + fused_tokens, state, embodiment_id, context_mask + ) + + action_dim_total = self.action_dim + per_action_dim = self.per_action_dim + + action = torch.rand(batch_size, action_dim_total, device=device, dtype=context_tokens.dtype) * 2 - 1 + action_seq = action.view(batch_size, self.horizon, per_action_dim) + action_mask = self._expand_action_mask( + action_mask, + batch_size=batch_size, + per_action_dim=per_action_dim, + device=action_seq.device, + dtype=action_seq.dtype, + ) + action_seq = action_seq * action_mask + + target_dtype = self.dtype + context_tokens = context_tokens.to(dtype=target_dtype) + + num_steps = int(self.num_inference_timesteps) + if num_steps <= 0: + raise ValueError(f"num_inference_timesteps must be positive, got {num_steps}") + dt = 1.0 / num_steps + + use_rtc = rtc_processor is not None and ( + inference_delay is not None or prev_chunk_left_over is not None + ) + + def predict_velocity(seq: torch.Tensor, step_time_emb: torch.Tensor) -> torch.Tensor: + """Predict the masked flow velocity (x1 - x0 convention) for one integration step.""" + seq = seq * action_mask + action_tokens = self._project_actions(seq, embodiment_id).to(dtype=target_dtype) + x = action_tokens + for block in self.transformer_blocks: + x = block(x, context_tokens, step_time_emb, key_padding_mask) + x = self.norm_out(x) + x_pooled = self.seq_pool_proj(x.reshape(batch_size, -1)) if self.horizon > 1 else x.squeeze(1) + pred = self.mlp_head(x_pooled, embodiment_id) + return pred.view(batch_size, self.horizon, per_action_dim) * action_mask + + for i in range(num_steps): + t = i / num_steps + time_index = min(int(t * 999), 999) + time_emb = self.time_pos_enc(1000)[:, time_index, :].to(device).squeeze(0).to(dtype=target_dtype) + time_emb = time_emb.unsqueeze(0).repeat(batch_size, 1) + + if use_rtc: + # RTCProcessor assumes the pi0 flow convention: its `time` runs 1 -> 0 and the + # clean-action estimate is x1 = x_t - time * v. EVO1 integrates t: 0 -> 1 with + # velocity v = x1 - x0 (so x1 = x_t + (1 - t) * v); passing time = 1 - t and + # flipping the velocity sign in both directions maps one convention onto the other. + guided = rtc_processor.denoise_step( + x_t=action_seq, + prev_chunk_left_over=prev_chunk_left_over, + inference_delay=inference_delay, + time=1.0 - t, + original_denoise_step_partial=lambda seq, emb=time_emb: -predict_velocity(seq, emb), + execution_horizon=execution_horizon, + ) + velocity = -guided + else: + velocity = predict_velocity(action_seq, time_emb) + + action_seq = action_seq + dt * velocity + + action_seq = action_seq * action_mask + return action_seq.reshape(batch_size, -1) + + @property + def device(self): + return next(self.parameters()).device + + @property + def dtype(self): + return next(self.parameters()).dtype diff --git a/src/lerobot/policies/evo1/internvl3_embedder.py b/src/lerobot/policies/evo1/internvl3_embedder.py new file mode 100644 index 000000000..d47105d96 --- /dev/null +++ b/src/lerobot/policies/evo1/internvl3_embedder.py @@ -0,0 +1,369 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +from collections.abc import Sequence +from typing import TYPE_CHECKING + +import torch +import torch.nn as nn +import torchvision.transforms.functional as tvf +from torchvision.transforms.functional import InterpolationMode + +from lerobot.utils.import_utils import _transformers_available, require_package + +if TYPE_CHECKING or _transformers_available: + from transformers import AutoModel, AutoTokenizer +else: + AutoModel = None + AutoTokenizer = None + +IMAGENET_MEAN = (0.485, 0.456, 0.406) +IMAGENET_STD = (0.229, 0.224, 0.225) +IMG_CONTEXT_TOKEN = "" # nosec B105 +IMG_START_TOKEN = "" # nosec B105 +IMG_END_TOKEN = "" # nosec B105 + +logger = logging.getLogger(__name__) + + +def _batched_resize_01(images: torch.Tensor, image_size: int) -> torch.Tensor: + """Resize a batch of ``[0, 1]`` images to ``(image_size, image_size)`` on-device. + + Numerically mirrors InternVL3's reference PIL preprocessing + (``to_pil_image`` -> ``Image.resize`` -> ``to_tensor``): the float input is quantized to uint8 + exactly as ``to_pil_image`` does, then resized with bicubic interpolation and antialiasing, + which matches PIL's default resampler. Matching the reference pixel-for-pixel keeps the policy + interchangeable with checkpoints produced by the upstream EVO1 preprocessing. + + Args: + images: float tensor of shape ``(N, C, H, W)`` with values in ``[0, 1]``. + + Returns: + float32 tensor of shape ``(N, C, image_size, image_size)`` with values in ``[0, 1]``. + """ + # to_pil_image() quantizes float [0, 1] to uint8 (x * 255, truncated); replicate that so the + # bicubic resample sees the same integer pixels PIL would. + pixels_u8 = (images * 255.0).clamp(0, 255).to(torch.uint8) + resized = tvf.resize( + pixels_u8, [image_size, image_size], interpolation=InterpolationMode.BICUBIC, antialias=True + ) + return resized.to(torch.float32) / 255.0 + + +def _batched_pixel_values( + camera_images: Sequence[torch.Tensor], + max_views: int, + image_size: int, + mean: torch.Tensor, + std: torch.Tensor, + dtype: torch.dtype, + device: torch.device | str, +) -> torch.Tensor: + """Build InternVL3 ``pixel_values`` from per-camera ``[0, 1]`` image batches without leaving the device. + + Each image is resized, converted to ``dtype``, and ImageNet-normalized (a single tile per + image), batched across the whole minibatch. Absent views (fewer cameras than ``max_views``) + are filled with zero images; their placeholder tokens are masked out of attention downstream + via ``_mask_absent_image_tokens``. + + Returns: + ``pixel_values`` of shape ``(B * max_views, C, image_size, image_size)``, ordered row-major + over ``(sample, view)`` to line up with the per-view image placeholders in the prompt. + """ + resized: list[torch.Tensor] = [] + for image in camera_images: + resized.append(_batched_resize_01(image.to(device=device), image_size).to(dtype)) + + batch_size = resized[0].shape[0] + channels = resized[0].shape[1] + while len(resized) < max_views: + resized.append(torch.zeros(batch_size, channels, image_size, image_size, dtype=dtype, device=device)) + + stacked = torch.stack(resized[:max_views], dim=1) # (B, V, C, H, W) + mean = mean.to(device=device, dtype=dtype).view(1, 1, -1, 1, 1) + std = std.to(device=device, dtype=dtype).view(1, 1, -1, 1, 1) + normalized = (stacked - mean) / std + return normalized.reshape(batch_size * max_views, channels, image_size, image_size) + + +class InternVL3Embedder(nn.Module): + """Vision-language embedder using the native HF InternVL3 model (no trust_remote_code).""" + + def __init__( + self, + model_name="OpenGVLab/InternVL3-1B-hf", + image_size=448, + device="cuda", + num_language_layers: int | None = 14, + model_dtype: str | torch.dtype = "bfloat16", + use_flash_attn: bool = True, + max_text_length: int = 1024, + enable_gradient_checkpointing: bool = True, + gradient_checkpointing_use_reentrant: bool = False, + hub_kwargs: dict | None = None, + ): + super().__init__() + self._requested_device = device + self.image_size = image_size + self.num_language_layers = num_language_layers + self.max_text_length = max_text_length + self.enable_gradient_checkpointing = bool(enable_gradient_checkpointing) + self.gradient_checkpointing_use_reentrant = bool(gradient_checkpointing_use_reentrant) + hub_kwargs = hub_kwargs or {} + + require_package("transformers", extra="evo1") + + self.tokenizer = AutoTokenizer.from_pretrained(model_name, **hub_kwargs) + if isinstance(model_dtype, str): + try: + model_dtype = getattr(torch, model_dtype) + except AttributeError as exc: + raise ValueError(f"Unsupported EVO1 vlm_dtype '{model_dtype}'") from exc + self.model_dtype = model_dtype + + attn_implementation = "flash_attention_2" if (use_flash_attn and _flash_attn_available()) else "eager" + if use_flash_attn and attn_implementation == "eager": + logger.warning("flash_attn is not installed. Falling back to eager attention.") + + self.model = AutoModel.from_pretrained( + model_name, + torch_dtype=model_dtype, + attn_implementation=attn_implementation, + low_cpu_mem_usage=True, + **hub_kwargs, + ).to(self._requested_device) + + checkpoint_image_size = getattr(self.model.config.vision_config, "image_size", None) + if isinstance(checkpoint_image_size, (list, tuple)): + checkpoint_image_size = checkpoint_image_size[0] + if checkpoint_image_size is not None and int(checkpoint_image_size) != int(image_size): + raise ValueError( + f"EVO1 image_resolution ({image_size}) must match the InternVL checkpoint's native " + f"image size ({checkpoint_image_size}): the checkpoint's image_seq_length assumes " + "its native resolution, so other sizes would desync the image placeholder tokens " + "from the vision features." + ) + + self.num_image_token = self.model.config.image_seq_length + + # Truncate language model to the requested number of layers + layers = self.model.language_model.layers + if self.num_language_layers is not None: + layers = layers[: self.num_language_layers] + self.model.language_model.layers = torch.nn.ModuleList(layers) + + self._configure_memory_features() + self.img_context_token_id = self.tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) + + def _configure_memory_features(self) -> None: + checkpoint_kwargs = {"use_reentrant": self.gradient_checkpointing_use_reentrant} + + if not self.enable_gradient_checkpointing: + language_model = self.model.language_model + if hasattr(language_model, "gradient_checkpointing_disable"): + language_model.gradient_checkpointing_disable() + vision_tower = getattr(self.model, "vision_tower", None) + if vision_tower is not None and hasattr(vision_tower, "encoder"): + vision_tower.encoder.gradient_checkpointing = False + return + + def _enable_ckpt(module: nn.Module | None) -> bool: + if module is None: + return False + if hasattr(module, "gradient_checkpointing_enable"): + try: + module.gradient_checkpointing_enable(gradient_checkpointing_kwargs=checkpoint_kwargs) + except TypeError: + module.gradient_checkpointing_enable() + return True + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = True + return True + return False + + enabled_any = _enable_ckpt(self.model) + + vision_tower = getattr(self.model, "vision_tower", None) + if vision_tower is not None: + enabled_any = _enable_ckpt(vision_tower) or enabled_any + + language_model = self.model.language_model + enabled_any = _enable_ckpt(language_model) or enabled_any + if hasattr(language_model, "config"): + language_model.config.use_cache = False + + if hasattr(self.model, "config"): + self.model.config.use_cache = False + if hasattr(self.model, "enable_input_require_grads"): + self.model.enable_input_require_grads() + + if enabled_any: + logger.info("Gradient checkpointing enabled for InternVL3 embedder.") + else: + logger.warning( + "Requested gradient checkpointing, but model does not expose checkpointing controls." + ) + + def _build_multimodal_prompts( + self, + batch_num_tiles_list: list[list[int]], + text_prompts: Sequence[str], + ) -> list[str]: + prompts = [] + for num_tiles_list, text_prompt in zip(batch_num_tiles_list, text_prompts, strict=True): + prompt_segments = [] + for i, tile_count in enumerate(num_tiles_list): + token_count = self.num_image_token * tile_count + image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * token_count + IMG_END_TOKEN + prompt_segments.append(f"Image-{i + 1}: {image_tokens}\n") + prompts.append("".join(prompt_segments) + text_prompt.strip()) + return prompts + + def get_fused_image_text_embedding_batched( + self, + camera_images: Sequence[torch.Tensor], + image_masks: torch.Tensor, + text_prompts: Sequence[str], + return_cls_only: bool = True, + ): + """Fused VL embedding from per-camera ``[0, 1]`` image batches (no PIL, no host round-trip). + + Args: + camera_images: list of per-camera tensors, each shaped ``(B, C, H, W)`` in ``[0, 1]``. + image_masks: bool tensor ``(B, max_views)`` marking present views. + + Returns: + A ``(embeddings, valid_mask)`` tuple. With ``return_cls_only=False``, ``embeddings`` is + ``(B, L, H)`` and ``valid_mask`` is a ``(B, L)`` bool tensor marking tokens downstream + attention may attend to (padding and absent-view tokens are False). With + ``return_cls_only=True``, ``embeddings`` is the pooled ``(B, H)`` last-valid-token state + and ``valid_mask`` is None. + """ + max_views = int(image_masks.shape[1]) + batch_size = int(image_masks.shape[0]) + mean = torch.tensor(IMAGENET_MEAN, device=self.device, dtype=self.model_dtype) + std = torch.tensor(IMAGENET_STD, device=self.device, dtype=self.model_dtype) + pixel_values = _batched_pixel_values( + camera_images, max_views, self.image_size, mean, std, self.model_dtype, self.device + ) + # InternVL3 preprocessing uses a single tile per image (max_num=1). + batch_num_tiles_list = [[1] * max_views for _ in range(batch_size)] + return self._forward_vlm( + pixel_values, batch_num_tiles_list, image_masks, text_prompts, return_cls_only + ) + + def _mask_absent_image_tokens( + self, + input_ids: torch.Tensor, + attention_mask: torch.Tensor, + image_masks: torch.Tensor, + batch_num_tiles_list: list[list[int]], + ) -> torch.Tensor: + """Zero attention over the image-context tokens of absent (zero-padded) views. + + Fully vectorized: runs without any host<->device synchronization. + """ + # A single tile per image (max_num=1), so every image occupies the same number of + # context tokens. + tiles_per_image = ( + batch_num_tiles_list[0][0] if batch_num_tiles_list and batch_num_tiles_list[0] else 1 + ) + tokens_per_image = self.num_image_token * tiles_per_image + + image_masks = image_masks.to(device=input_ids.device).bool() + img_token_mask = input_ids == self.img_context_token_id # (B, L) + # keep[b, k] tells whether the k-th image-context token (ordered view0, view1, ...) survives. + per_token_keep = image_masks.repeat_interleave(tokens_per_image, dim=1) # (B, V * tokens_per_image) + # Rank each context token by its running position among the row's context tokens. + ctx_index = img_token_mask.to(torch.long).cumsum(dim=1) - 1 + ctx_index = ctx_index.clamp(min=0, max=per_token_keep.shape[1] - 1) + keep_here = torch.gather(per_token_keep, 1, ctx_index) # (B, L) + drop = img_token_mask & ~keep_here + return attention_mask.masked_fill(drop, 0) + + def _forward_vlm( + self, + pixel_values: torch.Tensor, + batch_num_tiles_list: list[list[int]], + image_masks: torch.Tensor, + text_prompts: Sequence[str], + return_cls_only: bool, + ): + if pixel_values.shape[0] == 0: + logger.warning("InternVL3 received an empty image batch after preprocessing.") + hidden_size = getattr(self.model.config, "hidden_size", None) + if hidden_size is None: + hidden_size = getattr(self.model.config.text_config, "hidden_size", None) + if hidden_size is None: + raise RuntimeError("Unable to infer hidden size for empty InternVL3 batch.") + return torch.empty(0, hidden_size, device=self.device, dtype=torch.float32), None + + prompts = self._build_multimodal_prompts(batch_num_tiles_list, text_prompts) + + model_inputs = self.tokenizer( + list(prompts), + return_tensors="pt", + padding=True, + truncation=True, + max_length=self.max_text_length, + ).to(self.device) + input_ids = model_inputs["input_ids"] + if input_ids.shape[1] >= self.max_text_length: + # Truncation cuts from the right, so text is dropped before image placeholders — but a + # large max_views * image_seq_length budget can still eat into them. Fail loudly instead + # of letting the VLM crash on a placeholder/vision-feature count mismatch. + expected_image_tokens = self.num_image_token * sum(batch_num_tiles_list[0]) + image_token_counts = (input_ids == self.img_context_token_id).sum(dim=1) + if not bool((image_token_counts == expected_image_tokens).all()): + raise ValueError( + f"Prompt truncation at max_text_length={self.max_text_length} cut into the " + f"image placeholder tokens ({expected_image_tokens} expected per sample). " + "Increase max_text_length or reduce max_views." + ) + attention_mask = self._mask_absent_image_tokens( + input_ids, model_inputs["attention_mask"], image_masks, batch_num_tiles_list + ) + + outputs = self.model( + input_ids=input_ids, + pixel_values=pixel_values, + attention_mask=attention_mask, + output_hidden_states=True, + return_dict=True, + ) + fused_hidden = outputs.hidden_states[-1].to(torch.float32) + valid_mask = attention_mask.to(torch.bool) + if return_cls_only: + # Right-padded causal decoder: the last valid token is the only one that has attended + # to the full image + text prompt. + positions = torch.arange(valid_mask.shape[1], device=valid_mask.device) + last_valid = (valid_mask.long() * positions).argmax(dim=1) + batch_index = torch.arange(fused_hidden.shape[0], device=fused_hidden.device) + return fused_hidden[batch_index, last_valid], None + return fused_hidden, valid_mask + + @property + def device(self) -> torch.device: + return next(self.model.parameters()).device + + +def _flash_attn_available() -> bool: + try: + import flash_attn # noqa: F401 + except ModuleNotFoundError: + return False + return True diff --git a/src/lerobot/policies/evo1/modeling_evo1.py b/src/lerobot/policies/evo1/modeling_evo1.py new file mode 100644 index 000000000..a81a0705a --- /dev/null +++ b/src/lerobot/policies/evo1/modeling_evo1.py @@ -0,0 +1,532 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import builtins +from collections import deque +from contextlib import nullcontext +from pathlib import Path +from typing import TypedDict, Unpack + +import torch +from torch import Tensor + +from lerobot.configs.policies import PreTrainedConfig +from lerobot.policies.pretrained import PreTrainedPolicy, T +from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE + +from ..rtc.modeling_rtc import RTCProcessor +from .configuration_evo1 import Evo1Config +from .evo1_model import Evo1Model + + +class ActionSelectKwargs(TypedDict, total=False): + inference_delay: int | None + prev_chunk_left_over: Tensor | None + execution_horizon: int | None + + +class Evo1Policy(PreTrainedPolicy): + config_class = Evo1Config + name = "evo1" + + def __init__(self, config: Evo1Config, *, vlm_hub_kwargs: dict | None = None, **kwargs): + super().__init__(config) + config.validate_features() + + if len(config.image_features) > config.max_views: + raise ValueError( + f"EVO1 supports at most {config.max_views} camera streams, got {len(config.image_features)}" + ) + + self.config = config + self.model = Evo1Model(config, vlm_hub_kwargs=vlm_hub_kwargs) + self.model.set_finetune_flags() + self._keep_frozen_embedder_eval() + self.init_rtc_processor() + self.reset() + + def init_rtc_processor(self): + """Create the RTC processor when config.rtc_config is set. + + The RTC rollout backend assigns config.rtc_config after loading the policy and re-invokes + this method. + """ + self.rtc_processor = None + if self.config.rtc_config is not None: + self.rtc_processor = RTCProcessor(self.config.rtc_config) + model = getattr(self, "model", None) + if model is not None: + model.rtc_processor = self.rtc_processor + + def _rtc_enabled(self) -> bool: + return self.config.rtc_config is not None and self.config.rtc_config.enabled + + @classmethod + def from_pretrained( + cls: builtins.type[T], + pretrained_name_or_path: str | Path, + *, + config: PreTrainedConfig | None = None, + force_download: bool = False, + resume_download: bool | None = None, + proxies: dict | None = None, + token: str | bool | None = None, + cache_dir: str | Path | None = None, + local_files_only: bool = False, + revision: str | None = None, + strict: bool | None = None, + **kwargs, + ) -> T: + if strict is None: + strict = True + vlm_hub_kwargs = kwargs.pop("vlm_hub_kwargs", None) + if config is None: + config = PreTrainedConfig.from_pretrained( + pretrained_name_or_path=pretrained_name_or_path, + force_download=force_download, + resume_download=resume_download, + proxies=proxies, + token=token, + cache_dir=cache_dir, + local_files_only=local_files_only, + revision=revision, + **kwargs, + ) + if vlm_hub_kwargs is None: + # Forward the hub download options to the base-VLM download as well; `revision` is not + # forwarded because it identifies the policy repo, not the VLM repo. + vlm_hub_kwargs = { + key: value + for key, value in ( + ("token", token), + ("cache_dir", cache_dir), + ("local_files_only", local_files_only), + ("proxies", proxies), + ) + if value not in (None, False) + } + kwargs["vlm_hub_kwargs"] = vlm_hub_kwargs + return super().from_pretrained( + pretrained_name_or_path=pretrained_name_or_path, + config=config, + force_download=force_download, + resume_download=resume_download, + proxies=proxies, + token=token, + cache_dir=cache_dir, + local_files_only=local_files_only, + revision=revision, + strict=strict, + **kwargs, + ) + + @property + def _camera_keys(self) -> list[str]: + return list(self.config.image_features) + + @property + def _env_action_dim(self) -> int: + action_feature = self.config.action_feature + if action_feature is None: + return self.config.max_action_dim + return int(action_feature.shape[0]) + + @property + def _compute_dtype(self) -> torch.dtype: + return next(self.model.action_head.parameters()).dtype + + @property + def _device(self) -> torch.device: + # The device the policy actually lives on. Derived from the parameters rather than + # config.device so the policy keeps working after accelerate (or a plain .to()) moves it. + return next(self.model.action_head.parameters()).device + + @property + def _amp_enabled(self) -> bool: + return bool(self.config.use_amp) and self._device.type == "cuda" + + def _maybe_autocast(self): + # EVO1 manages its own mixed precision: an explicit bf16 autocast that also overrides any + # outer autocast context (e.g. lerobot-eval's fp16 default), keeping train and eval + # numerics identical. + if self._amp_enabled: + return torch.autocast(device_type="cuda", dtype=torch.bfloat16) + return nullcontext() + + def get_optim_params(self) -> list[dict]: + decay, no_decay = [], [] + for name, param in self.named_parameters(): + if not param.requires_grad: + continue + is_bias = name.endswith("bias") or ".bias" in name + is_norm = param.dim() == 1 or "norm" in name.lower() + if is_bias or is_norm: + no_decay.append(param) + else: + decay.append(param) + return [ + {"params": decay, "weight_decay": self.config.optimizer_weight_decay}, + {"params": no_decay, "weight_decay": 0.0}, + ] + + def reset(self): + self._action_queue = deque([], maxlen=self.config.n_action_steps) + + def _normalize_task_batch(self, batch: dict[str, Tensor | list[str] | str]) -> list[str]: + prompts = batch.get(self.config.task_field) + if prompts is None and self.config.task_field != "task": + prompts = batch.get("task") + if prompts is None: + raise ValueError(f"EVO1 expects a '{self.config.task_field}' text field in the batch.") + if isinstance(prompts, str): + return [prompts] + if isinstance(prompts, (list, tuple)): + return [str(prompt) for prompt in prompts] + raise TypeError(f"Unsupported prompt batch type: {type(prompts)}") + + def _prepare_state(self, batch: dict[str, Tensor]) -> tuple[Tensor, Tensor]: + if OBS_STATE not in batch: + raise ValueError(f"EVO1 requires '{OBS_STATE}' in the batch.") + state = batch[OBS_STATE] + if state.dim() == 1: + state = state.unsqueeze(0) + elif state.dim() == 3: + state = state[:, -1] + elif state.dim() != 2: + raise ValueError(f"Unsupported state tensor shape for EVO1: {tuple(state.shape)}") + batch_size, state_dim = state.shape + if state_dim > self.config.max_state_dim: + raise ValueError( + f"State dim {state_dim} exceeds configured max_state_dim {self.config.max_state_dim}" + ) + explicit_mask = batch.get("state_mask") + if explicit_mask is not None: + if explicit_mask.dim() == 1: + explicit_mask = explicit_mask.unsqueeze(0) + elif explicit_mask.dim() == 3: + explicit_mask = explicit_mask[:, -1] + elif explicit_mask.dim() != 2: + raise ValueError( + f"Unsupported state_mask tensor shape for EVO1: {tuple(explicit_mask.shape)}" + ) + if explicit_mask.shape != (batch_size, state_dim): + raise ValueError( + f"state_mask shape {tuple(explicit_mask.shape)} does not match state shape {(batch_size, state_dim)}" + ) + device = self._device + padded = torch.zeros( + batch_size, + self.config.max_state_dim, + dtype=state.dtype, + device=device, + ) + padded[:, :state_dim] = state.to(device=device) + mask = torch.zeros( + batch_size, + self.config.max_state_dim, + dtype=torch.bool, + device=device, + ) + if explicit_mask is None: + mask[:, :state_dim] = True + else: + mask[:, :state_dim] = explicit_mask.to(device=device, dtype=torch.bool) + # Zero out masked state dims so an explicit state_mask actually affects the model input + # (the state encoder has no mask argument of its own). + padded = padded * mask.to(dtype=padded.dtype) + return padded.to(dtype=self._compute_dtype), mask + + def _prepare_actions(self, batch: dict[str, Tensor]) -> tuple[Tensor, Tensor]: + if ACTION not in batch: + raise ValueError(f"EVO1 requires '{ACTION}' in the batch for training.") + action = batch[ACTION] + if action.dim() == 2: + action = action.unsqueeze(1) + batch_size, horizon, action_dim = action.shape + if horizon != self.config.chunk_size: + raise ValueError( + f"EVO1 expects chunk_size={self.config.chunk_size}, got action horizon {horizon}" + ) + if action_dim > self.config.max_action_dim: + raise ValueError( + f"Action dim {action_dim} exceeds configured max_action_dim {self.config.max_action_dim}" + ) + explicit_mask = batch.get("action_mask") + if explicit_mask is not None: + if explicit_mask.dim() == 2: + if horizon == 1: + explicit_mask = explicit_mask.unsqueeze(1) + else: + raise ValueError( + f"2D action_mask is only supported when chunk_size=1, got action horizon {horizon}" + ) + elif explicit_mask.dim() != 3: + raise ValueError( + f"Unsupported action_mask tensor shape for EVO1: {tuple(explicit_mask.shape)}" + ) + if explicit_mask.shape != (batch_size, horizon, action_dim): + raise ValueError( + "action_mask shape " + f"{tuple(explicit_mask.shape)} does not match action shape {(batch_size, horizon, action_dim)}" + ) + device = self._device + padded = torch.zeros( + batch_size, + horizon, + self.config.max_action_dim, + dtype=action.dtype, + device=device, + ) + padded[:, :, :action_dim] = action.to(device=device) + mask = torch.zeros( + batch_size, + horizon, + self.config.max_action_dim, + dtype=torch.bool, + device=device, + ) + if explicit_mask is None: + mask[:, :, :action_dim] = True + else: + mask[:, :, :action_dim] = explicit_mask.to(device=device, dtype=torch.bool) + + # Timesteps beyond the episode end hold fabricated (repeated) actions; exclude them from + # the loss like the other chunked policies do. + action_is_pad = batch.get("action_is_pad") + if action_is_pad is not None: + if action_is_pad.shape != (batch_size, horizon): + raise ValueError( + f"action_is_pad shape {tuple(action_is_pad.shape)} does not match " + f"(batch_size, chunk_size)={(batch_size, horizon)}" + ) + in_episode = ~action_is_pad.to(device=device, dtype=torch.bool) + mask = mask & in_episode.unsqueeze(-1) + return padded.to(dtype=self._compute_dtype), mask + + def _prepare_inference_action_mask(self, batch_size: int) -> Tensor: + mask = torch.zeros( + batch_size, + self.config.max_action_dim, + dtype=torch.bool, + device=self._device, + ) + mask[:, : self._env_action_dim] = True + return mask + + def _get_embodiment_ids(self, batch: dict[str, Tensor], batch_size: int) -> Tensor: + embodiment_ids = batch.get("embodiment_id") + if embodiment_ids is None and self.config.embodiment_id_field: + embodiment_ids = batch.get(self.config.embodiment_id_field) + if embodiment_ids is None: + return torch.full( + (batch_size,), + self.config.default_embodiment_id, + dtype=torch.long, + device=self._device, + ) + if embodiment_ids.dim() == 0: + embodiment_ids = embodiment_ids.unsqueeze(0) + elif embodiment_ids.dim() > 1: + embodiment_ids = embodiment_ids[:, -1] + return embodiment_ids.to(device=self._device, dtype=torch.long) + + @property + def _tracks_vlm_gradients(self) -> bool: + return bool( + self.config.finetune_vlm + or self.config.finetune_language_model + or self.config.finetune_vision_model + ) + + def _keep_frozen_embedder_eval(self) -> None: + if self._tracks_vlm_gradients: + return + embedder = getattr(self.model, "embedder", None) + if embedder is not None: + embedder.eval() + + def train(self, mode: bool = True): + super().train(mode) + self._keep_frozen_embedder_eval() + return self + + def _collect_image_batches(self, batch: dict[str, Tensor]) -> tuple[list[Tensor], Tensor]: + camera_keys = self._camera_keys or sorted(key for key in batch if key.startswith(f"{OBS_IMAGES}.")) + if not camera_keys: + raise ValueError("EVO1 requires at least one visual observation feature.") + camera_keys = list(camera_keys)[: self.config.max_views] + + # Configured cameras may be absent from the batch up to the empty_cameras budget (e.g. the + # placeholder features added by validate_features); they become masked-out views that the + # embedder zero-pads. Any other absent camera is an error. + present_keys = [key for key in camera_keys if key in batch] + missing_keys = [key for key in camera_keys if key not in batch] + if len(missing_keys) > self.config.empty_cameras: + raise ValueError( + f"Missing camera features {missing_keys} in batch; at most " + f"empty_cameras={self.config.empty_cameras} may be absent." + ) + if not present_keys: + raise ValueError("EVO1 requires at least one visual observation in the batch.") + + # Keep each present camera as a batched (B, C, H, W) tensor on its current (GPU) device. + # Resizing/normalization and zero-padding of absent views happen batched inside the + # embedder, so images never leave the device here. + camera_images: list[Tensor] = [] + for camera_key in present_keys: + image = batch[camera_key] + if image.dim() == 3: + # Promote an unbatched (C, H, W) frame so batch_size is read from a real batch dim. + image = image.unsqueeze(0) + elif image.dim() == 5: + image = image[:, -1] + elif image.dim() != 4: + raise ValueError( + f"Unsupported image tensor shape for EVO1: key={camera_key} shape={tuple(image.shape)}" + ) + camera_images.append(image) + + batch_size = camera_images[0].shape[0] + n_present = len(camera_images) + image_masks = torch.zeros( + batch_size, self.config.max_views, dtype=torch.bool, device=camera_images[0].device + ) + image_masks[:, :n_present] = True + + return camera_images, image_masks + + def _compute_fused_tokens( + self, + prompts: list[str], + image_batches: list[Tensor], + image_masks: Tensor, + ) -> tuple[Tensor, Tensor | None]: + track_vlm_gradients = self._tracks_vlm_gradients + grad_context = nullcontext() if track_vlm_gradients else torch.no_grad() + with grad_context: + fused_tokens, context_mask = self.model.get_vl_embeddings( + images=image_batches, + image_mask=image_masks, + prompt=prompts, + return_cls_only=self.config.return_cls_only, + ) + + if not track_vlm_gradients: + fused_tokens = fused_tokens.detach() + fused_tokens = fused_tokens.to(device=self._device, dtype=self._compute_dtype) + if context_mask is not None: + context_mask = context_mask.to(device=self._device) + return fused_tokens, context_mask + + def _compute_masked_loss( + self, + pred_velocity: Tensor, + target_velocity: Tensor, + action_mask: Tensor, + reduction: str, + ) -> Tensor: + flat_mask = action_mask.view(action_mask.shape[0], -1).to(dtype=pred_velocity.dtype) + sq_error = ((pred_velocity - target_velocity) * flat_mask).pow(2) + active = flat_mask.sum(dim=1).clamp_min(1.0) + per_sample_loss = sq_error.sum(dim=1) / active + if reduction == "none": + return per_sample_loss + if reduction != "mean": + raise ValueError(f"Unsupported reduction '{reduction}'") + return sq_error.sum() / active.sum() + + def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict]: + prompts = self._normalize_task_batch(batch) + image_batches, image_masks = self._collect_image_batches(batch) + states, _state_mask = self._prepare_state(batch) + actions_gt, action_mask = self._prepare_actions(batch) + embodiment_ids = self._get_embodiment_ids(batch, states.shape[0]) + + with self._maybe_autocast(): + fused_tokens, context_mask = self._compute_fused_tokens(prompts, image_batches, image_masks) + pred_velocity, noise = self.model( + fused_tokens, + state=states, + actions_gt=actions_gt, + action_mask=action_mask.to(device=self._device, dtype=self._compute_dtype), + embodiment_ids=embodiment_ids, + context_mask=context_mask, + ) + + # Compute the flow-matching regression loss in fp32, outside the autocast block. + pred_velocity = pred_velocity.float() + noise = noise.float() + flat_action_mask = action_mask.view(action_mask.shape[0], -1).to(dtype=torch.float32) + # Flow-matching velocity target. Padded (masked-out) action dims are already zero on both sides + # here (`actions_gt` is zero-padded in `_prepare_actions`, and `noise` is masked inside the head), + # and the whole difference is multiplied by `flat_action_mask`, so padded dims contribute nothing. + target_velocity = (actions_gt.float() - noise).view(actions_gt.shape[0], -1) * flat_action_mask + loss = self._compute_masked_loss(pred_velocity, target_velocity, action_mask, reduction) + loss_mean = loss.mean().item() if loss.ndim > 0 else loss.item() + return loss, { + "loss": loss_mean, + "active_action_dims": float(action_mask.sum(dim=(1, 2)).float().mean().item()), + } + + @torch.no_grad() + def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor: + inference_delay = kwargs.get("inference_delay") + prev_chunk_left_over = kwargs.get("prev_chunk_left_over") + execution_horizon = kwargs.get("execution_horizon") + if (inference_delay is not None or prev_chunk_left_over is not None) and not self._rtc_enabled(): + raise RuntimeError( + "Received RTC arguments but RTC is not configured for this EVO1 policy: set " + "config.rtc_config and call init_rtc_processor() (lerobot-rollout does this for " + "--inference.type=rtc)." + ) + self.eval() + + prompts = self._normalize_task_batch(batch) + image_batches, image_masks = self._collect_image_batches(batch) + states, _state_mask = self._prepare_state(batch) + embodiment_ids = self._get_embodiment_ids(batch, states.shape[0]) + action_mask = self._prepare_inference_action_mask(states.shape[0]) + if prev_chunk_left_over is not None: + prev_chunk_left_over = prev_chunk_left_over.to(device=self._device) + + with self._maybe_autocast(): + fused_tokens, context_mask = self._compute_fused_tokens(prompts, image_batches, image_masks) + actions = self.model( + fused_tokens, + state=states, + action_mask=action_mask, + embodiment_ids=embodiment_ids, + context_mask=context_mask, + inference_delay=inference_delay, + prev_chunk_left_over=prev_chunk_left_over, + execution_horizon=execution_horizon, + ) + actions = actions.view(states.shape[0], self.config.chunk_size, self.config.max_action_dim) + return actions.to(dtype=torch.float32) + + @torch.no_grad() + def select_action(self, batch: dict[str, Tensor], **kwargs) -> Tensor: + assert not self._rtc_enabled(), ( + "RTC is not supported for select_action, use it with predict_action_chunk" + ) + self.eval() + if len(self._action_queue) == 0: + action_chunk = self.predict_action_chunk(batch)[:, : self.config.n_action_steps] + self._action_queue.extend(action_chunk.transpose(0, 1)) + # Returns one step of shape (B, max_action_dim): actions are emitted at the padded max_action_dim + # width and cropped to the real action dim downstream by the postprocessor (Evo1ActionProcessorStep). + # Callers that bypass the postprocessor receive the padded width. + return self._action_queue.popleft() diff --git a/src/lerobot/policies/evo1/processor_evo1.py b/src/lerobot/policies/evo1/processor_evo1.py new file mode 100644 index 000000000..adff8b66a --- /dev/null +++ b/src/lerobot/policies/evo1/processor_evo1.py @@ -0,0 +1,400 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from copy import deepcopy +from dataclasses import dataclass +from typing import Any + +import torch + +from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature +from lerobot.processor import ( + AddBatchDimensionProcessorStep, + DeviceProcessorStep, + NormalizerProcessorStep, + ObservationProcessorStep, + PolicyAction, + PolicyActionProcessorStep, + PolicyProcessorPipeline, + ProcessorStep, + ProcessorStepRegistry, + RenameObservationsProcessorStep, + UnnormalizerProcessorStep, +) +from lerobot.processor.converters import ( + batch_to_transition, + create_transition, + policy_action_to_transition, + transition_to_policy_action, +) +from lerobot.types import EnvTransition, TransitionKey +from lerobot.utils.constants import ( + ACTION, + DONE, + INFO, + OBS_PREFIX, + OBS_STATE, + POLICY_POSTPROCESSOR_DEFAULT_NAME, + POLICY_PREPROCESSOR_DEFAULT_NAME, + REWARD, + TRUNCATED, +) + +from .configuration_evo1 import Evo1Config + + +def evo1_batch_to_transition(batch: dict[str, Any]): + transition = batch_to_transition(batch) + complementary_data = dict(transition.get("complementary_data") or {}) + reserved = {ACTION, REWARD, DONE, TRUNCATED, INFO} + for key, value in batch.items(): + if key in reserved or key.startswith(OBS_PREFIX): + continue + complementary_data.setdefault(key, value) + return create_transition( + observation=transition.get("observation"), + action=transition.get("action"), + reward=transition.get("reward", 0.0), + done=transition.get("done", False), + truncated=transition.get("truncated", False), + info=transition.get("info", {}), + complementary_data=complementary_data, + ) + + +@dataclass +@ProcessorStepRegistry.register(name="evo1_pad_state_processor") +class Evo1PadStateProcessorStep(ObservationProcessorStep): + """Pad policy observations to EVO1's fixed state width before normalization.""" + + max_state_dim: int = 24 + + def observation(self, observation: dict[str, Any]) -> dict[str, Any]: + if OBS_STATE not in observation: + return observation + + state = observation[OBS_STATE] + state_dim = state.shape[-1] + if state_dim > self.max_state_dim: + raise ValueError( + f"EVO1 state has {state_dim} dims, which exceeds max_state_dim={self.max_state_dim}." + ) + if state_dim < self.max_state_dim: + observation = observation.copy() + observation[OBS_STATE] = torch.nn.functional.pad(state, (0, self.max_state_dim - state_dim)) + return observation + + def transform_features( + self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]] + ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]: + new_features = {ft: feats.copy() for ft, feats in features.items()} + obs_feats = new_features.setdefault(PipelineFeatureType.OBSERVATION, {}) + if OBS_STATE in obs_feats: + obs_feats[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(self.max_state_dim,)) + return new_features + + def get_config(self) -> dict[str, Any]: + return {"max_state_dim": self.max_state_dim} + + +@dataclass +@ProcessorStepRegistry.register(name="evo1_pad_action_processor") +class Evo1PadActionProcessorStep(ProcessorStep): + """Pad training actions and preserve the active action dimensions with action_mask.""" + + max_action_dim: int = 24 + + def __call__(self, transition: EnvTransition) -> EnvTransition: + action = transition.get(TransitionKey.ACTION) + if action is None: + return transition + if not isinstance(action, PolicyAction): + raise ValueError(f"EVO1 action should be a PolicyAction tensor, but got {type(action)}.") + + action_dim = action.shape[-1] + if action_dim > self.max_action_dim: + raise ValueError( + f"EVO1 action has {action_dim} dims, which exceeds max_action_dim={self.max_action_dim}." + ) + + new_transition = transition.copy() + new_action = action + if action_dim < self.max_action_dim: + new_action = torch.nn.functional.pad(action, (0, self.max_action_dim - action_dim)) + + complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}) + action_mask = complementary_data.get("action_mask") + if action_mask is None: + action_mask = torch.ones(action.shape, dtype=torch.bool, device=action.device) + else: + action_mask = torch.as_tensor(action_mask, dtype=torch.bool, device=action.device) + if action_mask.shape != action.shape: + raise ValueError( + f"action_mask shape {tuple(action_mask.shape)} does not match action shape {tuple(action.shape)}." + ) + if action_dim < self.max_action_dim: + action_mask = torch.nn.functional.pad(action_mask, (0, self.max_action_dim - action_dim)) + + complementary_data["action_mask"] = action_mask + new_transition[TransitionKey.ACTION] = new_action + new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data + return new_transition + + def transform_features( + self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]] + ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]: + new_features = {ft: feats.copy() for ft, feats in features.items()} + action_feats = new_features.setdefault(PipelineFeatureType.ACTION, {}) + action_feats[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(self.max_action_dim,)) + return new_features + + def get_config(self) -> dict[str, Any]: + return {"max_action_dim": self.max_action_dim} + + +@dataclass +@ProcessorStepRegistry.register(name="evo1_action_processor") +class Evo1ActionProcessorStep(PolicyActionProcessorStep): + """Crop padded EVO1 actions and optionally binarize the LIBERO gripper channel.""" + + action_dim: int + binarize_gripper: bool = False + gripper_index: int = 6 + gripper_threshold: float = 0.5 + gripper_below_threshold_value: float = 1.0 + gripper_above_threshold_value: float = -1.0 + + def action(self, action: PolicyAction) -> PolicyAction: + if action.shape[-1] < self.action_dim: + raise ValueError( + f"EVO1 action has {action.shape[-1]} dims, which is smaller than action_dim={self.action_dim}." + ) + + action = action[..., : self.action_dim] + if not self.binarize_gripper: + return action + + if not 0 <= self.gripper_index < self.action_dim: + raise ValueError( + f"gripper_index={self.gripper_index} must be within action_dim={self.action_dim}." + ) + + action = action.clone() + below = torch.as_tensor( + self.gripper_below_threshold_value, + dtype=action.dtype, + device=action.device, + ) + above = torch.as_tensor( + self.gripper_above_threshold_value, + dtype=action.dtype, + device=action.device, + ) + action[..., self.gripper_index] = torch.where( + action[..., self.gripper_index] > self.gripper_threshold, + above, + below, + ) + return action + + def transform_features( + self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]] + ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]: + new_features = {ft: feats.copy() for ft, feats in features.items()} + action_feats = new_features.setdefault(PipelineFeatureType.ACTION, {}) + action_feats[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(self.action_dim,)) + return new_features + + def get_config(self) -> dict[str, Any]: + return { + "action_dim": self.action_dim, + "binarize_gripper": self.binarize_gripper, + "gripper_index": self.gripper_index, + "gripper_threshold": self.gripper_threshold, + "gripper_below_threshold_value": self.gripper_below_threshold_value, + "gripper_above_threshold_value": self.gripper_above_threshold_value, + } + + +def _evo1_action_dim(config: Evo1Config) -> int: + if config.postprocess_action_dim is not None: + return config.postprocess_action_dim + action_feature = config.action_feature + if action_feature is None: + return config.max_action_dim + return int(action_feature.shape[0]) + + +def _evo1_normalization_features(config: Evo1Config) -> dict[str, PolicyFeature]: + features = {**config.input_features, **config.output_features} + features[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(config.max_state_dim,)) + features[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(config.max_action_dim,)) + return features + + +def _evo1_action_features(config: Evo1Config) -> dict[str, PolicyFeature]: + return {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(config.max_action_dim,))} + + +_STAT_PAD_VALUES = { + "mean": 0.0, + "std": 1.0, + "min": -1.0, + "max": 1.0, + "q01": -1.0, + "q99": 1.0, + "q10": -1.0, + "q90": 1.0, +} + + +def _pad_stat_value(value: Any, target_dim: int, stat_name: str) -> torch.Tensor: + tensor = torch.as_tensor(value) + if not tensor.is_floating_point(): + tensor = tensor.to(dtype=torch.float32) + if tensor.ndim == 0 or tensor.shape[-1] >= target_dim: + return tensor + + pad_shape = (*tensor.shape[:-1], target_dim - tensor.shape[-1]) + pad_value = _STAT_PAD_VALUES.get(stat_name, 0.0) + padding = torch.full(pad_shape, pad_value, dtype=tensor.dtype, device=tensor.device) + return torch.cat([tensor, padding], dim=-1) + + +def _pad_feature_stats( + stats: dict[str, dict[str, Any]], + feature_key: str, + target_dim: int, +) -> None: + if feature_key not in stats: + return + stats[feature_key] = { + stat_name: _pad_stat_value(stat_value, target_dim, stat_name) + for stat_name, stat_value in stats[feature_key].items() + } + + +def _pad_evo1_stats( + config: Evo1Config, + stats: dict[str, dict[str, Any]] | None, +) -> dict[str, dict[str, Any]] | None: + if stats is None: + return None + + padded_stats = deepcopy(stats) + # Added dimensions represent zero-padding inside EVO1. These neutral stats keep + # padded observations at normalized zero and only provide shape compatibility. + _pad_feature_stats(padded_stats, OBS_STATE, config.max_state_dim) + _pad_feature_stats(padded_stats, ACTION, config.max_action_dim) + return padded_stats + + +def reconcile_evo1_processors( + config: Evo1Config, + preprocessor: PolicyProcessorPipeline, + postprocessor: PolicyProcessorPipeline, +) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]: + """Reconcile checkpoint-loaded pipelines with the current EVO1 config. + + Two things cannot be restored from a serialized pipeline alone: the EVO1 batch converter + (converters are plain functions and are never serialized), and eval-time CLI overrides of the + action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`). This + restores the converter and rebuilds the action step from the current config so those overrides + take effect. + """ + # Pipelines reloaded from a checkpoint come back with the default batch converter, which drops + # non-observation extras (embodiment_id, state_mask, custom task fields) needed by EVO1. + preprocessor.to_transition = evo1_batch_to_transition + + action_step = Evo1ActionProcessorStep( + action_dim=_evo1_action_dim(config), + binarize_gripper=config.binarize_gripper, + gripper_index=config.gripper_index, + gripper_threshold=config.gripper_threshold, + gripper_below_threshold_value=config.gripper_below_threshold_value, + gripper_above_threshold_value=config.gripper_above_threshold_value, + ) + steps = list(postprocessor.steps) + action_step_idx = next( + (idx for idx, step in enumerate(steps) if isinstance(step, Evo1ActionProcessorStep)), None + ) + if action_step_idx is None: + insert_idx = next( + (idx + 1 for idx, step in enumerate(steps) if isinstance(step, UnnormalizerProcessorStep)), + 0, + ) + steps.insert(insert_idx, action_step) + else: + steps[action_step_idx] = action_step + postprocessor.steps = steps + + return preprocessor, postprocessor + + +def make_evo1_pre_post_processors( + config: Evo1Config, + dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None, +) -> tuple[ + PolicyProcessorPipeline[dict[str, Any], dict[str, Any]], + PolicyProcessorPipeline[PolicyAction, PolicyAction], +]: + normalization_features = _evo1_normalization_features(config) + action_features = _evo1_action_features(config) + normalization_stats = _pad_evo1_stats(config, dataset_stats) + + input_steps = [ + RenameObservationsProcessorStep(rename_map={}), + AddBatchDimensionProcessorStep(), + Evo1PadStateProcessorStep(max_state_dim=config.max_state_dim), + Evo1PadActionProcessorStep(max_action_dim=config.max_action_dim), + NormalizerProcessorStep( + features=normalization_features, + norm_map=config.normalization_mapping, + stats=normalization_stats, + ), + DeviceProcessorStep(device=config.device), + ] + output_steps = [ + UnnormalizerProcessorStep( + features=action_features, + norm_map=config.normalization_mapping, + stats=normalization_stats, + ), + Evo1ActionProcessorStep( + action_dim=_evo1_action_dim(config), + binarize_gripper=config.binarize_gripper, + gripper_index=config.gripper_index, + gripper_threshold=config.gripper_threshold, + gripper_below_threshold_value=config.gripper_below_threshold_value, + gripper_above_threshold_value=config.gripper_above_threshold_value, + ), + # float32 so downstream numpy conversion works even when the policy computes in bf16. + DeviceProcessorStep(device="cpu", float_dtype="float32"), + ] + + return ( + PolicyProcessorPipeline[dict[str, Any], dict[str, Any]]( + steps=input_steps, + name=POLICY_PREPROCESSOR_DEFAULT_NAME, + to_transition=evo1_batch_to_transition, + ), + PolicyProcessorPipeline[PolicyAction, PolicyAction]( + steps=output_steps, + name=POLICY_POSTPROCESSOR_DEFAULT_NAME, + to_transition=policy_action_to_transition, + to_output=transition_to_policy_action, + ), + ) diff --git a/src/lerobot/policies/factory.py b/src/lerobot/policies/factory.py index a42b38ba4..73fd9455f 100644 --- a/src/lerobot/policies/factory.py +++ b/src/lerobot/policies/factory.py @@ -47,8 +47,11 @@ from lerobot.utils.feature_utils import dataset_to_policy_features from .act.configuration_act import ACTConfig from .diffusion.configuration_diffusion import DiffusionConfig from .eo1.configuration_eo1 import EO1Config +from .evo1.configuration_evo1 import Evo1Config +from .fastwam.configuration_fastwam import FastWAMConfig from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig from .groot.configuration_groot import GrootConfig +from .lingbot_va.configuration_lingbot_va import LingBotVAConfig from .molmoact2.configuration_molmoact2 import MolmoAct2Config from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig from .pi0.configuration_pi0 import PI0Config @@ -91,7 +94,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]: Args: name: The name of the policy. Supported names are "tdmpc", "diffusion", "act", "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x", - "molmoact2". + "molmoact2", "eo1", "evo1". Returns: The policy class corresponding to the given name. @@ -162,6 +165,18 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]: from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy return VLAJEPAPolicy + elif name == "lingbot_va": + from .lingbot_va.modeling_lingbot_va import LingBotVAPolicy + + return LingBotVAPolicy + elif name == "fastwam": + from .fastwam.modeling_fastwam import FastWAMPolicy + + return FastWAMPolicy + elif name == "evo1": + from .evo1.modeling_evo1 import Evo1Policy + + return Evo1Policy else: try: return _get_policy_cls_from_policy_name(name=name) @@ -179,7 +194,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig: Args: policy_type: The type of the policy. Supported types include "tdmpc", "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor", - "smolvla", "wall_x", "molmoact2". + "smolvla", "wall_x", "molmoact2", "eo1", "evo1". **kwargs: Keyword arguments to be passed to the configuration class constructor. Returns: @@ -218,6 +233,12 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig: return MolmoAct2Config(**kwargs) elif policy_type == "vla_jepa": return VLAJEPAConfig(**kwargs) + elif policy_type == "lingbot_va": + return LingBotVAConfig(**kwargs) + elif policy_type == "fastwam": + return FastWAMConfig(**kwargs) + elif policy_type == "evo1": + return Evo1Config(**kwargs) else: try: config_cls = PreTrainedConfig.get_choice_class(policy_type) @@ -252,6 +273,7 @@ class ProcessorConfigKwargs(TypedDict, total=False): def make_pre_post_processors( policy_cfg: PreTrainedConfig, pretrained_path: str | None = None, + pretrained_revision: str | None = None, **kwargs: Unpack[ProcessorConfigKwargs], ) -> tuple[ PolicyProcessorPipeline[dict[str, Any], dict[str, Any]], @@ -280,26 +302,23 @@ def make_pre_post_processors( policy configuration type. """ if pretrained_path: - # TODO(Steven): Temporary patch, implement correctly the processors for Gr00t if isinstance(policy_cfg, GrootConfig): - # GROOT handles normalization in groot_pack_inputs_v3 step - # Need to override both stats AND normalize_min_max since saved config might be empty - preprocessor_overrides = {} - postprocessor_overrides = {} - preprocessor_overrides["groot_pack_inputs_v3"] = { - "stats": kwargs.get("dataset_stats"), - "normalize_min_max": True, - } + from .groot.processor_groot import make_groot_pre_post_processors_from_pretrained - # Also ensure postprocessing slices to env action dim and unnormalizes with dataset stats - env_action_dim = policy_cfg.output_features[ACTION].shape[0] - postprocessor_overrides["groot_action_unpack_unnormalize_v1"] = { - "stats": kwargs.get("dataset_stats"), - "normalize_min_max": True, - "env_action_dim": env_action_dim, - } - kwargs["preprocessor_overrides"] = preprocessor_overrides - kwargs["postprocessor_overrides"] = postprocessor_overrides + return make_groot_pre_post_processors_from_pretrained( + config=policy_cfg, + pretrained_path=pretrained_path, + dataset_stats=kwargs.get("dataset_stats"), + dataset_meta=kwargs.get("dataset_meta"), + preprocessor_overrides=kwargs.get("preprocessor_overrides"), + postprocessor_overrides=kwargs.get("postprocessor_overrides"), + preprocessor_config_filename=kwargs.get( + "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json" + ), + postprocessor_config_filename=kwargs.get( + "postprocessor_config_filename", f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json" + ), + ) preprocessor = PolicyProcessorPipeline.from_pretrained( pretrained_model_name_or_path=pretrained_path, @@ -309,6 +328,7 @@ def make_pre_post_processors( overrides=kwargs.get("preprocessor_overrides", {}), to_transition=batch_to_transition, to_output=transition_to_batch, + revision=pretrained_revision, ) postprocessor = PolicyProcessorPipeline.from_pretrained( pretrained_model_name_or_path=pretrained_path, @@ -318,8 +338,17 @@ def make_pre_post_processors( overrides=kwargs.get("postprocessor_overrides", {}), to_transition=policy_action_to_transition, to_output=transition_to_policy_action, + revision=pretrained_revision, ) _reconnect_relative_absolute_steps(preprocessor, postprocessor) + if isinstance(policy_cfg, Evo1Config): + from .evo1.processor_evo1 import reconcile_evo1_processors + + preprocessor, postprocessor = reconcile_evo1_processors( + policy_cfg, + preprocessor, + postprocessor, + ) return preprocessor, postprocessor # Create a new processor based on policy type @@ -403,6 +432,7 @@ def make_pre_post_processors( processors = make_groot_pre_post_processors( config=policy_cfg, dataset_stats=kwargs.get("dataset_stats"), + dataset_meta=kwargs.get("dataset_meta"), ) elif isinstance(policy_cfg, XVLAConfig): @@ -430,6 +460,13 @@ def make_pre_post_processors( config=policy_cfg, dataset_stats=kwargs.get("dataset_stats"), ) + elif isinstance(policy_cfg, Evo1Config): + from .evo1.processor_evo1 import make_evo1_pre_post_processors + + processors = make_evo1_pre_post_processors( + config=policy_cfg, + dataset_stats=kwargs.get("dataset_stats"), + ) elif isinstance(policy_cfg, MolmoAct2Config): from .molmoact2.processor_molmoact2 import make_molmoact2_pre_post_processors @@ -448,6 +485,22 @@ def make_pre_post_processors( dataset_stats=kwargs.get("dataset_stats"), ) + elif isinstance(policy_cfg, LingBotVAConfig): + from .lingbot_va.processor_lingbot_va import make_lingbot_va_pre_post_processors + + processors = make_lingbot_va_pre_post_processors( + config=policy_cfg, + dataset_stats=kwargs.get("dataset_stats"), + ) + + elif isinstance(policy_cfg, FastWAMConfig): + from .fastwam.processor_fastwam import make_fastwam_pre_post_processors + + processors = make_fastwam_pre_post_processors( + config=policy_cfg, + dataset_stats=kwargs.get("dataset_stats"), + ) + else: try: processors = _make_processors_from_policy_config( @@ -537,6 +590,7 @@ def make_policy( set_dataset_feature_metadata = getattr(cfg, "set_dataset_feature_metadata", None) if callable(set_dataset_feature_metadata): set_dataset_feature_metadata(ds_meta.features) + cfg._runtime_dataset_meta = ds_meta kwargs["config"] = cfg @@ -557,6 +611,7 @@ def make_policy( # Load a pretrained policy and override the config if needed (for example, if there are inference-time # hyperparameters that we want to vary). kwargs["pretrained_name_or_path"] = cfg.pretrained_path + kwargs["revision"] = cfg.pretrained_revision policy = policy_cls.from_pretrained(**kwargs) elif cfg.pretrained_path and cfg.use_peft: # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo diff --git a/src/lerobot/policies/fastwam/README.md b/src/lerobot/policies/fastwam/README.md new file mode 120000 index 000000000..d78b9ef36 --- /dev/null +++ b/src/lerobot/policies/fastwam/README.md @@ -0,0 +1 @@ +../../../../docs/source/policy_fastwam_README.md \ No newline at end of file diff --git a/src/lerobot/policies/fastwam/__init__.py b/src/lerobot/policies/fastwam/__init__.py new file mode 100644 index 000000000..8488e7b78 --- /dev/null +++ b/src/lerobot/policies/fastwam/__init__.py @@ -0,0 +1,23 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from .configuration_fastwam import FastWAMConfig +from .modeling_fastwam import FastWAMPolicy +from .processor_fastwam import make_fastwam_pre_post_processors + +__all__ = [ + "FastWAMConfig", + "FastWAMPolicy", + "make_fastwam_pre_post_processors", +] diff --git a/src/lerobot/policies/fastwam/configuration_fastwam.py b/src/lerobot/policies/fastwam/configuration_fastwam.py new file mode 100644 index 000000000..a3ef4f602 --- /dev/null +++ b/src/lerobot/policies/fastwam/configuration_fastwam.py @@ -0,0 +1,399 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass, field +from pathlib import Path +from typing import Any + +from lerobot.configs import ( + FeatureType, + NormalizationMode, + PolicyFeature, + PreTrainedConfig, +) +from lerobot.optim import AdamWConfig +from lerobot.utils.constants import ACTION, OBS_STATE + +WAN22_MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B" +WAN22_DIFFUSERS_MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B-Diffusers" +FASTWAM_BASE_MODEL_ID = "lerobot/fastwam_base" +WAN_T5_TOKENIZER_ID = "google/umt5-xxl" + + +_FASTWAM_VIDEO_BASE_COMPAT_KEYS = ( + "patch_size", + "in_dim", + "hidden_dim", + "ffn_dim", + "freq_dim", + "text_dim", + "out_dim", + "num_heads", + "attn_head_dim", + "num_layers", +) + +_FASTWAM_ACTION_BASE_COMPAT_KEYS = ( + "hidden_dim", + "ffn_dim", + "num_heads", + "attn_head_dim", + "num_layers", + "text_dim", + "freq_dim", +) + + +def default_video_dit_config(action_dim: int) -> dict[str, Any]: + return { + "patch_size": [1, 2, 2], + "in_dim": 48, + "hidden_dim": 3072, + "ffn_dim": 14336, + "freq_dim": 256, + "text_dim": 4096, + "out_dim": 48, + "num_heads": 24, + "attn_head_dim": 128, + "num_layers": 30, + "eps": 1.0e-6, + "seperated_timestep": True, + "use_gradient_checkpointing": False, + "video_attention_mask_mode": "first_frame_causal", + "action_conditioned": False, + "action_dim": action_dim, + "action_group_causal_mask_mode": "group_diagonal", + "fp32_attention": True, + } + + +def default_action_dit_config(action_dim: int) -> dict[str, Any]: + return { + "action_dim": action_dim, + "hidden_dim": 1024, + "ffn_dim": 4096, + "num_heads": 24, + "attn_head_dim": 128, + "num_layers": 30, + "text_dim": 4096, + "freq_dim": 256, + "eps": 1.0e-6, + "use_gradient_checkpointing": False, + "fp32_attention": True, + } + + +def _coerce_enum(enum_cls: type, value: Any) -> Any: + if isinstance(value, enum_cls): + return value + try: + return enum_cls(value) + except (TypeError, ValueError) as exc: + member = getattr(enum_cls, str(value), None) + if member is None: + raise ValueError(f"Cannot coerce {value!r} into {enum_cls.__name__}.") from exc + return member + + +def _coerce_policy_features(features: dict[str, Any] | None) -> dict[str, PolicyFeature] | None: + if features is None: + return None + coerced = {} + for name, feature in features.items(): + if isinstance(feature, PolicyFeature): + coerced[name] = feature + continue + coerced[name] = PolicyFeature( + type=_coerce_enum(FeatureType, feature["type"]), + shape=tuple(feature["shape"]), + ) + return coerced + + +def _is_local_model_id(value: str) -> bool: + path = Path(value).expanduser() + return path.is_absolute() or value.startswith(("./", "../", "~")) or path.exists() + + +def _validate_wan_model_id(value: str, field_name: str) -> str: + if value == WAN22_MODEL_ID or _is_local_model_id(value): + return value + raise ValueError(f"`{field_name}` must be `{WAN22_MODEL_ID}` or an explicit local path, got `{value}`.") + + +def is_fastwam_base_compatible_config(config: FastWAMConfig) -> bool: + """Return whether `fastwam_base` partial weights can initialize this config.""" + + default_video_config = default_video_dit_config(config.action_dim) + default_action_config = default_action_dit_config(config.action_dim) + return all( + config.video_dit_config.get(key) == default_video_config.get(key) + for key in _FASTWAM_VIDEO_BASE_COMPAT_KEYS + ) and all( + config.action_dit_config.get(key) == default_action_config.get(key) + for key in _FASTWAM_ACTION_BASE_COMPAT_KEYS + ) + + +@PreTrainedConfig.register_subclass("fastwam") +@dataclass +class FastWAMConfig(PreTrainedConfig): + """Configuration for the FastWAM LeRobot policy. + + Args: + action_dim (int): Number of scalar action channels per timestep. + proprio_dim (int | None): Number of proprioception channels used as an + extra text-context token. `None` disables proprio conditioning. + action_horizon (int): Number of actions predicted by one policy call. + num_video_frames (int): Raw video sampling window (in dataset frames). The + model actually operates on `model_video_frames` frames after subsampling + by `action_video_freq_ratio`. + action_video_freq_ratio (int): Actions are sampled at this multiple of the + video frame rate. Video frames are taken every `action_video_freq_ratio`-th + raw frame, so the model sees `(num_video_frames - 1) // ratio + 1` frames + spanning the same time window as `action_horizon` actions (ratio actions + per video frame). + image_size (tuple[int, int]): Concatenated image size as `(height, width)`. + context_len (int): Maximum text embedding token length. + video_dit_config (dict[str, Any] | None): Wan video expert config. + action_dit_config (dict[str, Any] | None): Action expert config. + use_gradient_checkpointing (bool): Enable activation checkpointing in both DiT + experts (trades compute for memory; propagated into the DiT configs). + freeze_video_expert (bool): Freeze the ~5B Wan video expert + (`model.video_expert`) so only the action expert + proprio encoder train. + Cuts the AdamW optimizer footprint substantially; the video expert keeps its + pretrained weights. (If enabled, also set `loss.lambda_video=0` to skip the + now-gradient-free video loss compute.) + """ + + n_obs_steps: int = 1 + action_dim: int = 7 + proprio_dim: int | None = 8 + action_horizon: int = 32 + n_action_steps: int = 32 + num_video_frames: int = 33 + action_video_freq_ratio: int = 4 + image_size: tuple[int, int] = (224, 448) + context_len: int = 128 + model_id: str = WAN22_MODEL_ID + tokenizer_model_id: str = WAN_T5_TOKENIZER_ID + text_encoder_model_id: str = WAN22_DIFFUSERS_MODEL_ID + base_model_id: str | None = FASTWAM_BASE_MODEL_ID + tokenizer_max_len: int = 128 + load_text_encoder: bool = True + mot_checkpoint_mixed_attn: bool = False + torch_dtype: str = "bfloat16" + prompt_template: str = ( + "A video recorded from a robot's point of view executing the following instruction: {task}" + ) + num_inference_steps: int = 10 + inference_seed: int | None = 42 + rand_device: str = "cpu" + text_cfg_scale: float = 1.0 + negative_prompt: str = "" + sigma_shift: float | None = None + tiled: bool = False + fp32_attention: bool = True + use_gradient_checkpointing: bool = False + freeze_video_expert: bool = False + toggle_action_dimensions: list[int] = field(default_factory=list) + video_scheduler: dict[str, float | int] = field( + default_factory=lambda: {"train_shift": 5.0, "infer_shift": 5.0, "num_train_timesteps": 1000} + ) + action_scheduler: dict[str, float | int] = field( + default_factory=lambda: {"train_shift": 5.0, "infer_shift": 5.0, "num_train_timesteps": 1000} + ) + loss: dict[str, float] = field(default_factory=lambda: {"lambda_video": 1.0, "lambda_action": 1.0}) + video_dit_config: dict[str, Any] | None = None + action_dit_config: dict[str, Any] | None = None + normalization_mapping: dict[str, NormalizationMode] = field( + default_factory=lambda: { + "VISUAL": NormalizationMode.IDENTITY, + "STATE": NormalizationMode.MEAN_STD, + "ACTION": NormalizationMode.MEAN_STD, + } + ) + input_features: dict[str, PolicyFeature] | None = None + output_features: dict[str, PolicyFeature] | None = None + optimizer_lr: float = 1.0e-4 + optimizer_weight_decay: float = 1.0e-2 + + def __post_init__(self) -> None: + super().__post_init__() + self.image_size = tuple(self.image_size) + self.model_id = _validate_wan_model_id(self.model_id, "model_id") + self.input_features = _coerce_policy_features(self.input_features) + self.output_features = _coerce_policy_features(self.output_features) + self.toggle_action_dimensions = [int(dim) for dim in self.toggle_action_dimensions] + self.video_dit_config = self.video_dit_config or default_video_dit_config(self.action_dim) + self.action_dit_config = self.action_dit_config or default_action_dit_config(self.action_dim) + self.video_dit_config["fp32_attention"] = bool(self.fp32_attention) + self.action_dit_config["fp32_attention"] = bool(self.fp32_attention) + self.video_dit_config["use_gradient_checkpointing"] = bool(self.use_gradient_checkpointing) + self.action_dit_config["use_gradient_checkpointing"] = bool(self.use_gradient_checkpointing) + if self.input_features is None: + height, width = self.image_size + self.input_features = { + "observation.images.image": PolicyFeature( + type=FeatureType.VISUAL, + shape=(3, height, width), + ) + } + if self.proprio_dim is not None: + self.input_features[OBS_STATE] = PolicyFeature( + type=FeatureType.STATE, + shape=(self.proprio_dim,), + ) + if self.output_features is None: + self.output_features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(self.action_dim,))} + self.validate_features() + if self.pretrained_path or self.use_peft or not self.base_model_id: + return + if not is_fastwam_base_compatible_config(self): + return + self.pretrained_path = Path(self.base_model_id) + self._auto_pretrained_path = True + + def _save_pretrained(self, save_directory: Path) -> None: + if not getattr(self, "_auto_pretrained_path", False): + super()._save_pretrained(save_directory) + return + + pretrained_path = self.pretrained_path + self.pretrained_path = None + try: + super()._save_pretrained(save_directory) + finally: + self.pretrained_path = pretrained_path + + def get_optimizer_preset(self) -> AdamWConfig: + return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay) + + def get_scheduler_preset(self) -> None: + return None + + def set_dataset_feature_metadata(self, dataset_features: dict[str, Any]) -> None: + """Rebuild visual input features from the dataset's real camera keys. + + FastWAM's `__post_init__` installs a synthetic single-image default + (`observation.images.image` at full `image_size` width). For datasets + with one or more separately-named cameras (e.g. `observation.images.top`, + `observation.images.wrist`), this hook — invoked by `make_policy` once the + dataset metadata is known — replaces that default with the actual camera + keys, each declared at the policy's native per-camera resolution + (`image_size[0]` x `image_size[1] // num_cameras`). The accompanying + resize step in `make_fastwam_pre_post_processors` resizes raw frames to + match, so heterogeneous source resolutions (e.g. 480x640) are supported. + """ + image_keys = sorted( + key + for key, feature in dataset_features.items() + if key.startswith("observation.images.") and feature.get("dtype") in ("video", "image") + ) + if not image_keys: + return + height, total_width = self.image_size + per_cam_width = total_width // len(image_keys) + new_inputs: dict[str, PolicyFeature] = { + key: PolicyFeature(type=FeatureType.VISUAL, shape=(3, height, per_cam_width)) + for key in image_keys + } + if self.proprio_dim is not None and OBS_STATE in dataset_features: + new_inputs[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(self.proprio_dim,)) + self.input_features = new_inputs + self.validate_features() + + def validate_features(self) -> None: + if self.action_dim <= 0: + raise ValueError(f"`action_dim` must be positive, got {self.action_dim}.") + if self.action_horizon <= 0: + raise ValueError(f"`action_horizon` must be positive, got {self.action_horizon}.") + if self.n_action_steps > self.action_horizon: + raise ValueError("`n_action_steps` cannot exceed `action_horizon`.") + if self.action_video_freq_ratio <= 0: + raise ValueError( + f"`action_video_freq_ratio` must be positive, got {self.action_video_freq_ratio}." + ) + # Video frames are subsampled by action_video_freq_ratio; the resulting model frame + # count must satisfy T % 4 == 1 for the VAE temporal tokenization (mirrors the + # original FastWAM dataset asserts). + if (self.num_video_frames - 1) % self.action_video_freq_ratio != 0: + raise ValueError( + f"`num_video_frames - 1` ({self.num_video_frames - 1}) must be divisible by " + f"`action_video_freq_ratio` ({self.action_video_freq_ratio})." + ) + if ((self.num_video_frames - 1) // self.action_video_freq_ratio) % 4 != 0: + raise ValueError( + f"Subsampled video transitions ({(self.num_video_frames - 1) // self.action_video_freq_ratio}) " + "must be divisible by 4 for VAE tokenization (i.e. model_video_frames % 4 == 1)." + ) + if self.action_horizon % ((self.num_video_frames - 1) // self.action_video_freq_ratio) != 0: + raise ValueError( + f"`action_horizon` ({self.action_horizon}) must be divisible by the number of " + f"video transitions ({(self.num_video_frames - 1) // self.action_video_freq_ratio})." + ) + if not self.image_features: + raise ValueError("FastWAM requires at least one image feature.") + if self.action_feature is None: + raise ValueError("FastWAM requires `action` in output_features.") + action_shape = tuple(self.action_feature.shape) + if action_shape != (self.action_dim,): + raise ValueError( + f"FastWAM action feature shape must be ({self.action_dim},), got {action_shape}." + ) + if self.proprio_dim is not None: + state_feature = self.robot_state_feature + if state_feature is None: + raise ValueError("FastWAM requires `observation.state` when `proprio_dim` is set.") + state_shape = tuple(state_feature.shape) + if state_shape != (self.proprio_dim,): + raise ValueError( + f"FastWAM state feature shape must be ({self.proprio_dim},), got {state_shape}." + ) + height, width = self.image_size + image_width_sum = 0 + for name, feature in self.image_features.items(): + shape = tuple(feature.shape) + if len(shape) != 3 or shape[0] != 3: + raise ValueError(f"FastWAM image feature `{name}` must have shape (3, H, W), got {shape}.") + if shape[1] != height: + raise ValueError(f"FastWAM image feature `{name}` height must be {height}, got {shape[1]}.") + image_width_sum += shape[2] + if image_width_sum != width: + raise ValueError(f"FastWAM image feature widths must sum to {width}, got {image_width_sum}.") + + @property + def model_video_frames(self) -> int: + """Number of video frames the model actually operates on, after subsampling the + raw `num_video_frames` window by `action_video_freq_ratio` (e.g. 33 -> 9).""" + return (self.num_video_frames - 1) // self.action_video_freq_ratio + 1 + + @property + def observation_delta_indices(self) -> list[int]: + # Load the video frames the model is supervised on: the future window subsampled by + # action_video_freq_ratio (e.g. [0, 4, 8, ..., 32] -> 9 frames). Each video frame is + # thus `action_video_freq_ratio` actions apart, while actions load at the full rate + # (`action_delta_indices` = range(action_horizon)). Returning None would load only the + # current frame, making the video target a static repeat (degenerate supervision). + return list(range(0, self.num_video_frames, self.action_video_freq_ratio)) + + @property + def action_delta_indices(self) -> list[int]: + return list(range(self.action_horizon)) + + @property + def reward_delta_indices(self) -> None: + return None diff --git a/src/lerobot/policies/fastwam/modeling_fastwam.py b/src/lerobot/policies/fastwam/modeling_fastwam.py new file mode 100644 index 000000000..10671e717 --- /dev/null +++ b/src/lerobot/policies/fastwam/modeling_fastwam.py @@ -0,0 +1,440 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +from collections import deque +from typing import Any + +import torch +from torch import Tensor + +from lerobot.policies.pretrained import PreTrainedPolicy +from lerobot.utils.constants import OBS_STATE +from lerobot.utils.import_utils import require_package + +from .configuration_fastwam import FastWAMConfig +from .wan import ( + ActionDiT, + FastWAM, + MoT, + WanVideoDiT, + build_wan_tokenizer, + load_pretrained_wan_text_encoder, + load_pretrained_wan_vae, +) + + +class FastWAMPolicy(PreTrainedPolicy): + """LeRobot policy wrapper for FastWAM. + + Attention backend: FastWAM's DiT uses ``torch.nn.functional.scaled_dot_product_attention`` + (SDPA) for all attention. It does not use FlashAttention, because MoT routing requires + arbitrary boolean ``[query, key]`` masks that the FlashAttention varlen API cannot express; + installing ``flash-attn`` has no effect on the FastWAM path. (SDPA may still dispatch to + PyTorch's own flash/mem-efficient/math kernel internally, unrelated to the ``flash-attn`` package.) + + Args: + config (FastWAMConfig): FastWAM policy configuration. + dataset_stats (dict[str, dict[str, Tensor]] | None): Optional LeRobot + dataset statistics passed by the training/evaluation stack. + """ + + config_class = FastWAMConfig + name = "fastwam" + + def __init__( + self, + config: FastWAMConfig, + dataset_stats: dict[str, dict[str, Tensor]] | None = None, + **kwargs: Any, + ): + # FastWAM's Wan2.2 backbone needs transformers (UMT5 text encoder/tokenizer) and + # diffusers (Wan VAE), both behind the `fastwam` extra. Fail fast with an actionable + # message in base installs rather than deep in Wan component construction. + require_package("transformers", extra="fastwam") + require_package("diffusers", extra="fastwam") + # `make_policy`/`from_pretrained` forward extra kwargs (e.g. `dataset_meta`); the + # dataset feature metadata is already applied to `config` by make_policy upstream, + # so we accept and ignore them, matching the other LeRobot policies. + super().__init__(config, dataset_stats) + config.validate_features() + self.config = config + self.dataset_stats = dataset_stats + self.model = self._build_core_model(config) + if config.freeze_video_expert and getattr(self.model, "video_expert", None) is not None: + # Freeze the ~5B Wan video expert; get_optim_params filters on requires_grad, + # so its params drop out of the optimizer (and DDP skips them). + self.model.video_expert.requires_grad_(False) + # The transformer blocks are re-parented onto the MoTLayers (single FSDP owner), so + # `video_expert.requires_grad_` no longer reaches them — freeze them via the layers. + mot = getattr(self.model, "mot", None) + if mot is not None and getattr(mot, "layers", None) is not None: + for layer in mot.layers: + if "video" in layer.blocks: + layer.blocks["video"].requires_grad_(False) + self.reset() + + @classmethod + def _load_as_safetensor(cls, model, model_file: str, map_location: str, strict: bool): + """Shape-aware load that supports cross-embodiment fine-tuning. + + `safetensors.load_model(strict=False)` ignores missing/unexpected keys but + still raises on a shape mismatch for a shared key. When fine-tuning from a + checkpoint trained on a different embodiment (e.g. the LIBERO 7-DoF / 8-dim + checkpoint adapted to a 6-DoF / 6-dim arm), the action encoder/head and + proprio encoder legitimately differ in shape. With `strict=False` we drop + only those shape-mismatched tensors — leaving them at their freshly + initialized values — and load every compatible tensor. With `strict=True` + the standard exact-match loader is used. + """ + from safetensors import safe_open + + model_state_dict = model.state_dict() + mismatched = [] + with safe_open(model_file, framework="pt") as f: + checkpoint_keys = list(f.keys()) + for key in checkpoint_keys: + if key in model_state_dict and tuple(model_state_dict[key].shape) != tuple( + f.get_slice(key).get_shape() + ): + mismatched.append(key) + + if not mismatched: + return super()._load_as_safetensor(model, model_file, map_location, strict) + if strict: + raise RuntimeError( + f"FastWAM: {len(mismatched)} checkpoint tensors have a shape mismatch under " + f"strict=True: {mismatched}" + ) + + from safetensors.torch import load_file + + logging.warning( + "FastWAM cross-embodiment load: reinitializing %d shape-mismatched tensor(s), keeping " + "every compatible weight: %s", + len(mismatched), + mismatched, + ) + state_dict = load_file(model_file, device="cpu") + for key in mismatched: + state_dict.pop(key, None) + model.load_state_dict(state_dict, strict=False) + if map_location and map_location != "cpu": + model.to(map_location) + return model + + def get_optim_params(self) -> list[Tensor]: + # Return the trainable tensors directly (a single param group). The optimizer + # builder wraps these in a param group; returning a bare {"params": [...]} dict + # instead would make `list(...)` yield the key string "params". + params = ( + list(self.model.dit.parameters()) if hasattr(self.model, "dit") else list(self.model.parameters()) + ) + proprio_encoder = getattr(self.model, "proprio_encoder", None) + if proprio_encoder is not None: + params.extend(list(proprio_encoder.parameters())) + return [p for p in params if p.requires_grad] + + def reset(self) -> None: + self._action_queue: deque[Tensor] = deque([], maxlen=self.config.n_action_steps) + + def _batch_to_training_sample(self, batch: dict[str, Tensor]) -> dict[str, Tensor]: + """Adapt a standard LeRobot batch to the FastWAM-native sample that + `FastWAM.build_inputs` consumes (`video`, `action`, `context`/`context_mask`, + per-frame `proprio`). + + The LeRobot training loop passes raw `observation.images.*`, a single-step + `observation.state` `[B, D]`, `action`, and a language `task` string. We do + only the translation `build_inputs` can't: stack the camera frames into a + video, encode the prompt with the (frozen) text encoder (mirroring inference, + so language-conditioned datasets need no precomputed context), and give proprio + the per-frame axis `build_inputs` indexes. All shape/presence validation is + left to `build_inputs`, the single authority on the contract. + """ + sample = dict(batch) + if "video" not in sample: + sample["video"] = _stack_video_from_images(batch, self.config) + if "context" not in sample or "context_mask" not in sample: + prompt = _prompt_from_batch(batch=batch, config=self.config) + if prompt is None: + raise KeyError( + "FastWAM training requires a `task`/`prompt` to encode text context, " + "or precomputed `context`/`context_mask` in the batch." + ) + sample["context"], sample["context_mask"] = self.model.encode_prompt(prompt) + if self.config.proprio_dim is not None and "proprio" not in sample: + state = sample.get(OBS_STATE) + if state is not None: + # LeRobot gives a single-step state [B, D]; build_inputs expects + # per-frame [B, T, D] and uses frame 0, so add a T=1 axis. + sample["proprio"] = state.unsqueeze(1) if state.ndim == 2 else state + return sample + + def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Any]]: + """Compute FastWAM training loss for a LeRobot batch. + + Args: + batch (dict[str, Tensor]): Batch containing FastWAM-ready keys + (`video`, `action`, `context`, `context_mask`) or LeRobot keys + that can be adapted (`observation.images.*`, `observation.state`, + `action`, `action_is_pad`). + + Returns: + tuple[Tensor, dict[str, Any]]: The scalar loss to backprop, and a dict of + logging metrics (e.g. `loss_video`, `loss_action`) — the `(loss, output_dict)` + contract the LeRobot training loop expects. + """ + + sample = self._batch_to_training_sample(batch) + loss, metrics = self.model.training_loss(sample) + return loss, dict(metrics or {}) + + @torch.no_grad() + def predict_action_chunk(self, batch: dict[str, Tensor], **_: Any) -> Tensor: + """Predict a chunk of actions from the current FastWAM observation. + + Args: + batch (dict[str, Tensor]): Inference batch with `input_image` or + image observation keys, plus `context/context_mask` or `prompt`. + + Returns: + Tensor: Action chunk with shape `[B, action_horizon, action_dim]`. + """ + + self.eval() + infer_kwargs = _batch_to_infer_kwargs(batch=batch, config=self.config) + batch_size = _infer_kwargs_batch_size(infer_kwargs) + if batch_size == 1: + action = _action_from_model_output(self.model.infer_action(**infer_kwargs)) + else: + action = torch.cat( + [ + _action_from_model_output( + self.model.infer_action( + **_slice_infer_kwargs(infer_kwargs, index=i, batch_size=batch_size) + ) + ) + for i in range(batch_size) + ], + dim=0, + ) + return action.to(device=batch_device(batch), dtype=torch.float32) + + @torch.no_grad() + def select_action(self, batch: dict[str, Tensor], **kwargs: Any) -> Tensor: + self.eval() + if len(self._action_queue) == 0: + actions = self.predict_action_chunk(batch, **kwargs)[:, : self.config.n_action_steps] + self._action_queue.extend(actions.transpose(0, 1)) + return self._action_queue.popleft() + + def _build_core_model(self, config: FastWAMConfig) -> FastWAM: + """Build the FastWAM core for training / inference. + + Only the trainable parts (the MoT DiT and the proprio encoder) are + materialized empty here and then filled from the policy's + `model.safetensors` by the base `from_pretrained`. The *frozen* Wan2.2 VAE + and UMT5 text encoder are loaded with their real weights from the + `Wan-AI/Wan2.2-TI2V-5B-Diffusers` repo (cached in the HF cache, shared + across checkpoints) and are intentionally excluded from `model.safetensors` + — see `FastWAM.__init__`. The tokenizer comes from `google/umt5-xxl`. + """ + dtype = _dtype_from_name(config.torch_dtype) + device = config.device + video_expert = WanVideoDiT(**config.video_dit_config).to(device=device, dtype=dtype) + action_expert = ActionDiT(**config.action_dit_config).to(device=device, dtype=dtype) + mot = MoT( + mixtures={"video": video_expert, "action": action_expert}, + mot_checkpoint_mixed_attn=config.mot_checkpoint_mixed_attn, + ) + text_encoder = ( + load_pretrained_wan_text_encoder( + model_id=config.text_encoder_model_id, torch_dtype=dtype, device=device + ) + if config.load_text_encoder + else None + ) + return FastWAM( + video_expert=video_expert, + action_expert=action_expert, + mot=mot, + vae=load_pretrained_wan_vae(torch_dtype=dtype, device=device), + text_encoder=text_encoder, + tokenizer=build_wan_tokenizer( + model_id=config.tokenizer_model_id, tokenizer_max_len=config.tokenizer_max_len + ), + text_dim=int(config.video_dit_config["text_dim"]), + proprio_dim=config.proprio_dim, + device=device, + torch_dtype=dtype, + video_train_shift=float(config.video_scheduler["train_shift"]), + video_infer_shift=float(config.video_scheduler["infer_shift"]), + video_num_train_timesteps=int(config.video_scheduler["num_train_timesteps"]), + action_train_shift=float(config.action_scheduler["train_shift"]), + action_infer_shift=float(config.action_scheduler["infer_shift"]), + action_num_train_timesteps=int(config.action_scheduler["num_train_timesteps"]), + loss_lambda_video=float(config.loss["lambda_video"]), + loss_lambda_action=float(config.loss["lambda_action"]), + ) + + +def _scalar(value: Any) -> Any: + """Unwrap a 0-/1-element tensor (e.g. from DataLoader collation) to a Python scalar.""" + return value.item() if isinstance(value, Tensor) else value + + +def _batch_to_infer_kwargs(batch: dict[str, Tensor], config: FastWAMConfig) -> dict[str, Any]: + return { + "prompt": _prompt_from_batch(batch=batch, config=config), + "input_image": _input_image_from_batch(batch, config), + "action_horizon": config.action_horizon, + "proprio": batch.get("proprio", batch.get(OBS_STATE)), + "context": batch.get("context"), + "context_mask": batch.get("context_mask"), + "negative_prompt": batch.get("negative_prompt", config.negative_prompt), + "text_cfg_scale": float(_scalar(batch.get("text_cfg_scale", config.text_cfg_scale))), + "num_inference_steps": int(_scalar(batch.get("num_inference_steps", config.num_inference_steps))), + "sigma_shift": batch.get("sigma_shift", config.sigma_shift), + "seed": batch.get("seed", config.inference_seed), + "rand_device": batch.get("rand_device", config.rand_device), + "tiled": bool(batch.get("tiled", config.tiled)), + } + + +def _prompt_from_batch(batch: dict[str, Tensor], config: FastWAMConfig) -> Any: + prompt = batch.get("prompt") + if prompt is not None: + return prompt + + task = batch.get("task") + if task is None: + return None + if isinstance(task, str): + return config.prompt_template.format(task=task) + if isinstance(task, (list, tuple)): + return [config.prompt_template.format(task=str(item)) for item in task] + return config.prompt_template.format(task=str(task)) + + +def _action_from_model_output(output: Any) -> Tensor: + action = output["action"] if isinstance(output, dict) else output + if action.ndim == 2: + action = action.unsqueeze(0) + return action + + +def _infer_kwargs_batch_size(infer_kwargs: dict[str, Any]) -> int: + image = infer_kwargs["input_image"] + if not isinstance(image, Tensor): + raise TypeError(f"`input_image` must be a tensor, got {type(image).__name__}.") + if image.ndim == 3: + return 1 + if image.ndim == 4: + return int(image.shape[0]) + raise ValueError(f"`input_image` must be [B,C,H,W] or [C,H,W], got {tuple(image.shape)}.") + + +def _slice_infer_kwargs(infer_kwargs: dict[str, Any], *, index: int, batch_size: int) -> dict[str, Any]: + return { + key: _slice_infer_value(value, index=index, batch_size=batch_size) + for key, value in infer_kwargs.items() + } + + +def _slice_infer_value(value: Any, *, index: int, batch_size: int) -> Any: + if isinstance(value, Tensor) and value.ndim > 0 and value.shape[0] == batch_size: + return value[index : index + 1] + if isinstance(value, (list, tuple)) and len(value) == batch_size: + return value[index] + return value + + +def _dtype_from_name(name: str) -> torch.dtype: + dtype_map = {"float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16} + if name not in dtype_map: + raise ValueError(f"Unsupported torch dtype `{name}`.") + return dtype_map[name] + + +def batch_device(batch: dict[str, Any]) -> torch.device: + for value in batch.values(): + if isinstance(value, Tensor): + return value.device + return torch.device("cpu") + + +def _resize_frames(frames: Tensor, size: tuple[int, int]) -> Tensor: + """Resize a frame tensor to `size` (H, W), tolerating a leading temporal/batch stack. + + `interpolate` only accepts a single leading batch dim (`[N, C, H, W]`), but FastWAM camera + tensors arrive as `[B, C, H, W]` (live eval) or `[B, T, C, H, W]` (temporal stack), so flatten + any leading dims into the batch, resize, then restore. A no-op when already at `size`. + """ + if tuple(frames.shape[-2:]) == size: + return frames + lead = frames.shape[:-3] + flat = frames.reshape(-1, *frames.shape[-3:]) + flat = torch.nn.functional.interpolate( + flat, size=size, mode="bilinear", align_corners=False, antialias=True + ) + return flat.reshape(*lead, *flat.shape[-3:]) + + +def _stack_video_from_images(batch: dict[str, Tensor], config: FastWAMConfig) -> Tensor: + # Exclude the `*_is_pad` companion tensors that delta-timestamp loading adds alongside + # each camera (shape [B, T]); they share the `observation.images.` prefix but are not frames. + image_keys = sorted(k for k in batch if k.startswith("observation.images.") and not k.endswith("_is_pad")) + if not image_keys: + raise KeyError("FastWAM batch must contain `video` or `observation.images.*` keys.") + per_cam = (int(config.image_size[0]), int(config.image_size[1]) // len(image_keys)) + images = [_resize_frames(batch[key], per_cam) for key in image_keys] + # Cameras concatenate along width (last dim) in both the single-frame and temporal case. + image = torch.cat(images, dim=-1) if len(images) > 1 else images[0] + if image.ndim == 4: + # [B, C, H, W]: a single frame (e.g. the live eval observation) -> repeat across time. + image = image.unsqueeze(2).repeat(1, 1, config.model_video_frames, 1, 1) + elif image.ndim == 5: + # [B, T, C, H, W]: temporal stack from delta-timestamp loading -> [B, C, T, H, W]. + image = image.permute(0, 2, 1, 3, 4) + else: + raise ValueError(f"Expected image batch [B,C,H,W] or temporal [B,T,C,H,W], got {tuple(image.shape)}.") + return image + + +def _input_image_from_batch(batch: dict[str, Tensor], config: FastWAMConfig) -> Tensor: + if "input_image" in batch: + return _prepare_infer_image(batch["input_image"], config) + video = batch.get("video") + if video is None: + video = _stack_video_from_images(batch, config) + if video.ndim == 5: + return _prepare_infer_image(video[:, :, 0], config) + if video.ndim == 4: + return _prepare_infer_image(video, config) + raise ValueError(f"Cannot build input image from tensor with shape {tuple(video.shape)}.") + + +def _prepare_infer_image(image: Tensor, config: FastWAMConfig) -> Tensor: + if image.ndim == 3: + image = image.unsqueeze(0) + if image.ndim != 4: + raise ValueError(f"Expected image tensor [B,C,H,W] or [C,H,W], got {tuple(image.shape)}.") + + # Resize to the full configured resolution (no-op when the video path already produced it, but + # also covers a directly-supplied `input_image`). The model owns its input resolution — see + # `_stack_video_from_images` — so we resize rather than assert on a mismatch. + target_h, target_w = int(config.image_size[0]), int(config.image_size[1]) + return _resize_frames(image, (target_h, target_w)) diff --git a/src/lerobot/policies/fastwam/processor_fastwam.py b/src/lerobot/policies/fastwam/processor_fastwam.py new file mode 100644 index 000000000..31f3b9277 --- /dev/null +++ b/src/lerobot/policies/fastwam/processor_fastwam.py @@ -0,0 +1,142 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass +from typing import Any + +import torch + +from lerobot.configs import PipelineFeatureType, PolicyFeature +from lerobot.processor import ( + ActionProcessorStep, + AddBatchDimensionProcessorStep, + DeviceProcessorStep, + NormalizerProcessorStep, + PolicyAction, + PolicyProcessorPipeline, + ProcessorStepRegistry, + RenameObservationsProcessorStep, + UnnormalizerProcessorStep, + policy_action_to_transition, + transition_to_policy_action, +) +from lerobot.utils.constants import ( + POLICY_POSTPROCESSOR_DEFAULT_NAME, + POLICY_PREPROCESSOR_DEFAULT_NAME, +) + +from .configuration_fastwam import FastWAMConfig + + +@dataclass +@ProcessorStepRegistry.register(name="fastwam_action_toggle_processor") +class FastWAMActionToggleProcessorStep(ActionProcessorStep): + """Apply FastWAM LIBERO toggle semantics to configured action dimensions.""" + + toggle_dimensions: list[int] + + def action(self, action: PolicyAction) -> PolicyAction: + if not self.toggle_dimensions: + return action + processed_action = action.clone() + action_dim = int(processed_action.shape[-1]) + for dim in self.toggle_dimensions: + resolved_dim = dim if dim >= 0 else action_dim + dim + if resolved_dim < 0 or resolved_dim >= action_dim: + raise ValueError( + f"FastWAM action toggle dimension {dim} is out of bounds for action dim {action_dim}." + ) + value = processed_action[..., resolved_dim] + value = value * 2.0 - 1.0 + processed_action[..., resolved_dim] = torch.sign(-value) + return processed_action + + def get_config(self) -> dict[str, Any]: + return {"toggle_dimensions": self.toggle_dimensions} + + def transform_features( + self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]] + ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]: + return features + + +def make_fastwam_pre_post_processors( + config: FastWAMConfig, + dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None, +) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]: + """Create LeRobot pre- and post-processing pipelines for FastWAM. + + Args: + config (FastWAMConfig): Policy configuration controlling device and + normalization feature metadata. + dataset_stats (dict[str, dict[str, torch.Tensor]] | None): Optional + LeRobot dataset statistics used by normalization processors. + + Returns: + tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]: Input and + output processor pipelines discoverable by LeRobot. + """ + + # NOTE: no visual normalization here. VISUAL is IDENTITY (see configuration_fastwam.normalization_mapping) + # — images pass through in [0, 1] and the model maps them to the Wan VAE's [-1, 1] at the encode + # boundary. This is deliberate: `lerobot_train.py` overrides the normalizer stats with + # `dataset.meta.stats` when fine-tuning, and a real dataset's per-channel image std is the tiny + # frame-to-frame brightness variance, which would blow images far outside [-1,1] and saturate them. + # STATE/ACTION still normalize with dataset stats below. + normalization_stats: dict[str, dict[str, Any]] = dict(dataset_stats or {}) + + # NOTE: no resize step here. The model is the single authority on input resolution: it resizes + # each camera to the per-camera target (image_size split across cameras) in + # `_stack_video_from_images` / `_prepare_infer_image`, on every path (train forward, rollout and + # eval select_action). A preprocessor resize step would be both redundant (the model re-resizes + # anyway) and unsafe across fine-tuning: its `resize_size` would be inherited from the base + # checkpoint's camera geometry, not this dataset's, making the concatenation N_cameras x too wide. + + input_steps = [ + RenameObservationsProcessorStep(rename_map={}), + AddBatchDimensionProcessorStep(), + DeviceProcessorStep(device=config.device), + NormalizerProcessorStep( + features={**config.input_features, **config.output_features}, + norm_map=config.normalization_mapping, + stats=normalization_stats, + device=config.device, + ), + ] + output_steps = [ + UnnormalizerProcessorStep( + features=config.output_features, + norm_map=config.normalization_mapping, + stats=normalization_stats, + ), + ] + if config.toggle_action_dimensions: + output_steps.append( + FastWAMActionToggleProcessorStep(toggle_dimensions=config.toggle_action_dimensions) + ) + output_steps.append(DeviceProcessorStep(device="cpu")) + return ( + PolicyProcessorPipeline[dict[str, Any], dict[str, Any]]( + steps=input_steps, + name=POLICY_PREPROCESSOR_DEFAULT_NAME, + ), + PolicyProcessorPipeline[PolicyAction, PolicyAction]( + steps=output_steps, + name=POLICY_POSTPROCESSOR_DEFAULT_NAME, + to_transition=policy_action_to_transition, + to_output=transition_to_policy_action, + ), + ) diff --git a/src/lerobot/policies/fastwam/wan/README.md b/src/lerobot/policies/fastwam/wan/README.md new file mode 100644 index 000000000..7b7d61033 --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/README.md @@ -0,0 +1,34 @@ +# FastWAM `wan` package + +This package holds FastWAM's model implementation. It mixes a small **vendored +subset of the official Wan2.2 source tree** with FastWAM's own code, kept flat in +a single directory. + +## Vendored from Wan2.2 + +- Upstream repository: https://github.com/Wan-Video/Wan2.2 +- Upstream commit: `42bf4cfaa384bc21833865abc2f9e6c0e67233dc` +- License: Apache-2.0, matching the license in `LICENSE.txt` from the upstream repository + +Copied files: + +- `model.py` (was `wan/modules/model.py`), trimmed: the flash-attention path + (the vendored `attention.py` and the block/model `forward`s) was removed. + FastWAM's DiT uses SDPA instead (see `video_dit.py`). +- `get_sampling_sigmas` in `video_dit.py` (was `wan/utils/fm_solvers.py`), inlined + next to its only caller. + +This subset only backs FastWAM's **custom MoT video DiT**. The Wan2.2 VAE, +UMT5 text encoder, and tokenizer are no longer vendored - they come from +`diffusers.AutoencoderKLWan`, `transformers.UMT5EncoderModel`, and +`transformers.AutoTokenizer` (see `components.py` and `adapters.py`). + +## FastWAM's own code + +- `video_dit.py` builds on `model` (`sinusoidal_embedding_1d`, `rope_params`, + `rope_apply`, …) and computes attention with SDPA (`fastwam_masked_attention`). Its + `WanContinuousFlowMatchScheduler` uses `get_sampling_sigmas` for Wan-compatible + inference timesteps. +- `components.py` / `adapters.py` load the VAE, text encoder, tokenizer, and the + custom DiT weights. +- `modular.py` defines the FastWAM model (`ActionDiT`, `MoT`, `FastWAM`, …). diff --git a/src/lerobot/policies/fastwam/wan/__init__.py b/src/lerobot/policies/fastwam/wan/__init__.py new file mode 100644 index 000000000..11c0aaf6f --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/__init__.py @@ -0,0 +1,33 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from .adapters import WanVideoVAE38 +from .components import ( + build_wan_tokenizer, + load_pretrained_wan_text_encoder, + load_pretrained_wan_vae, +) +from .modular import ActionDiT, FastWAM, MoT +from .video_dit import WanVideoDiT + +__all__ = [ + "ActionDiT", + "FastWAM", + "MoT", + "WanVideoDiT", + "WanVideoVAE38", + "build_wan_tokenizer", + "load_pretrained_wan_text_encoder", + "load_pretrained_wan_vae", +] diff --git a/src/lerobot/policies/fastwam/wan/adapters.py b/src/lerobot/policies/fastwam/wan/adapters.py new file mode 100644 index 000000000..a2e8a1068 --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/adapters.py @@ -0,0 +1,108 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from typing import TYPE_CHECKING + +import torch + +if TYPE_CHECKING: + from diffusers import AutoencoderKLWan + + +class WanVideoVAE38(torch.nn.Module): + """FastWAM VAE contract over `diffusers.AutoencoderKLWan` (Wan2.2-TI2V-5B). + + 16x spatial / 4x temporal compression, 48 latent channels. diffusers' + `AutoencoderKLWan` returns *raw* latents (it does not apply `latents_mean`/ + `latents_std`), so `encode`/`decode` here apply the same standardization the + Wan reference uses — `(latents - mean) / std` — done in fp32 for stability. + `encode` uses the deterministic posterior mode, matching the original VAE + which returned the latent mean `mu`. + """ + + upsampling_factor = 16 + temporal_downsample_factor = 4 + z_dim = 48 + + def __init__( + self, + dtype: torch.dtype = torch.float32, + device: str | torch.device = "cuda", + *, + pretrained: AutoencoderKLWan, + ) -> None: + super().__init__() + # The Wan2.2 VAE is a fixed pretrained model — it is never trained from scratch, + # so a real `AutoencoderKLWan` (with weights) must always be supplied (loaded from + # the diffusers repo by `load_pretrained_wan_vae`). No random/offline build path. + self.vae = pretrained.to(device=device, dtype=dtype) + + # Read the standardization stats from the VAE's own config (diffusers populates + # these from vae/config.json) — single source of truth, no local copy. diffusers' + # encode/decode return *raw* latents, so we apply (latent - mean) / std ourselves. + # Non-persistent: kept out of state_dict. + self.register_buffer( + "latents_mean", + torch.tensor(self.vae.config.latents_mean).view(1, self.z_dim, 1, 1, 1), + persistent=False, + ) + self.register_buffer( + "latents_std", + torch.tensor(self.vae.config.latents_std).view(1, self.z_dim, 1, 1, 1), + persistent=False, + ) + + def _device_dtype(self) -> tuple[torch.device, torch.dtype]: + param = next(self.vae.parameters()) + return param.device, param.dtype + + def encode( + self, + videos: list[torch.Tensor] | torch.Tensor, + device: str | torch.device | None = None, + tiled: bool = False, + tile_size: tuple[int, int] = (34, 34), + tile_stride: tuple[int, int] = (18, 16), + ) -> torch.Tensor: + del device, tile_size, tile_stride + if tiled: + raise NotImplementedError("Tiled Wan2.2 VAE encoding is not supported by the FastWAM adapter.") + if isinstance(videos, (list, tuple)): + videos = torch.stack(list(videos)) + dev, dtype = self._device_dtype() + mu = self.vae.encode(videos.to(device=dev, dtype=dtype)).latent_dist.mode().float() + mean = self.latents_mean.float().to(mu.device) + std = self.latents_std.float().to(mu.device) + return (mu - mean) / std + + def decode( + self, + hidden_states: list[torch.Tensor] | torch.Tensor, + device: str | torch.device | None = None, + tiled: bool = False, + tile_size: tuple[int, int] = (34, 34), + tile_stride: tuple[int, int] = (18, 16), + ) -> torch.Tensor: + del device, tile_size, tile_stride + if tiled: + raise NotImplementedError("Tiled Wan2.2 VAE decoding is not supported by the FastWAM adapter.") + if isinstance(hidden_states, (list, tuple)): + hidden_states = torch.stack(list(hidden_states)) + dev, dtype = self._device_dtype() + z = hidden_states.float() + z = z * self.latents_std.float().to(z.device) + self.latents_mean.float().to(z.device) + out = self.vae.decode(z.to(device=dev, dtype=dtype)).sample + return out.float().clamp_(-1.0, 1.0) diff --git a/src/lerobot/policies/fastwam/wan/components.py b/src/lerobot/policies/fastwam/wan/components.py new file mode 100644 index 000000000..59a9b610e --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/components.py @@ -0,0 +1,175 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +from collections.abc import Sequence +from pathlib import Path +from typing import TYPE_CHECKING, Any + +import torch +from huggingface_hub import snapshot_download +from safetensors.torch import load_file + +from lerobot.utils.import_utils import _diffusers_available, _transformers_available, require_package + +if TYPE_CHECKING or _transformers_available: + from transformers import AutoTokenizer, UMT5EncoderModel +else: + AutoTokenizer = None + UMT5EncoderModel = None + +if TYPE_CHECKING or _diffusers_available: + from diffusers import AutoencoderKLWan +else: + AutoencoderKLWan = None + +from .adapters import WanVideoVAE38 +from .video_dit import WanVideoDiT + +logger = logging.getLogger(__name__) + +# The custom MoT video DiT still ships in the original (non-diffusers) Wan2.2 +# repo as sharded `diffusion_pytorch_model*.safetensors`; the VAE and UMT5 text +# encoder come from the diffusers conversion. Tokenizer is the stock UMT5 one. +WAN_DIT_PATTERN = "diffusion_pytorch_model*.safetensors" +WAN_T5_TOKENIZER = "google/umt5-xxl" +WAN22_DIFFUSERS_MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B-Diffusers" + + +class WanTextEncoder(torch.nn.Module): + """FastWAM text-encoder contract over `transformers.UMT5EncoderModel`. + + Exposes `.dim` (hidden size) and `forward(ids, mask) -> [B, L, dim]`, matching + the call in `FastWAM.encode_prompt`. + """ + + def __init__( + self, + dtype: torch.dtype = torch.bfloat16, + device: str | torch.device = "cuda", + *, + pretrained: torch.nn.Module, + ) -> None: + super().__init__() + # UMT5-XXL is a fixed pretrained encoder — never trained from scratch, so a real + # `UMT5EncoderModel` (with weights) must always be supplied (loaded from the + # diffusers repo by `load_pretrained_wan_text_encoder`). No random/offline build. + self.model = pretrained.to(device=device, dtype=dtype) + self.dim = int(self.model.config.d_model) + + def forward(self, ids: torch.Tensor, mask: torch.Tensor) -> torch.Tensor: + return self.model(input_ids=ids, attention_mask=mask.long()).last_hidden_state + + +class WanTokenizer: + """UMT5 tokenizer wrapper returning `(input_ids, attention_mask)` like the + FastWAM call site expects.""" + + def __init__(self, name: str = WAN_T5_TOKENIZER, seq_len: int = 512) -> None: + require_package("transformers", extra="fastwam") + self.tokenizer = AutoTokenizer.from_pretrained(name) + self.seq_len = int(seq_len) + + def __call__( + self, + sequence: str | Sequence[str], + return_mask: bool = False, + add_special_tokens: bool = True, + **_: Any, + ): + if isinstance(sequence, str): + sequence = [sequence] + out = self.tokenizer( + list(sequence), + padding="max_length", + truncation=True, + max_length=self.seq_len, + add_special_tokens=add_special_tokens, + return_tensors="pt", + ) + if return_mask: + return out.input_ids, out.attention_mask + return out.input_ids + + +def build_wan_tokenizer(*, model_id: str = WAN_T5_TOKENIZER, tokenizer_max_len: int) -> WanTokenizer: + return WanTokenizer(name=model_id, seq_len=int(tokenizer_max_len)) + + +def load_pretrained_wan_vae(*, torch_dtype: torch.dtype, device: str) -> WanVideoVAE38: + """Load real Wan2.2 VAE weights from the diffusers repo (offline base creation).""" + require_package("diffusers", extra="fastwam") + vae = AutoencoderKLWan.from_pretrained(WAN22_DIFFUSERS_MODEL_ID, subfolder="vae", torch_dtype=torch_dtype) + return WanVideoVAE38(dtype=torch_dtype, device=device, pretrained=vae) + + +def load_pretrained_wan_text_encoder( + *, + model_id: str = WAN22_DIFFUSERS_MODEL_ID, + subfolder: str | None = "text_encoder", + torch_dtype: torch.dtype, + device: str, +) -> WanTextEncoder: + """Load UMT5-XXL encoder weights (defaults to the Wan2.2 diffusers repo). + + Must stay compatible with the tokenizer (see `build_wan_tokenizer`): the encoder's + embedding table is indexed by the tokenizer's vocabulary. + """ + require_package("transformers", extra="fastwam") + encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder=subfolder, torch_dtype=torch_dtype) + return WanTextEncoder(dtype=torch_dtype, device=device, pretrained=encoder) + + +def resolve_wan_dit_paths( + model_id_or_path: str | Path, + *, + cache_dir: str | Path | None = None, + local_files_only: bool = False, + revision: str | None = None, +) -> list[Path]: + """Resolve the custom MoT DiT shards from the original Wan2.2 repo or a local dir.""" + path = Path(model_id_or_path).expanduser() + if path.is_dir(): + return sorted(path.glob(WAN_DIT_PATTERN)) + + snapshot_path = snapshot_download( + repo_id=str(model_id_or_path), + revision=revision, + cache_dir=cache_dir, + local_files_only=local_files_only, + allow_patterns=[WAN_DIT_PATTERN], + ) + return sorted(Path(snapshot_path).glob(WAN_DIT_PATTERN)) + + +def load_wan_video_dit( + paths: list[str | Path], + *, + dit_config: dict[str, Any], + torch_dtype: torch.dtype, + device: str, +) -> WanVideoDiT: + model = WanVideoDiT(**dit_config) + state_dict = _read_wan_dit_safetensors(paths) + model.load_state_dict(state_dict, strict=False) + return model.to(device=device, dtype=torch_dtype) + + +def _read_wan_dit_safetensors(paths: list[str | Path]) -> dict[str, torch.Tensor]: + state_dict = {} + for path in paths: + state_dict.update(load_file(str(path), device="cpu")) + return state_dict diff --git a/src/lerobot/policies/fastwam/wan/model.py b/src/lerobot/policies/fastwam/wan/model.py new file mode 100644 index 000000000..329d1c48a --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/model.py @@ -0,0 +1,341 @@ +# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved. +import math + +import torch +import torch.nn as nn + + +def sinusoidal_embedding_1d(dim, position): + # preprocess + if dim % 2 != 0: + raise ValueError(f"dim must be even, got {dim}.") + half = dim // 2 + position = position.type(torch.float64) + + # calculation + sinusoid = torch.outer(position, torch.pow(10000, -torch.arange(half).to(position).div(half))) + x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1) + return x + + +@torch.amp.autocast("cuda", enabled=False) +def rope_params(max_seq_len, dim, theta=10000): + if dim % 2 != 0: + raise ValueError(f"dim must be even, got {dim}.") + freqs = torch.outer( + torch.arange(max_seq_len), 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float64).div(dim)) + ) + freqs = torch.polar(torch.ones_like(freqs), freqs) + return freqs + + +@torch.amp.autocast("cuda", enabled=False) +def rope_apply(x, grid_sizes, freqs): + n, c = x.size(2), x.size(3) // 2 + + # split freqs + freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1) + + # loop over samples + output = [] + for i, (f, h, w) in enumerate(grid_sizes.tolist()): + seq_len = f * h * w + + # precompute multipliers + x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape(seq_len, n, -1, 2)) + freqs_i = torch.cat( + [ + freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1), + freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1), + freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1), + ], + dim=-1, + ).reshape(seq_len, 1, -1) + + # apply rotary embedding + x_i = torch.view_as_real(x_i * freqs_i).flatten(2) + x_i = torch.cat([x_i, x[i, seq_len:]]) + + # append to collection + output.append(x_i) + return torch.stack(output).float() + + +class WanRMSNorm(nn.Module): + def __init__(self, dim, eps=1e-5): + super().__init__() + self.dim = dim + self.eps = eps + self.weight = nn.Parameter(torch.ones(dim)) + + def forward(self, x): + r""" + Args: + x(Tensor): Shape [B, L, C] + """ + return self._norm(x.float()).type_as(x) * self.weight + + def _norm(self, x): + return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps) + + +class WanLayerNorm(nn.LayerNorm): + def __init__(self, dim, eps=1e-6, elementwise_affine=False): + super().__init__(dim, elementwise_affine=elementwise_affine, eps=eps) + + def forward(self, x): + r""" + Args: + x(Tensor): Shape [B, L, C] + """ + return super().forward(x.float()).type_as(x) + + +class WanSelfAttention(nn.Module): + def __init__(self, dim, num_heads, qk_norm=True, eps=1e-6): + if dim % num_heads != 0: + raise ValueError(f"dim ({dim}) must be divisible by num_heads ({num_heads}).") + super().__init__() + self.num_heads = num_heads + self.head_dim = dim // num_heads + + # layers + self.q = nn.Linear(dim, dim) + self.k = nn.Linear(dim, dim) + self.v = nn.Linear(dim, dim) + self.o = nn.Linear(dim, dim) + self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() + self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() + + # NOTE: FastWAM never runs the upstream Wan attention forward. FastWAMAttentionBlock + # reuses only the q/k/v/o/norm submodules defined above and computes attention via + # `fastwam_masked_attention` (SDPA). The original flash-attention forward was removed, + # which also collapsed the former WanCrossAttention subclass into this class (it only + # differed by its forward): self- and cross-attention now share the same projection module. + + +class WanAttentionBlock(nn.Module): + def __init__(self, dim, ffn_dim, num_heads, qk_norm=True, cross_attn_norm=False, eps=1e-6): + super().__init__() + self.dim = dim + self.ffn_dim = ffn_dim + self.num_heads = num_heads + self.qk_norm = qk_norm + self.cross_attn_norm = cross_attn_norm + self.eps = eps + + # layers + self.norm1 = WanLayerNorm(dim, eps) + self.self_attn = WanSelfAttention(dim, num_heads, qk_norm, eps) + self.norm3 = WanLayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity() + self.cross_attn = WanSelfAttention(dim, num_heads, qk_norm, eps) + self.norm2 = WanLayerNorm(dim, eps) + self.ffn = nn.Sequential( + nn.Linear(dim, ffn_dim), nn.GELU(approximate="tanh"), nn.Linear(ffn_dim, dim) + ) + + # modulation + self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5) + + # NOTE: The upstream Wan block forward (self-attention + cross-attention + FFN via + # flash-attention) was removed. FastWAM subclasses this block as FastWAMAttentionBlock + # and overrides forward to use SDPA with explicit boolean masks; only __init__ (the + # norm/attention/ffn submodules) is reused here. + + +class Head(nn.Module): + def __init__(self, dim, out_dim, patch_size, eps=1e-6): + super().__init__() + self.dim = dim + self.out_dim = out_dim + self.patch_size = patch_size + self.eps = eps + + # layers + out_dim = math.prod(patch_size) * out_dim + self.norm = WanLayerNorm(dim, eps) + self.head = nn.Linear(dim, out_dim) + + # modulation + self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5) + + def forward(self, x, e): + r""" + Args: + x(Tensor): Shape [B, L1, C] + e(Tensor): Shape [B, L1, C] + """ + with torch.amp.autocast("cuda", dtype=torch.float32): + e = (self.modulation.unsqueeze(0) + e.unsqueeze(2)).chunk(2, dim=2) + x = self.head(self.norm(x) * (1 + e[1].squeeze(2)) + e[0].squeeze(2)) + return x + + +class WanModel(nn.Module): + r""" + Wan diffusion backbone supporting both text-to-video and image-to-video. + """ + + def __init__( + self, + model_type="t2v", + patch_size=(1, 2, 2), + text_len=512, + in_dim=16, + dim=2048, + ffn_dim=8192, + freq_dim=256, + text_dim=4096, + out_dim=16, + num_heads=16, + num_layers=32, + qk_norm=True, + cross_attn_norm=True, + eps=1e-6, + ): + r""" + Initialize the diffusion model backbone. + + Args: + model_type (`str`, *optional*, defaults to 't2v'): + Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video) + patch_size (`tuple`, *optional*, defaults to (1, 2, 2)): + 3D patch dimensions for video embedding (t_patch, h_patch, w_patch) + text_len (`int`, *optional*, defaults to 512): + Fixed length for text embeddings + in_dim (`int`, *optional*, defaults to 16): + Input video channels (C_in) + dim (`int`, *optional*, defaults to 2048): + Hidden dimension of the transformer + ffn_dim (`int`, *optional*, defaults to 8192): + Intermediate dimension in feed-forward network + freq_dim (`int`, *optional*, defaults to 256): + Dimension for sinusoidal time embeddings + text_dim (`int`, *optional*, defaults to 4096): + Input dimension for text embeddings + out_dim (`int`, *optional*, defaults to 16): + Output video channels (C_out) + num_heads (`int`, *optional*, defaults to 16): + Number of attention heads + num_layers (`int`, *optional*, defaults to 32): + Number of transformer blocks + qk_norm (`bool`, *optional*, defaults to True): + Enable query/key normalization + cross_attn_norm (`bool`, *optional*, defaults to False): + Enable cross-attention normalization + eps (`float`, *optional*, defaults to 1e-6): + Epsilon value for normalization layers + """ + + super().__init__() + + if model_type not in ["t2v", "i2v", "ti2v", "s2v"]: + raise ValueError(f"model_type must be one of ['t2v', 'i2v', 'ti2v', 's2v'], got {model_type!r}.") + self.model_type = model_type + + self.patch_size = patch_size + self.text_len = text_len + self.in_dim = in_dim + self.dim = dim + self.ffn_dim = ffn_dim + self.freq_dim = freq_dim + self.text_dim = text_dim + self.out_dim = out_dim + self.num_heads = num_heads + self.num_layers = num_layers + self.qk_norm = qk_norm + self.cross_attn_norm = cross_attn_norm + self.eps = eps + + # embeddings + self.patch_embedding = nn.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size) + self.text_embedding = nn.Sequential( + nn.Linear(text_dim, dim), nn.GELU(approximate="tanh"), nn.Linear(dim, dim) + ) + + self.time_embedding = nn.Sequential(nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim)) + self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6)) + + # blocks + self.blocks = nn.ModuleList( + [ + WanAttentionBlock(dim, ffn_dim, num_heads, qk_norm, cross_attn_norm, eps) + for _ in range(num_layers) + ] + ) + + # head + self.head = Head(dim, out_dim, patch_size, eps) + + # buffers (don't use register_buffer otherwise dtype will be changed in to()) + if (dim % num_heads) != 0 or (dim // num_heads) % 2 != 0: + raise ValueError( + f"dim ({dim}) must be divisible by num_heads ({num_heads}) with an even head dim." + ) + d = dim // num_heads + self.freqs = torch.cat( + [ + rope_params(1024, d - 4 * (d // 6)), + rope_params(1024, 2 * (d // 6)), + rope_params(1024, 2 * (d // 6)), + ], + dim=1, + ) + + # initialize weights + self.init_weights() + + # NOTE: The upstream Wan diffusion forward (flash-attention based) was removed. + # FastWAM's WanVideoDiT subclasses this model, rebuilds `self.blocks` with + # FastWAMAttentionBlock, and provides its own SDPA-based forward. Only the + # constructor (embeddings, blocks, head, rope buffers) and the helpers below + # (unpatchify / init_weights) are reused. WanModel is never run directly. + + def unpatchify(self, x, grid_sizes): + r""" + Reconstruct video tensors from patch embeddings. + + Args: + x (List[Tensor]): + List of patchified features, each with shape [L, C_out * prod(patch_size)] + grid_sizes (Tensor): + Original spatial-temporal grid dimensions before patching, + shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches) + + Returns: + List[Tensor]: + Reconstructed video tensors with shape [C_out, F, H / 8, W / 8] + """ + + c = self.out_dim + out = [] + for u, v in zip(x, grid_sizes.tolist(), strict=False): + u = u[: math.prod(v)].view(*v, *self.patch_size, c) + u = torch.einsum("fhwpqrc->cfphqwr", u) + u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size, strict=False)]) + out.append(u) + return out + + def init_weights(self): + r""" + Initialize model parameters using Xavier initialization. + """ + + # basic init + for m in self.modules(): + if isinstance(m, nn.Linear): + nn.init.xavier_uniform_(m.weight) + if m.bias is not None: + nn.init.zeros_(m.bias) + + # init embeddings + nn.init.xavier_uniform_(self.patch_embedding.weight.flatten(1)) + for m in self.text_embedding.modules(): + if isinstance(m, nn.Linear): + nn.init.normal_(m.weight, std=0.02) + for m in self.time_embedding.modules(): + if isinstance(m, nn.Linear): + nn.init.normal_(m.weight, std=0.02) + + # init output layer + nn.init.zeros_(self.head.head.weight) diff --git a/src/lerobot/policies/fastwam/wan/modular.py b/src/lerobot/policies/fastwam/wan/modular.py new file mode 100644 index 000000000..fac96776b --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/modular.py @@ -0,0 +1,1912 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import logging +import re +from collections.abc import Sequence +from typing import Any + +import torch +import torch.nn as nn +import torch.nn.functional as functional +from PIL import Image + +from .components import ( + WAN22_DIFFUSERS_MODEL_ID, + WAN_T5_TOKENIZER, + build_wan_tokenizer, + load_pretrained_wan_text_encoder, + load_pretrained_wan_vae, + load_wan_video_dit, + resolve_wan_dit_paths, +) +from .video_dit import ( + FastWAMAttentionBlock, + WanContinuousFlowMatchScheduler, + fastwam_masked_attention, + gradient_checkpoint_forward, + modulate, + precompute_freqs_cis, + sinusoidal_embedding_1d, +) + +logger = logging.getLogger(__name__) + + +def _apply_block_norm(block, name: str, x: torch.Tensor) -> torch.Tensor: + apply_norm = getattr(block, f"apply_{name}", None) + if apply_norm is not None: + return apply_norm(x) + return getattr(block, name)(x) + + +class ActionHead(nn.Module): + def __init__(self, hidden_dim: int, out_dim: int, eps: float): + super().__init__() + self.norm = nn.LayerNorm(hidden_dim, eps=eps, elementwise_affine=False) + self.proj = nn.Linear(hidden_dim, out_dim) + self.modulation = nn.Parameter(torch.randn(1, 2, hidden_dim) / hidden_dim**0.5) + + def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor: + shift, scale = (self.modulation.to(dtype=t.dtype, device=t.device) + t.unsqueeze(1)).chunk(2, dim=1) + shift = shift.squeeze(1) + scale = scale.squeeze(1) + return self.proj(self.norm(x) * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)) + + +class ActionDiT(nn.Module): + def __init__( + self, + hidden_dim: int, + action_dim: int, + ffn_dim: int, + text_dim: int, + freq_dim: int, + eps: float, + num_heads: int, + attn_head_dim: int, + num_layers: int, + use_gradient_checkpointing: bool = False, + fp32_attention: bool = True, + ): + super().__init__() + self.hidden_dim = hidden_dim + self.action_dim = action_dim + self.ffn_dim = ffn_dim + self.text_dim = text_dim + self.freq_dim = freq_dim + self.num_heads = num_heads + self.attn_head_dim = attn_head_dim + + if num_heads <= 0: + raise ValueError(f"`num_heads` must be > 0, got {num_heads}") + if attn_head_dim <= 0: + raise ValueError(f"`attn_head_dim` must be > 0, got {attn_head_dim}") + if attn_head_dim % 2 != 0: + raise ValueError(f"`attn_head_dim` must be even for RoPE, got {attn_head_dim}") + + self.action_encoder = nn.Linear(action_dim, hidden_dim) + self.text_embedding = nn.Sequential( + nn.Linear(text_dim, hidden_dim), + nn.GELU(approximate="tanh"), + nn.Linear(hidden_dim, hidden_dim), + ) + self.time_embedding = nn.Sequential( + nn.Linear(freq_dim, hidden_dim), + nn.SiLU(), + nn.Linear(hidden_dim, hidden_dim), + ) + self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(hidden_dim, hidden_dim * 6)) + self.blocks = nn.ModuleList( + [ + FastWAMAttentionBlock( + hidden_dim=hidden_dim, + attn_head_dim=attn_head_dim, + num_heads=num_heads, + ffn_dim=ffn_dim, + eps=eps, + fp32_attention=fp32_attention, + ) + for _ in range(num_layers) + ] + ) + self.head = nn.Linear(hidden_dim, action_dim) + self.freqs = precompute_freqs_cis(attn_head_dim, end=1024) + self.fp32_attention = bool(fp32_attention) + + self.use_gradient_checkpointing = use_gradient_checkpointing + + def pre_dit( + self, + action_tokens: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor | None = None, + ) -> dict[str, Any]: + if action_tokens.ndim != 3: + raise ValueError( + f"`action_tokens` must be 3D [B, T, action_dim], got shape {tuple(action_tokens.shape)}" + ) + if action_tokens.shape[2] != self.action_dim: + raise ValueError( + f"`action_tokens` last dim must be {self.action_dim}, got {action_tokens.shape[2]}" + ) + if timestep.ndim != 1: + raise ValueError(f"`timestep` must be 1D [B] or [1], got shape {tuple(timestep.shape)}") + if context.ndim != 3: + raise ValueError(f"`context` must be 3D [B, L, D], got shape {tuple(context.shape)}") + + batch_size = action_tokens.shape[0] + if context.shape[0] != batch_size: + raise ValueError( + f"Batch mismatch between action tokens and text context: {batch_size} vs {context.shape[0]}" + ) + if timestep.shape[0] not in (1, batch_size): + raise ValueError( + f"`timestep` length must be 1 or batch_size({batch_size}), got {timestep.shape[0]}" + ) + if timestep.shape[0] == 1 and batch_size > 1: + if self.training: + raise ValueError("During training, action timestep length must match batch_size.") + timestep = timestep.expand(batch_size) + + if context_mask is None: + context_mask = torch.ones((batch_size, context.shape[1]), dtype=torch.bool, device=context.device) + else: + if context_mask.ndim != 2: + raise ValueError(f"`context_mask` must be 2D [B, L], got shape {tuple(context_mask.shape)}") + if context_mask.shape[0] != batch_size or context_mask.shape[1] != context.shape[1]: + raise ValueError( + f"`context_mask` shape must match `context` shape [B, L], got {tuple(context_mask.shape)} vs {tuple(context.shape)}" + ) + + seq_len = action_tokens.shape[1] + if seq_len > self.freqs.shape[0]: + raise ValueError(f"Action token length {seq_len} exceeds RoPE cache {self.freqs.shape[0]}.") + + model_dtype = self.action_encoder.weight.dtype + action_tokens = action_tokens.to(dtype=model_dtype) + context = context.to(dtype=model_dtype) + t_emb = sinusoidal_embedding_1d(self.freq_dim, timestep).to(dtype=model_dtype) + t = self.time_embedding(t_emb) + t_mod = self.time_projection(t).unflatten(1, (6, self.hidden_dim)) + + tokens = self.action_encoder(action_tokens) + context_emb = self.text_embedding(context) + context_attn_mask = context_mask.unsqueeze(1).expand(-1, seq_len, -1) + freqs = self.freqs[:seq_len].view(seq_len, 1, -1).to(tokens.device) + + return { + "tokens": tokens, + "freqs": freqs, + "t": t, + "t_mod": t_mod, + "context": context_emb, + "context_mask": context_attn_mask, + "meta": { + "batch_size": batch_size, + "seq_len": seq_len, + }, + } + + def post_dit(self, tokens: torch.Tensor, pre_state: dict[str, Any]) -> torch.Tensor: + return self.head(tokens) + + def forward( + self, + action_tokens: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor | None = None, + ) -> torch.Tensor: + pre_state = self.pre_dit( + action_tokens=action_tokens, + timestep=timestep, + context=context, + context_mask=context_mask, + ) + x = pre_state["tokens"] + context = pre_state["context"] + t_mod = pre_state["t_mod"] + freqs = pre_state["freqs"] + context_mask = pre_state["context_mask"] + + for block in self.blocks: + if self.use_gradient_checkpointing: + x = gradient_checkpoint_forward( + block, + self.use_gradient_checkpointing, + x, + context, + t_mod, + freqs, + context_mask=context_mask, + ) + else: + x = block(x, context, t_mod, freqs, context_mask=context_mask) + + return self.post_dit(x, pre_state) + + +class MoTLayer(nn.Module): + """A single MoT layer: owns one transformer block per expert and runs the cross-expert + mixed-attention step for that layer. + + This exists as a module — rather than the per-layer work being inlined in ``MoT``'s loop — + so FSDP can wrap each layer as its own unit. FSDP all-gathers a wrapped module's sharded + parameters via a hook on that module's ``forward``/``__call__``. ``MoT`` drives block + submodules directly (the joint mixed attention concatenates Q/K/V across experts, so no + single block's ``forward`` is ever called), so ``MoTLayer.forward`` is the only call + boundary FSDP can hook. All three per-layer paths therefore dispatch through + ``forward(mode=...)`` so each enters via ``__call__``. + """ + + def __init__( + self, + blocks: dict[str, nn.Module], + experts: dict[str, nn.Module], + num_heads: int, + attn_head_dim: int, + fp32_attention: bool, + mot_checkpoint_mixed_attn: bool, + ): + super().__init__() + # Registered owner of this layer's blocks (one per expert) — the FSDP wrap unit. + self.blocks = nn.ModuleDict(blocks) + self.expert_order = list(blocks.keys()) + # Unregistered back-references to the experts: used only to read the live + # `use_gradient_checkpointing` flag, kept out of parameters()/state_dict(). + object.__setattr__(self, "_experts", dict(experts)) + self.num_heads = num_heads + self.attn_head_dim = attn_head_dim + self.fp32_attention = bool(fp32_attention) + self.mot_checkpoint_mixed_attn = bool(mot_checkpoint_mixed_attn) + + @staticmethod + def _split_modulation(block, t_mod: torch.Tensor): + has_seq = len(t_mod.shape) == 4 + chunk_dim = 2 if has_seq else 1 + + base_mod = block.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (base_mod + t_mod).chunk( + 6, dim=chunk_dim + ) + if has_seq: + # means t_mod has separate modulation for each token, otherwise same modulation for all tokens in the block + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + shift_msa.squeeze(2), + scale_msa.squeeze(2), + gate_msa.squeeze(2), + shift_mlp.squeeze(2), + scale_mlp.squeeze(2), + gate_mlp.squeeze(2), + ) + return shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp + + def _mixed_attention( + self, + q_cat: torch.Tensor, + k_cat: torch.Tensor, + v_cat: torch.Tensor, + attention_mask: torch.Tensor, + ) -> torch.Tensor: + attn_mask = attention_mask.to(device=q_cat.device) + + def _forward(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor: + return fastwam_masked_attention( + q=q, + k=k, + v=v, + num_heads=self.num_heads, + ctx_mask=attn_mask, + fp32_attention=self.fp32_attention, + ) + + if self.mot_checkpoint_mixed_attn and self.training: + return torch.utils.checkpoint.checkpoint( + _forward, + q_cat, + k_cat, + v_cat, + use_reentrant=False, + ) + return _forward(q_cat, k_cat, v_cat) + + @staticmethod + def _apply_expert_post_block( + block, + residual_x: torch.Tensor, + mixed_attn_out: torch.Tensor, + gate_msa: torch.Tensor, + shift_mlp: torch.Tensor, + scale_mlp: torch.Tensor, + gate_mlp: torch.Tensor, + context_payload: dict | None, + ) -> torch.Tensor: + if hasattr(block, "project_self_attention_output"): + projected_attn = block.project_self_attention_output(mixed_attn_out) + else: + projected_attn = block.self_attn.o(mixed_attn_out.to(dtype=block.self_attn.o.weight.dtype)) + x = residual_x + gate_msa * projected_attn + + if context_payload is not None: + context = context_payload.get("context") + if context is not None: + context_mask = context_payload.get("mask") + if context_mask is not None and context_mask.dim() == 3: + context_mask = context_mask.unsqueeze(1) + x = x + block.apply_cross_attention( + _apply_block_norm(block, "norm3", x), + context, + context_mask=context_mask, + ) + + mlp_input = modulate(_apply_block_norm(block, "norm2", x), shift_mlp, scale_mlp) + x = x + gate_mlp * block.ffn(mlp_input) + return x + + def _build_expert_attention_io( + self, + name: str, + x: torch.Tensor, + freqs: torch.Tensor | dict[str, torch.Tensor], + t_mod: torch.Tensor, + ): + """Build this expert's attention tensors and post-block states for the layer. + + Returns (q, k, v, residual_x, gate_msa, shift_mlp, scale_mlp, gate_mlp, use_gc). + """ + block = self.blocks[name] + expert = self._experts[name] + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self._split_modulation(block, t_mod) + attn_input = modulate(_apply_block_norm(block, "norm1", x), shift_msa, scale_msa) + + q, k, v = block.project_self_attention(attn_input, freqs) + + use_gradient_checkpointing = bool(getattr(expert, "use_gradient_checkpointing", False)) + return ( + q, + k, + v, + x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + use_gradient_checkpointing, + ) + + def _apply_post_with_optional_checkpoint( + self, + block, + residual_x: torch.Tensor, + gate_msa: torch.Tensor, + shift_mlp: torch.Tensor, + scale_mlp: torch.Tensor, + gate_mlp: torch.Tensor, + use_gradient_checkpointing: bool, + mixed_slice: torch.Tensor, + context_payload: dict | None, + ) -> torch.Tensor: + def _post_fn( + _mixed_slice: torch.Tensor, + _x: torch.Tensor, + _gate_msa: torch.Tensor, + _shift_mlp: torch.Tensor, + _scale_mlp: torch.Tensor, + _gate_mlp: torch.Tensor, + _block=block, + _context_payload=context_payload, + ) -> torch.Tensor: + return self._apply_expert_post_block( + block=_block, + residual_x=_x, + mixed_attn_out=_mixed_slice, + gate_msa=_gate_msa, + shift_mlp=_shift_mlp, + scale_mlp=_scale_mlp, + gate_mlp=_gate_mlp, + context_payload=_context_payload, + ) + + if use_gradient_checkpointing and self.training: + return torch.utils.checkpoint.checkpoint( + _post_fn, + mixed_slice, + residual_x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + use_reentrant=False, + ) + return _post_fn( + mixed_slice, + residual_x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + ) + + def forward(self, mode: str, **kwargs): + if mode == "joint": + return self._forward_joint(**kwargs) + if mode == "video_prefill": + return self._forward_video_prefill(**kwargs) + if mode == "action_cached": + return self._forward_action_cached(**kwargs) + raise ValueError(f"Unknown MoTLayer forward mode: {mode!r}") + + def _forward_joint( + self, + tokens_all: dict[str, torch.Tensor], + attention_mask: torch.Tensor, + freqs_all: dict[str, torch.Tensor], + context_all: dict[str, dict | None], + t_mod_all: dict[str, torch.Tensor], + ) -> dict[str, torch.Tensor]: + q_chunks = [] + k_chunks = [] + v_chunks = [] + cached = {} + seq_lens = [] + + for name in self.expert_order: + ( + q, + k, + v, + residual_x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + use_gradient_checkpointing, + ) = self._build_expert_attention_io(name, tokens_all[name], freqs_all[name], t_mod_all[name]) + + q_chunks.append(q) + k_chunks.append(k) + v_chunks.append(v) + seq_lens.append(tokens_all[name].shape[1]) + cached[name] = { + "residual_x": residual_x, + "gate_msa": gate_msa, + "shift_mlp": shift_mlp, + "scale_mlp": scale_mlp, + "gate_mlp": gate_mlp, + "use_gradient_checkpointing": use_gradient_checkpointing, + } + + q_cat = torch.cat(q_chunks, dim=1) + k_cat = torch.cat(k_chunks, dim=1) + v_cat = torch.cat(v_chunks, dim=1) + + total_seq = q_cat.shape[1] + if attention_mask.shape[0] != total_seq: + raise ValueError( + f"Attention mask seq length mismatch: mask={attention_mask.shape[0]} vs tokens={total_seq}" + ) + + mixed = self._mixed_attention(q_cat=q_cat, k_cat=k_cat, v_cat=v_cat, attention_mask=attention_mask) + + out = {} + start = 0 + for name, seq_len in zip(self.expert_order, seq_lens, strict=True): + end = start + seq_len + mixed_slice = mixed[:, start:end, :] + cached_expert = cached[name] + out[name] = self._apply_post_with_optional_checkpoint( + block=self.blocks[name], + residual_x=cached_expert["residual_x"], + gate_msa=cached_expert["gate_msa"], + shift_mlp=cached_expert["shift_mlp"], + scale_mlp=cached_expert["scale_mlp"], + gate_mlp=cached_expert["gate_mlp"], + use_gradient_checkpointing=cached_expert["use_gradient_checkpointing"], + mixed_slice=mixed_slice, + context_payload=context_all.get(name), + ) + start = end + return out + + def _forward_video_prefill( + self, + x: torch.Tensor, + freqs: torch.Tensor, + t_mod: torch.Tensor, + context_payload: dict | None, + video_attention_mask: torch.Tensor, + ): + ( + q, + k, + v, + residual_x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + use_gradient_checkpointing, + ) = self._build_expert_attention_io("video", x, freqs, t_mod) + # Video prefill uses only video self-attention mask. + mixed = self._mixed_attention(q_cat=q, k_cat=k, v_cat=v, attention_mask=video_attention_mask) + x_out = self._apply_post_with_optional_checkpoint( + block=self.blocks["video"], + residual_x=residual_x, + gate_msa=gate_msa, + shift_mlp=shift_mlp, + scale_mlp=scale_mlp, + gate_mlp=gate_mlp, + use_gradient_checkpointing=use_gradient_checkpointing, + mixed_slice=mixed, + context_payload=context_payload, + ) + return x_out, k, v + + def _forward_action_cached( + self, + x: torch.Tensor, + freqs: torch.Tensor, + t_mod: torch.Tensor, + context_payload: dict | None, + k_video: torch.Tensor, + v_video: torch.Tensor, + action_attention_mask: torch.Tensor, + ) -> torch.Tensor: + ( + q_action, + k_action, + v_action, + residual_x, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + use_gradient_checkpointing, + ) = self._build_expert_attention_io("action", x, freqs, t_mod) + # Mixed attention: action queries attend to cached video K/V plus current action K/V. + k_cat = torch.cat([k_video, k_action], dim=1) + v_cat = torch.cat([v_video, v_action], dim=1) + mixed = self._mixed_attention( + q_cat=q_action, k_cat=k_cat, v_cat=v_cat, attention_mask=action_attention_mask + ) + return self._apply_post_with_optional_checkpoint( + block=self.blocks["action"], + residual_x=residual_x, + gate_msa=gate_msa, + shift_mlp=shift_mlp, + scale_mlp=scale_mlp, + gate_mlp=gate_mlp, + use_gradient_checkpointing=use_gradient_checkpointing, + mixed_slice=mixed, + context_payload=context_payload, + ) + + +class MoT(nn.Module): + def __init__( + self, + mixtures: dict[str, nn.Module], + mot_checkpoint_mixed_attn: bool = True, + ): + super().__init__() + if not mixtures: + raise ValueError("`mixtures` cannot be empty.") + if "video" not in mixtures or "action" not in mixtures: + raise ValueError("`mixtures` must include both 'video' and 'action' experts.") + + self.mixtures = nn.ModuleDict(mixtures) + self.expert_order = list(self.mixtures.keys()) + self.mot_checkpoint_mixed_attn = mot_checkpoint_mixed_attn + if mot_checkpoint_mixed_attn: + logger.info( + "Using gradient checkpointing for mixture attention. This will save memory but use more computation." + ) + + first_expert = self.mixtures[self.expert_order[0]] + self.num_layers = len(first_expert.blocks) + self.num_heads = first_expert.num_heads + self.attn_head_dim = first_expert.attn_head_dim + self.fp32_attention = bool(getattr(first_expert, "fp32_attention", True)) + + for name in self.expert_order[1:]: + expert = self.mixtures[name] + if len(expert.blocks) != self.num_layers: + raise ValueError( + f"All experts must have same number of layers; got {self.num_layers} and {len(expert.blocks)}" + ) + if expert.num_heads != self.num_heads: + raise ValueError( + f"All experts must have same num_heads; got {self.num_heads} and {expert.num_heads}" + ) + if expert.attn_head_dim != self.attn_head_dim: + raise ValueError( + "All experts must have same attn_head_dim; " + f"got {self.attn_head_dim} and {expert.attn_head_dim}" + ) + if bool(getattr(expert, "fp32_attention", True)) != self.fp32_attention: + raise ValueError("All experts must use the same `fp32_attention` setting.") + + logger.info(f"Initialized MoT with experts: {self.expert_order}, num_layers={self.num_layers}") + for name in self.expert_order: + expert = self.mixtures[name] + logger.info( + f" Expert '{name}': num_params={sum(p.numel() for p in expert.parameters()) / 1e9:.2f} B" + ) + + # One MoTLayer per layer, each owning that layer's block from every expert. This is the + # FSDP wrap unit: only MoTLayer.forward is ever called (MoT drives block submodules + # directly for the cross-expert mixed attention), so it is the boundary at which FSDP can + # all-gather a layer's params. The blocks are RE-PARENTED into the layers — removed from + # each expert's module registry — so they have a single owner; leaving them registered + # under both the expert and the layer would make FSDP try to manage the same params twice. + self.layers = nn.ModuleList( + [ + MoTLayer( + blocks={name: self.mixtures[name].blocks[layer_idx] for name in self.expert_order}, + experts={name: self.mixtures[name] for name in self.expert_order}, + num_heads=self.num_heads, + attn_head_dim=self.attn_head_dim, + fp32_attention=self.fp32_attention, + mot_checkpoint_mixed_attn=self.mot_checkpoint_mixed_attn, + ) + for layer_idx in range(self.num_layers) + ] + ) + for name in self.expert_order: + expert = self.mixtures[name] + kept_blocks = list(expert.blocks) + del expert._modules["blocks"] + # Keep an UNREGISTERED reference so the (unused) standalone `expert.forward` and any + # `len(expert.blocks)` still work, without re-adding the params to the expert's + # parameters()/state_dict() (which would double-register them with the MoTLayer owner). + object.__setattr__(expert, "blocks", kept_blocks) + + def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs): + # Backward-compat for checkpoints saved before the MoTLayer refactor. Then the per-layer + # blocks were keyed under the experts (`{prefix}mixtures..blocks..`, e.g. + # the released `ZibinDong/fastwam_libero_uncond_2cam224`); now they are owned by the layers + # (`{prefix}layers..blocks..`). Remap legacy keys in place so the recursion + # into `self.layers` finds them and the (now block-less) `self.mixtures` does not flag them. + legacy = re.compile(re.escape(prefix) + r"mixtures\.([^.]+)\.blocks\.(\d+)\.(.+)$") + moved = {} + for key in list(state_dict.keys()): + m = legacy.match(key) + if m is not None: + name, layer_idx, rest = m.group(1), m.group(2), m.group(3) + moved[f"{prefix}layers.{layer_idx}.blocks.{name}.{rest}"] = state_dict.pop(key) + state_dict.update(moved) + super()._load_from_state_dict(state_dict, prefix, *args, **kwargs) + + def prefill_video_cache( + self, + video_tokens: torch.Tensor, + video_freqs: torch.Tensor, + video_t_mod: torch.Tensor, + video_context_payload: dict | None, + video_attention_mask: torch.Tensor, + ) -> list[dict[str, torch.Tensor]]: + """Prefill video branch once and cache per-layer K/V for action denoising. + + Returns a list of length ``num_layers``, each entry ``{"k": ..., "v": ...}``. + """ + if "video" not in self.mixtures: + raise ValueError("MoT requires `video` expert for `prefill_video_cache`.") + if video_attention_mask.ndim != 2: + raise ValueError( + f"`video_attention_mask` must be 2D [S,S], got shape {tuple(video_attention_mask.shape)}" + ) + if video_attention_mask.shape[0] != video_attention_mask.shape[1]: + raise ValueError( + f"`video_attention_mask` must be square, got shape {tuple(video_attention_mask.shape)}" + ) + if video_attention_mask.shape[0] != video_tokens.shape[1]: + raise ValueError( + "`video_attention_mask` seq length mismatch: " + f"mask={video_attention_mask.shape[0]} vs tokens={video_tokens.shape[1]}" + ) + + x = video_tokens + kv_cache: list[dict[str, torch.Tensor]] = [] + for layer in self.layers: + x, k, v = layer( + mode="video_prefill", + x=x, + freqs=video_freqs, + t_mod=video_t_mod, + context_payload=video_context_payload, + video_attention_mask=video_attention_mask, + ) + kv_cache.append({"k": k, "v": v}) + return kv_cache + + def forward_action_with_video_cache( + self, + action_tokens: torch.Tensor, + action_freqs: torch.Tensor, + action_t_mod: torch.Tensor, + action_context_payload: dict | None, + video_kv_cache: list[dict[str, torch.Tensor]], + attention_mask: torch.Tensor, + video_seq_len: int, + ) -> torch.Tensor: + """Run action branch with cached video K/V instead of recomputing video tokens.""" + if "action" not in self.mixtures: + raise ValueError("MoT requires `action` expert for `forward_action_with_video_cache`.") + if len(video_kv_cache) != self.num_layers: + raise ValueError( + f"`video_kv_cache` must contain {self.num_layers} layers, got {len(video_kv_cache)}." + ) + if attention_mask.ndim != 2: + raise ValueError(f"`attention_mask` must be 2D [S,S], got shape {tuple(attention_mask.shape)}") + if attention_mask.shape[0] != attention_mask.shape[1]: + raise ValueError(f"`attention_mask` must be square, got shape {tuple(attention_mask.shape)}") + + action_seq_len = int(action_tokens.shape[1]) + total_seq_len = int(video_seq_len) + action_seq_len + if attention_mask.shape[0] != total_seq_len: + raise ValueError( + "`attention_mask` seq length mismatch: " + f"mask={attention_mask.shape[0]} vs expected_total={total_seq_len}" + ) + # Use the action query rows from the joint [video+action] mask. + action_attention_mask = attention_mask[video_seq_len:total_seq_len, :total_seq_len] + + x = action_tokens + for layer_idx, layer in enumerate(self.layers): + layer_cache = video_kv_cache[layer_idx] + if "k" not in layer_cache or "v" not in layer_cache: + raise ValueError(f"`video_kv_cache[{layer_idx}]` must contain `k` and `v`.") + k_video = layer_cache["k"] + v_video = layer_cache["v"] + if k_video.shape[1] != video_seq_len or v_video.shape[1] != video_seq_len: + raise ValueError(f"`video_kv_cache[{layer_idx}]` seq len mismatch, expected {video_seq_len}.") + x = layer( + mode="action_cached", + x=x, + freqs=action_freqs, + t_mod=action_t_mod, + context_payload=action_context_payload, + k_video=k_video, + v_video=v_video, + action_attention_mask=action_attention_mask, + ) + return x + + def forward( + self, + embeds_all: dict[str, torch.Tensor], + attention_mask: torch.Tensor, + freqs_all: dict[str, torch.Tensor], + context_all: dict[str, dict | None], + t_mod_all: dict[str, torch.Tensor], + ): + missing = [k for k in self.expert_order if k not in embeds_all] + if missing: + raise ValueError(f"Missing expert tokens for {missing}") + missing = [k for k in self.expert_order if k not in freqs_all] + if missing: + raise ValueError(f"Missing expert freqs for {missing}") + missing = [k for k in self.expert_order if k not in t_mod_all] + if missing: + raise ValueError(f"Missing expert t_mod for {missing}") + + if attention_mask.ndim != 2: + raise ValueError(f"`attention_mask` must be 2D [S, S], got shape {tuple(attention_mask.shape)}") + if attention_mask.shape[0] != attention_mask.shape[1]: + raise ValueError(f"`attention_mask` must be square, got shape {tuple(attention_mask.shape)}") + + # Each layer is a MoTLayer module; entering via __call__ lets FSDP all-gather that + # layer's params (the whole point of the per-layer split). + tokens_all = dict(embeds_all) + for layer in self.layers: + tokens_all = layer( + mode="joint", + tokens_all=tokens_all, + attention_mask=attention_mask, + freqs_all=freqs_all, + context_all=context_all, + t_mod_all=t_mod_all, + ) + return tokens_all + + +class FastWAM(torch.nn.Module): + """MoT world model with video/action experts.""" + + def __init__( + self, + video_expert, + action_expert: ActionDiT, + mot: MoT, + vae, + text_encoder=None, + tokenizer=None, + text_dim: int | None = None, + proprio_dim: int | None = None, + device: str = "cpu", + torch_dtype: torch.dtype = torch.float32, + video_train_shift: float = 5.0, + video_infer_shift: float = 5.0, + video_num_train_timesteps: int = 1000, + action_train_shift: float = 5.0, + action_infer_shift: float = 5.0, + action_num_train_timesteps: int = 1000, + loss_lambda_video: float = 1.0, + loss_lambda_action: float = 1.0, + ): + super().__init__() + self.mot = mot + # `video_expert` / `action_expert` are the very same module objects as + # `mot.mixtures["video"]` / `["action"]`, and `dit` is an alias of `mot`. Registering + # them as submodules too would give every expert tensor three names in `state_dict()` + # (`video_expert.*`, `mot.mixtures.video.*`, `dit.mixtures.video.*`) — a 3x-bloated + # gathered FSDP checkpoint and a doubled module tree for FSDP to traverse. Hold them as + # plain (unregistered) attributes instead — bypassing `nn.Module.__setattr__`, like the + # frozen vae/text_encoder below — so `mot` is the single registered owner and each tensor + # has one canonical name (`mot.mixtures.*` / `mot.layers.*`, matching the base checkpoint). + # Forward / freeze / optimizer code still reaches them by attribute, and device/dtype moves + # still apply via `mot`. (optimizer + freeze logic use `model.dit`.) + object.__setattr__(self, "video_expert", video_expert) + object.__setattr__(self, "action_expert", action_expert) + object.__setattr__(self, "dit", self.mot) + + # Frozen Wan2.2 components: bypass `nn.Module.__setattr__` so they are NOT + # registered as submodules. They are therefore excluded from `state_dict()` + # (lean checkpoints), `parameters()`, and DDP gradient sync, and are loaded + # with their real weights from the diffusers/transformers repos at construction. + # Device/dtype moves still reach them via the `_apply` override below. + object.__setattr__(self, "vae", vae) + object.__setattr__(self, "text_encoder", text_encoder) + self.tokenizer = tokenizer + vae.requires_grad_(False) + if text_encoder is not None: + text_encoder.requires_grad_(False) + if text_dim is None: + if self.text_encoder is None: + raise ValueError("`text_dim` is required when `text_encoder` is not loaded.") + text_dim = int(self.text_encoder.dim) + self.text_dim = int(text_dim) + self.proprio_dim = None if proprio_dim is None else int(proprio_dim) + if self.proprio_dim is not None: + self.proprio_encoder = nn.Linear(self.proprio_dim, self.text_dim).to(torch_dtype) + else: + self.proprio_encoder = None + + self.train_video_scheduler = WanContinuousFlowMatchScheduler( + num_train_timesteps=video_num_train_timesteps, + shift=video_train_shift, + ) + self.infer_video_scheduler = WanContinuousFlowMatchScheduler( + num_train_timesteps=video_num_train_timesteps, + shift=video_infer_shift, + ) + self.train_action_scheduler = WanContinuousFlowMatchScheduler( + num_train_timesteps=action_num_train_timesteps, + shift=action_train_shift, + ) + self.infer_action_scheduler = WanContinuousFlowMatchScheduler( + num_train_timesteps=action_num_train_timesteps, + shift=action_infer_shift, + ) + # Optional aliases for consistency with Wan22Core naming. + self.train_scheduler = self.train_video_scheduler + self.infer_scheduler = self.infer_video_scheduler + + self.device = torch.device(device) + self.torch_dtype = torch_dtype + self.loss_lambda_video = float(loss_lambda_video) + self.loss_lambda_action = float(loss_lambda_action) + + self.to(self.device) + + @classmethod + def from_wan22_pretrained( + cls, + device: str = "cuda", + torch_dtype: torch.dtype = torch.bfloat16, + model_id: str = "Wan-AI/Wan2.2-TI2V-5B", + tokenizer_model_id: str = WAN_T5_TOKENIZER, + text_encoder_model_id: str = WAN22_DIFFUSERS_MODEL_ID, + tokenizer_max_len: int = 512, + load_text_encoder: bool = True, + proprio_dim: int | None = None, + video_dit_config: dict[str, Any] | None = None, + action_dit_config: dict[str, Any] | None = None, + mot_checkpoint_mixed_attn: bool = True, + video_train_shift: float = 5.0, + video_infer_shift: float = 5.0, + video_num_train_timesteps: int = 1000, + action_train_shift: float = 5.0, + action_infer_shift: float = 5.0, + action_num_train_timesteps: int = 1000, + loss_lambda_video: float = 1.0, + loss_lambda_action: float = 1.0, + ): + if video_dit_config is None: + raise ValueError("`video_dit_config` is required for FastWAM.from_wan22_pretrained().") + if "text_dim" not in video_dit_config: + raise ValueError("`video_dit_config['text_dim']` is required for FastWAM.") + + # Custom MoT video DiT from the original Wan2.2 repo; frozen VAE / UMT5 from + # the diffusers conversion. This is the offline base-creation path; the + # weights it loads are then bundled into the FastWAM `model.safetensors`. + video_expert = load_wan_video_dit( + resolve_wan_dit_paths(model_id), + dit_config=video_dit_config, + torch_dtype=torch_dtype, + device=device, + ) + action_expert = ActionDiT(**action_dit_config).to(device=device, dtype=torch_dtype) + if int(action_expert.num_heads) != int(video_expert.num_heads): + raise ValueError("ActionDiT `num_heads` must match video expert for MoT mixed attention.") + if int(action_expert.attn_head_dim) != int(video_expert.attn_head_dim): + raise ValueError("ActionDiT `attn_head_dim` must match video expert for MoT mixed attention.") + if int(len(action_expert.blocks)) != int(len(video_expert.blocks)): + raise ValueError("ActionDiT `num_layers` must match video expert.") + + mot = MoT( + mixtures={"video": video_expert, "action": action_expert}, + mot_checkpoint_mixed_attn=mot_checkpoint_mixed_attn, + ) + + vae = load_pretrained_wan_vae(torch_dtype=torch_dtype, device=device) + text_encoder = ( + load_pretrained_wan_text_encoder( + model_id=text_encoder_model_id, torch_dtype=torch_dtype, device=device + ) + if load_text_encoder + else None + ) + tokenizer = build_wan_tokenizer(model_id=tokenizer_model_id, tokenizer_max_len=tokenizer_max_len) + + return cls( + video_expert=video_expert, + action_expert=action_expert, + mot=mot, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_dim=int(video_dit_config["text_dim"]), + proprio_dim=proprio_dim, + device=device, + torch_dtype=torch_dtype, + video_train_shift=video_train_shift, + video_infer_shift=video_infer_shift, + video_num_train_timesteps=video_num_train_timesteps, + action_train_shift=action_train_shift, + action_infer_shift=action_infer_shift, + action_num_train_timesteps=action_num_train_timesteps, + loss_lambda_video=loss_lambda_video, + loss_lambda_action=loss_lambda_action, + ) + + def _apply(self, fn, *args, **kwargs): + # `.to()` / `.cuda()` / `.cpu()` and accelerate/DDP device moves all funnel + # through `_apply`, and the parent policy reaches us via `child._apply(fn)` + # (not `child.to()`). Propagate `fn` to the *unregistered* frozen VAE / text + # encoder here so they follow the rest of the model onto the right device, + # while staying out of `state_dict()` / `parameters()`. + super()._apply(fn, *args, **kwargs) + self.vae._apply(fn) + if self.text_encoder is not None: + self.text_encoder._apply(fn) + return self + + @staticmethod + def _check_resize_height_width(height, width, num_frames): + if height % 16 != 0: + height = (height + 15) // 16 * 16 + if width % 16 != 0: + width = (width + 15) // 16 * 16 + if num_frames % 4 != 1: + num_frames = (num_frames + 3) // 4 * 4 + 1 + return height, width, num_frames + + @torch.no_grad() + def encode_prompt(self, prompt: str | Sequence[str]): + if self.text_encoder is None or self.tokenizer is None: + raise ValueError( + "Prompt encoding requires loaded text encoder/tokenizer. " + "Set `load_text_encoder=true` or provide precomputed `context/context_mask`." + ) + ids, mask = self.tokenizer(prompt, return_mask=True, add_special_tokens=True) + ids = ids.to(self.device) + mask = mask.to(self.device, dtype=torch.bool) + prompt_emb = self.text_encoder(ids, mask) + seq_lens = mask.gt(0).sum(dim=1).long() + for i, v in enumerate(seq_lens): + prompt_emb[i, v:] = 0 + # Match FastWAM/Wan2.2 context semantics: padding embeddings are zeroed, + # while cross-attention still sees a fixed-length context. + mask = torch.ones_like(mask) + return prompt_emb.to(device=self.device), mask + + def _append_proprio_to_context( + self, + context: torch.Tensor, + context_mask: torch.Tensor, + proprio: torch.Tensor | None, + ) -> tuple[torch.Tensor, torch.Tensor]: + if self.proprio_encoder is None or proprio is None: + return context, context_mask + if proprio.ndim != 2: + raise ValueError(f"`proprio` must be 2D [B, D], got shape {tuple(proprio.shape)}") + if self.proprio_dim is None or proprio.shape[1] != self.proprio_dim: + raise ValueError(f"`proprio` last dim must be {self.proprio_dim}, got {proprio.shape[1]}") + proprio_token = self.proprio_encoder( + proprio.to(device=self.device, dtype=context.dtype).unsqueeze(1) + ).to(dtype=context.dtype) # [B, 1, D] + proprio_mask = torch.ones((context_mask.shape[0], 1), dtype=torch.bool, device=context_mask.device) + return ( + torch.cat([context, proprio_token], dim=1), + torch.cat([context_mask, proprio_mask], dim=1), + ) + + @torch.no_grad() + def _encode_video_latents(self, video_tensor, tiled=False, tile_size=(30, 52), tile_stride=(15, 26)): + # The Wan VAE expects pixels in [-1, 1]; model inputs arrive in [0, 1] (VISUAL is IDENTITY in + # the preprocessor — see configuration_fastwam.normalization_mapping). Map here, at the single + # video-encode boundary, so it is applied exactly once on every path. + video_tensor = video_tensor * 2.0 - 1.0 + z = self.vae.encode( + video_tensor, + device=self.device, + tiled=tiled, + tile_size=tile_size, + tile_stride=tile_stride, + ) + return z + + @torch.no_grad() + def _encode_input_image_latents_tensor( + self, input_image: torch.Tensor, tiled=False, tile_size=(30, 52), tile_stride=(15, 26) + ): + if input_image.ndim == 3: + input_image = input_image.unsqueeze(0) + if input_image.ndim != 4 or input_image.shape[0] != 1 or input_image.shape[1] != 3: + raise ValueError( + f"`input_image` must have shape [1,3,H,W] or [3,H,W], got {tuple(input_image.shape)}" + ) + # [0, 1] -> [-1, 1] for the Wan VAE (mirrors `_encode_video_latents`); single image-encode boundary. + input_image = input_image * 2.0 - 1.0 + image = input_image.to(device=self.device)[0].unsqueeze(1) + z = self.vae.encode( + [image], device=self.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride + ) + if isinstance(z, list): + z = z[0].unsqueeze(0) + return z + + def _decode_latents(self, latents, tiled=False, tile_size=(30, 52), tile_stride=(15, 26)): + video_tensor = self.vae.decode( + latents, device=self.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride + ) + video_tensor = video_tensor.squeeze(0).detach().float().clamp(-1, 1) + video_tensor = ((video_tensor + 1.0) * 127.5).to(torch.uint8).cpu() + frames = [] + for t in range(video_tensor.shape[1]): + frame = video_tensor[:, t].permute(1, 2, 0).numpy() + frames.append(Image.fromarray(frame)) + return frames + + def build_inputs(self, sample, tiled: bool = False): + video = sample["video"] + if "context" not in sample or "context_mask" not in sample: + raise ValueError("FastWAM training requires `sample['context']` and `sample['context_mask']`.") + context = sample["context"] + context_mask = sample["context_mask"] + proprio = sample.get("proprio", None) + if video.ndim != 5: + raise ValueError(f"`sample['video']` must be 5D [B, 3, T, H, W], got shape {tuple(video.shape)}") + if video.shape[1] != 3: + raise ValueError(f"`sample['video']` channel dimension must be 3, got shape {tuple(video.shape)}") + + batch_size, _, num_frames, height, width = video.shape + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"Video spatial dims must be multiples of 16, got H={height}, W={width}") + if num_frames % 4 != 1: + raise ValueError(f"Video T must satisfy T % 4 == 1, got T={num_frames}") + if num_frames <= 1: + raise ValueError(f"Video T must be > 1 for action-conditioned training, got T={num_frames}") + + if "action" not in sample: + raise ValueError("`sample['action']` is required for FastWAM training.") + + action = sample["action"] + if action.ndim != 3: + raise ValueError(f"`sample['action']` must be 3D [B, T, a_dim], got shape {tuple(action.shape)}") + action_horizon = int(action.shape[1]) + if action_horizon % (num_frames - 1) != 0: + raise ValueError( + f"`sample['action']` temporal dimension must be divisible by video transitions ({num_frames - 1}), got {action_horizon}" + ) + + action_is_pad = sample.get("action_is_pad", None) + if action_is_pad is not None: + if action_is_pad.ndim != 2: + raise ValueError( + f"`sample['action_is_pad']` must be 2D [B, T], got shape {tuple(action_is_pad.shape)}" + ) + if action_is_pad.shape[0] != batch_size or action_is_pad.shape[1] != action_horizon: + raise ValueError( + "`sample['action_is_pad']` shape mismatch: " + f"got {tuple(action_is_pad.shape)} vs expected ({batch_size}, {action_horizon})" + ) + + image_is_pad = sample.get("image_is_pad", None) + if image_is_pad is not None: + if image_is_pad.ndim != 2: + raise ValueError( + f"`sample['image_is_pad']` must be 2D [B, T], got shape {tuple(image_is_pad.shape)}" + ) + if image_is_pad.shape[0] != batch_size or image_is_pad.shape[1] != num_frames: + raise ValueError( + "`sample['image_is_pad']` shape mismatch: " + f"got {tuple(image_is_pad.shape)} vs expected ({batch_size}, {num_frames})" + ) + + input_video = video.to(device=self.device, dtype=self.torch_dtype, non_blocking=True) + input_latents = self._encode_video_latents(input_video, tiled=tiled) + + first_frame_latents = None + fuse_flag = False + if getattr(self.video_expert, "fuse_vae_embedding_in_latents", False): + first_frame_latents = input_latents[:, :, 0:1] + fuse_flag = True + + if context.ndim != 3 or context_mask.ndim != 2: + raise ValueError( + f"`context/context_mask` must be [B,L,D]/[B,L], got {tuple(context.shape)} and {tuple(context_mask.shape)}" + ) + context = context.to(device=self.device, dtype=self.torch_dtype, non_blocking=True) + context_mask = context_mask.to(device=self.device, dtype=torch.bool, non_blocking=True) + if self.proprio_encoder is not None: + if proprio is None: + raise ValueError("`sample['proprio']` is required when `proprio_dim` is enabled.") + if proprio.ndim != 3: + raise ValueError( + f"`sample['proprio']` must be 3D [B, T, d], got shape {tuple(proprio.shape)}" + ) + if proprio.shape[2] != self.proprio_dim: + raise ValueError( + f"`sample['proprio']` last dim must be {self.proprio_dim}, got {proprio.shape[2]}" + ) + proprio = proprio[:, 0, :] # [B, D] + context, context_mask = self._append_proprio_to_context( + context=context, + context_mask=context_mask, + proprio=proprio.to(device=self.device, dtype=self.torch_dtype), + ) + action = action.to(device=self.device, dtype=self.torch_dtype, non_blocking=True) + + if action_is_pad is not None: + action_is_pad = action_is_pad.to(device=self.device, dtype=torch.bool, non_blocking=True) + if image_is_pad is not None: + image_is_pad = image_is_pad.to(device=self.device, dtype=torch.bool, non_blocking=True) + + return { + "context": context, + "context_mask": context_mask, + "input_latents": input_latents, + "first_frame_latents": first_frame_latents, + "fuse_vae_embedding_in_latents": fuse_flag, + "action": action, + "action_is_pad": action_is_pad, + "image_is_pad": image_is_pad, + } + + @torch.no_grad() + def _build_mot_attention_mask( + self, + video_seq_len: int, + action_seq_len: int, + video_tokens_per_frame: int, + device: torch.device, + ) -> torch.Tensor: + total_seq_len = video_seq_len + action_seq_len + mask = torch.zeros((total_seq_len, total_seq_len), dtype=torch.bool, device=device) + + # video -> video + mask[:video_seq_len, :video_seq_len] = self.video_expert.build_video_to_video_mask( + video_seq_len=video_seq_len, + video_tokens_per_frame=video_tokens_per_frame, + device=device, + ) + # action -> action + mask[video_seq_len:, video_seq_len:] = True + # action -> first-frame video only + first_frame_tokens = min(video_tokens_per_frame, video_seq_len) + mask[video_seq_len:, :first_frame_tokens] = True + return mask + + def _compute_video_loss_per_sample( + self, + pred_video: torch.Tensor, + target_video: torch.Tensor, + image_is_pad: torch.Tensor | None, + include_initial_video_step: bool, + ) -> torch.Tensor: + video_loss_token = functional.mse_loss( + pred_video.float(), target_video.float(), reduction="none" + ).mean(dim=(1, 3, 4)) + if image_is_pad is None: + return video_loss_token.mean(dim=1) + + temporal_factor = int(self.vae.temporal_downsample_factor) + if temporal_factor <= 0: + raise ValueError(f"`vae.temporal_downsample_factor` must be positive, got {temporal_factor}.") + if image_is_pad.shape[1] < 1: + raise ValueError("`image_is_pad` must contain at least one frame.") + if (image_is_pad.shape[1] - 1) % temporal_factor != 0: + raise ValueError( + "Cannot align `image_is_pad` with video latent steps: " + f"num_frames={image_is_pad.shape[1]}, temporal_downsample_factor={temporal_factor}." + ) + + tail_is_pad = image_is_pad[:, 1:] + latent_tail_is_pad = tail_is_pad.view(image_is_pad.shape[0], -1, temporal_factor).all(dim=2) + if include_initial_video_step: + video_is_pad = torch.cat([image_is_pad[:, :1], latent_tail_is_pad], dim=1) + else: + video_is_pad = latent_tail_is_pad + + if video_is_pad.shape[1] != video_loss_token.shape[1]: + raise ValueError( + "Video-loss mask shape mismatch: " + f"mask steps={video_is_pad.shape[1]}, loss steps={video_loss_token.shape[1]}." + ) + + valid = (~video_is_pad).to(device=video_loss_token.device, dtype=video_loss_token.dtype) + valid_sum = valid.sum(dim=1).clamp(min=1.0) + return (video_loss_token * valid).sum(dim=1) / valid_sum + + def _sample_training_targets(self, inputs: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]: + input_latents = inputs["input_latents"] + batch_size = input_latents.shape[0] + action = inputs["action"] + noise_video = torch.randn_like(input_latents) + timestep_video = self.train_video_scheduler.sample_training_t( + batch_size=batch_size, + device=self.device, + dtype=input_latents.dtype, + ) + latents = self.train_video_scheduler.add_noise(input_latents, noise_video, timestep_video) + target_video = self.train_video_scheduler.training_target(input_latents, noise_video, timestep_video) + + if inputs["first_frame_latents"] is not None: + latents[:, :, 0:1] = inputs["first_frame_latents"] + noise_action = torch.randn_like(action) + timestep_action = self.train_action_scheduler.sample_training_t( + batch_size=batch_size, + device=self.device, + dtype=action.dtype, + ) + noisy_action = self.train_action_scheduler.add_noise(action, noise_action, timestep_action) + target_action = self.train_action_scheduler.training_target(action, noise_action, timestep_action) + return { + "latents": latents, + "target_video": target_video, + "noisy_action": noisy_action, + "target_action": target_action, + "timestep_video": timestep_video, + "timestep_action": timestep_action, + } + + def _run_training_mot(self, inputs: dict[str, torch.Tensor], targets: dict[str, torch.Tensor]): + video_pre = self.video_expert.pre_dit( + x=targets["latents"], + timestep=targets["timestep_video"], + context=inputs["context"], + context_mask=inputs["context_mask"], + action=inputs["action"], + fuse_vae_embedding_in_latents=inputs["fuse_vae_embedding_in_latents"], + ) + action_pre = self.action_expert.pre_dit( + action_tokens=targets["noisy_action"], + timestep=targets["timestep_action"], + context=inputs["context"], + context_mask=inputs["context_mask"], + ) + video_tokens = video_pre["tokens"] + action_tokens = action_pre["tokens"] + attention_mask = self._build_mot_attention_mask( + video_seq_len=video_tokens.shape[1], + action_seq_len=action_tokens.shape[1], + video_tokens_per_frame=int(video_pre["meta"]["tokens_per_frame"]), + device=video_tokens.device, + ) + tokens_out = self.mot( + embeds_all={ + "video": video_tokens, + "action": action_tokens, + }, + attention_mask=attention_mask, + freqs_all={ + "video": video_pre["freqs"], + "action": action_pre["freqs"], + }, + context_all={ + "video": { + "context": video_pre["context"], + "mask": video_pre["context_mask"], + }, + "action": { + "context": action_pre["context"], + "mask": action_pre["context_mask"], + }, + }, + t_mod_all={ + "video": video_pre["t_mod"], + "action": action_pre["t_mod"], + }, + ) + pred_video = self.video_expert.post_dit(tokens_out["video"], video_pre) + pred_action = self.action_expert.post_dit(tokens_out["action"], action_pre) + return pred_video, pred_action + + def _compute_training_video_loss(self, inputs, pred_video, target_video, timestep_video): + include_initial_video_step = inputs["first_frame_latents"] is None + if inputs["first_frame_latents"] is not None: + pred_video = pred_video[:, :, 1:] + target_video = target_video[:, :, 1:] + loss_video_per_sample = self._compute_video_loss_per_sample( + pred_video=pred_video, + target_video=target_video, + image_is_pad=inputs["image_is_pad"], + include_initial_video_step=include_initial_video_step, + ) + video_weight = self.train_video_scheduler.training_weight(timestep_video).to( + loss_video_per_sample.device, + dtype=loss_video_per_sample.dtype, + ) + return (loss_video_per_sample * video_weight).mean() + + def _compute_training_action_loss(self, inputs, pred_action, target_action, timestep_action): + action_loss_token = functional.mse_loss( + pred_action.float(), target_action.float(), reduction="none" + ).mean(dim=2) + if inputs["action_is_pad"] is not None: + valid = (~inputs["action_is_pad"]).to( + device=action_loss_token.device, + dtype=action_loss_token.dtype, + ) + valid_sum = valid.sum(dim=1).clamp(min=1.0) + action_loss_per_sample = (action_loss_token * valid).sum(dim=1) / valid_sum + else: + action_loss_per_sample = action_loss_token.mean(dim=1) + action_weight = self.train_action_scheduler.training_weight(timestep_action).to( + action_loss_per_sample.device, + dtype=action_loss_per_sample.dtype, + ) + return (action_loss_per_sample * action_weight).mean() + + def training_loss(self, sample, tiled: bool = False): + inputs = self.build_inputs(sample, tiled=tiled) + targets = self._sample_training_targets(inputs) + pred_video, pred_action = self._run_training_mot(inputs=inputs, targets=targets) + loss_video = self._compute_training_video_loss( + inputs=inputs, + pred_video=pred_video, + target_video=targets["target_video"], + timestep_video=targets["timestep_video"], + ) + loss_action = self._compute_training_action_loss( + inputs=inputs, + pred_action=pred_action, + target_action=targets["target_action"], + timestep_action=targets["timestep_action"], + ) + loss_total = self.loss_lambda_video * loss_video + self.loss_lambda_action * loss_action + loss_dict = { + "loss_video": self.loss_lambda_video * float(loss_video.detach().item()), + "loss_action": self.loss_lambda_action * float(loss_action.detach().item()), + } + return loss_total, loss_dict + + @torch.no_grad() + def _predict_joint_noise( + self, + latents_video: torch.Tensor, + latents_action: torch.Tensor, + timestep_video: torch.Tensor, + timestep_action: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor, + fuse_vae_embedding_in_latents: bool, + gt_action: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + video_pre = self.video_expert.pre_dit( + x=latents_video, + timestep=timestep_video, + context=context, + context_mask=context_mask, + action=gt_action, + fuse_vae_embedding_in_latents=fuse_vae_embedding_in_latents, + ) + action_pre = self.action_expert.pre_dit( + action_tokens=latents_action, + timestep=timestep_action, + context=context, + context_mask=context_mask, + ) + + attention_mask = self._build_mot_attention_mask( + video_seq_len=video_pre["tokens"].shape[1], + action_seq_len=action_pre["tokens"].shape[1], + video_tokens_per_frame=int(video_pre["meta"]["tokens_per_frame"]), + device=video_pre["tokens"].device, + ) + + tokens_out = self.mot( + embeds_all={ + "video": video_pre["tokens"], + "action": action_pre["tokens"], + }, + attention_mask=attention_mask, + freqs_all={ + "video": video_pre["freqs"], + "action": action_pre["freqs"], + }, + context_all={ + "video": { + "context": video_pre["context"], + "mask": video_pre["context_mask"], + }, + "action": { + "context": action_pre["context"], + "mask": action_pre["context_mask"], + }, + }, + t_mod_all={ + "video": video_pre["t_mod"], + "action": action_pre["t_mod"], + }, + ) + + pred_video = self.video_expert.post_dit(tokens_out["video"], video_pre) + pred_action = self.action_expert.post_dit(tokens_out["action"], action_pre) + return pred_video, pred_action + + @torch.no_grad() + def _predict_action_noise( + self, + first_frame_latents: torch.Tensor, + latents_action: torch.Tensor, + timestep_action: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor, + fuse_vae_embedding_in_latents: bool, + ) -> torch.Tensor: + timestep_video = torch.zeros_like( + timestep_action, dtype=first_frame_latents.dtype, device=self.device + ) + video_pre = self.video_expert.pre_dit( + x=first_frame_latents, + timestep=timestep_video, + context=context, + context_mask=context_mask, + action=None, + fuse_vae_embedding_in_latents=fuse_vae_embedding_in_latents, + ) + action_pre = self.action_expert.pre_dit( + action_tokens=latents_action, + timestep=timestep_action, + context=context, + context_mask=context_mask, + ) + + attention_mask = self._build_mot_attention_mask( + video_seq_len=video_pre["tokens"].shape[1], + action_seq_len=action_pre["tokens"].shape[1], + video_tokens_per_frame=int(video_pre["meta"]["tokens_per_frame"]), + device=video_pre["tokens"].device, + ) + tokens_out = self.mot( + embeds_all={ + "video": video_pre["tokens"], + "action": action_pre["tokens"], + }, + attention_mask=attention_mask, + freqs_all={ + "video": video_pre["freqs"], + "action": action_pre["freqs"], + }, + context_all={ + "video": { + "context": video_pre["context"], + "mask": video_pre["context_mask"], + }, + "action": { + "context": action_pre["context"], + "mask": action_pre["context_mask"], + }, + }, + t_mod_all={ + "video": video_pre["t_mod"], + "action": action_pre["t_mod"], + }, + ) + pred_action = self.action_expert.post_dit(tokens_out["action"], action_pre) + return pred_action + + @torch.no_grad() + def _predict_action_noise_with_cache( + self, + latents_action: torch.Tensor, + timestep_action: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor, + video_kv_cache: list[dict[str, torch.Tensor]], + attention_mask: torch.Tensor, + video_seq_len: int, + ) -> torch.Tensor: + action_pre = self.action_expert.pre_dit( + action_tokens=latents_action, + timestep=timestep_action, + context=context, + context_mask=context_mask, + ) + action_tokens = self.mot.forward_action_with_video_cache( + action_tokens=action_pre["tokens"], + action_freqs=action_pre["freqs"], + action_t_mod=action_pre["t_mod"], + action_context_payload={ + "context": action_pre["context"], + "mask": action_pre["context_mask"], + }, + video_kv_cache=video_kv_cache, + attention_mask=attention_mask, + video_seq_len=video_seq_len, + ) + return self.action_expert.post_dit(action_tokens, action_pre) + + def _normalize_infer_input_image( + self, + input_image: torch.Tensor, + num_video_frames: int | None = None, + ) -> tuple[torch.Tensor, int, int]: + if input_image.ndim == 3: + input_image = input_image.unsqueeze(0) + if input_image.ndim != 4 or input_image.shape[0] != 1 or input_image.shape[1] != 3: + raise ValueError( + f"`input_image` must have shape [1,3,H,W] or [3,H,W], got {tuple(input_image.shape)}" + ) + _, _, height, width = input_image.shape + if height % 16 != 0 or width % 16 != 0: + raise ValueError( + f"`input_image` must be resized before infer, expected multiples of 16 but got HxW=({height},{width})" + ) + if num_video_frames is not None: + checked_h, checked_w, checked_t = self._check_resize_height_width(height, width, num_video_frames) + if (checked_h, checked_w) != (height, width): + raise ValueError( + f"`input_image` must be resized before infer, expected multiples of 16 but got HxW=({height},{width})" + ) + if checked_t != num_video_frames: + raise ValueError(f"`num_video_frames` must satisfy T % 4 == 1, got {num_video_frames}") + return input_image, height, width + + def _normalize_infer_proprio(self, proprio: torch.Tensor | None) -> torch.Tensor | None: + if proprio is None: + return None + if self.proprio_dim is None: + raise ValueError( + "`proprio` was provided but `proprio_dim=None` so `proprio_encoder` is disabled." + ) + if proprio.ndim == 1: + proprio = proprio.unsqueeze(0) + elif proprio.ndim == 2 and proprio.shape[0] == 1: + pass + else: + raise ValueError(f"`proprio` must be [D] or [1,D], got shape {tuple(proprio.shape)}") + if proprio.shape[1] != self.proprio_dim: + raise ValueError(f"`proprio` last dim must be {self.proprio_dim}, got {proprio.shape[1]}") + return proprio.to(device=self.device, dtype=self.torch_dtype) + + def _prepare_infer_context(self, prompt, context, context_mask, proprio): + use_prompt = prompt is not None + use_context = context is not None or context_mask is not None + if use_prompt and use_context: + raise ValueError("`prompt` and `context/context_mask` are mutually exclusive.") + if not use_prompt and not use_context: + raise ValueError("Either `prompt` or both `context/context_mask` must be provided.") + if use_prompt: + context, context_mask = self.encode_prompt(prompt) + else: + context, context_mask = self._normalize_context_tensors(context, context_mask) + if proprio is not None: + context, context_mask = self._append_proprio_to_context( + context=context, + context_mask=context_mask, + proprio=proprio, + ) + return context, context_mask + + def _normalize_context_tensors(self, context, context_mask): + if context is None or context_mask is None: + raise ValueError("`context` and `context_mask` must be both provided together.") + if context.ndim == 2: + context = context.unsqueeze(0) + if context_mask.ndim == 1: + context_mask = context_mask.unsqueeze(0) + if context.ndim != 3 or context_mask.ndim != 2: + raise ValueError( + f"`context/context_mask` must be [B,L,D]/[B,L], got {tuple(context.shape)} and {tuple(context_mask.shape)}" + ) + context = context.to(device=self.device, dtype=self.torch_dtype, non_blocking=True) + context_mask = context_mask.to(device=self.device, dtype=torch.bool, non_blocking=True) + return context, context_mask + + def _make_action_latents(self, action_horizon: int, seed: int | None, rand_device: str): + generator = None if seed is None else torch.Generator(device=rand_device).manual_seed(seed) + return torch.randn( + (1, action_horizon, self.action_expert.action_dim), + generator=generator, + device=rand_device, + dtype=torch.float32, + ).to(device=self.device, dtype=self.torch_dtype) + + def _make_video_latents(self, num_video_frames: int, height: int, width: int, seed, rand_device): + latent_t = (num_video_frames - 1) // self.vae.temporal_downsample_factor + 1 + latent_h = height // self.vae.upsampling_factor + latent_w = width // self.vae.upsampling_factor + generator = None if seed is None else torch.Generator(device=rand_device).manual_seed(seed) + return torch.randn( + (1, self.vae.z_dim, latent_t, latent_h, latent_w), + generator=generator, + device=rand_device, + dtype=torch.float32, + ).to(device=self.device, dtype=self.torch_dtype) + + @torch.no_grad() + def infer_joint( + self, + prompt: str | None, + input_image: torch.Tensor, + num_video_frames: int, + action_horizon: int, + action: torch.Tensor + | None = None, # NOTE: this is gt action for conditioning videos, not for action expert + proprio: torch.Tensor | None = None, + context: torch.Tensor | None = None, + context_mask: torch.Tensor | None = None, + negative_prompt: str | None = None, + text_cfg_scale: float = 1.0, + num_inference_steps: int = 20, + sigma_shift: float | None = None, + seed: int | None = None, + rand_device: str = "cpu", + tiled: bool = False, + test_action_with_infer_action: bool = True, + ) -> dict[str, Any]: + self.eval() + if test_action_with_infer_action: + if seed is None: + raise ValueError("`test_action_with_infer_action=True` requires non-null `seed`.") + action_only_out = self.infer_action( + prompt=prompt, + input_image=input_image.clone(), + action_horizon=action_horizon, + context=context.clone() if context is not None else None, + context_mask=context_mask.clone() if context_mask is not None else None, + num_inference_steps=num_inference_steps, + sigma_shift=sigma_shift, + seed=seed, + rand_device=rand_device, + tiled=tiled, + proprio=proprio.clone() if proprio is not None else None, + )["action"] + + input_image, height, width = self._normalize_infer_input_image(input_image, num_video_frames) + if action is not None: + if action.ndim == 2: + action = action.unsqueeze(0) + if action.ndim != 3 or action.shape[0] != 1 or action.shape[1] != action_horizon: + # NOTE: This enforces action condition to have the same shape as action horizon to predict, which may be unnecessary + raise ValueError( + f"`action` must have shape [1, T, a_dim] or [T, a_dim], got {tuple(action.shape)} with action_horizon={action_horizon}" + ) + action = action.to(device=self.device, dtype=self.torch_dtype) + proprio = self._normalize_infer_proprio(proprio) + latents_video = self._make_video_latents(num_video_frames, height, width, seed, rand_device) + latents_action = self._make_action_latents(action_horizon, seed, rand_device) + + input_image = input_image.to(device=self.device, dtype=self.torch_dtype) + first_frame_latents = self._encode_input_image_latents_tensor(input_image=input_image, tiled=tiled) + latents_video[:, :, 0:1] = first_frame_latents.clone() + fuse_flag = bool(getattr(self.video_expert, "fuse_vae_embedding_in_latents", False)) + context, context_mask = self._prepare_infer_context(prompt, context, context_mask, proprio) + + infer_timesteps_video, infer_deltas_video = self.infer_video_scheduler.build_inference_schedule( + num_inference_steps=num_inference_steps, + device=self.device, + dtype=latents_video.dtype, + shift_override=sigma_shift, + ) + infer_timesteps_action, infer_deltas_action = self.infer_action_scheduler.build_inference_schedule( + num_inference_steps=num_inference_steps, + device=self.device, + dtype=latents_action.dtype, + shift_override=sigma_shift, + ) + for step_t_video, step_delta_video, step_t_action, step_delta_action in zip( + infer_timesteps_video, + infer_deltas_video, + infer_timesteps_action, + infer_deltas_action, + strict=True, + ): + timestep_video = step_t_video.unsqueeze(0).to(dtype=latents_video.dtype, device=self.device) + timestep_action = step_t_action.unsqueeze(0).to(dtype=latents_action.dtype, device=self.device) + + pred_video, pred_action = self._predict_joint_noise( + latents_video=latents_video, + latents_action=latents_action, + timestep_video=timestep_video, + timestep_action=timestep_action, + context=context, + context_mask=context_mask, + fuse_vae_embedding_in_latents=fuse_flag, + gt_action=action, + ) + + latents_video = self.infer_video_scheduler.step(pred_video, step_delta_video, latents_video) + latents_action = self.infer_action_scheduler.step(pred_action, step_delta_action, latents_action) + latents_video[:, :, 0:1] = first_frame_latents.clone() + + action_out = latents_action[0].detach().to(device="cpu", dtype=torch.float32) + if test_action_with_infer_action and not torch.allclose( + action_out, action_only_out, atol=1e-2, rtol=1e-2 + ): + max_abs_diff = (action_out - action_only_out).abs().max().item() + logger.warning( + f"Action from infer_joint and infer_action differ with max abs diff {max_abs_diff:.6f}. " + ) + + return { + "video": self._decode_latents(latents_video, tiled=tiled), + "action": action_out, + } + + @torch.no_grad() + def infer_action( + self, + prompt: str | None, + input_image: torch.Tensor, + action_horizon: int, + proprio: torch.Tensor | None = None, + context: torch.Tensor | None = None, + context_mask: torch.Tensor | None = None, + negative_prompt: str | None = None, + text_cfg_scale: float = 1.0, + num_inference_steps: int = 20, + sigma_shift: float | None = None, + seed: int | None = None, + rand_device: str = "cpu", + tiled: bool = False, + ) -> dict[str, Any]: + self.eval() + if str(getattr(self.video_expert, "video_attention_mask_mode", "")) != "first_frame_causal": + raise ValueError("`infer_action` requires `video_attention_mask_mode='first_frame_causal'`.") + + input_image, _, _ = self._normalize_infer_input_image(input_image) + proprio = self._normalize_infer_proprio(proprio) + latents_action = self._make_action_latents(action_horizon, seed, rand_device) + + input_image = input_image.to(device=self.device, dtype=self.torch_dtype) + first_frame_latents = self._encode_input_image_latents_tensor(input_image=input_image, tiled=tiled) + fuse_flag = bool(getattr(self.video_expert, "fuse_vae_embedding_in_latents", False)) + + context, context_mask = self._prepare_infer_context(prompt, context, context_mask, proprio) + + timestep_video = torch.zeros( + (first_frame_latents.shape[0],), + dtype=first_frame_latents.dtype, + device=self.device, + ) + video_pre = self.video_expert.pre_dit( + x=first_frame_latents, + timestep=timestep_video, + context=context, + context_mask=context_mask, + action=None, + fuse_vae_embedding_in_latents=fuse_flag, + ) + video_seq_len = int(video_pre["tokens"].shape[1]) + attention_mask = self._build_mot_attention_mask( + video_seq_len=video_seq_len, + action_seq_len=latents_action.shape[1], + video_tokens_per_frame=int(video_pre["meta"]["tokens_per_frame"]), + device=video_pre["tokens"].device, + ) + video_kv_cache = self.mot.prefill_video_cache( + video_tokens=video_pre["tokens"], + video_freqs=video_pre["freqs"], + video_t_mod=video_pre["t_mod"], + video_context_payload={ + "context": video_pre["context"], + "mask": video_pre["context_mask"], + }, + video_attention_mask=attention_mask[:video_seq_len, :video_seq_len], + ) + + infer_timesteps_action, infer_deltas_action = self.infer_action_scheduler.build_inference_schedule( + num_inference_steps=num_inference_steps, + device=self.device, + dtype=latents_action.dtype, + shift_override=sigma_shift, + ) + for step_t_action, step_delta_action in zip(infer_timesteps_action, infer_deltas_action, strict=True): + timestep_action = step_t_action.unsqueeze(0).to(dtype=latents_action.dtype, device=self.device) + + pred_action = self._predict_action_noise_with_cache( + latents_action=latents_action, + timestep_action=timestep_action, + context=context, + context_mask=context_mask, + video_kv_cache=video_kv_cache, + attention_mask=attention_mask, + video_seq_len=video_seq_len, + ) + + latents_action = self.infer_action_scheduler.step(pred_action, step_delta_action, latents_action) + + return { + "action": latents_action[0].detach().to(device="cpu", dtype=torch.float32), + } + + @torch.no_grad() + def infer( + self, + prompt: str | None, + input_image: torch.Tensor, + num_frames: int, + action: torch.Tensor | None = None, + action_horizon: int | None = None, + proprio: torch.Tensor | None = None, + context: torch.Tensor | None = None, + context_mask: torch.Tensor | None = None, + negative_prompt: str | None = None, + text_cfg_scale: float = 5.0, + action_cfg_scale: float = 1.0, + num_inference_steps: int = 20, + sigma_shift: float | None = None, + seed: int | None = None, + rand_device: str = "cpu", + tiled: bool = False, + ): + return self.infer_joint( + prompt=prompt, + input_image=input_image, + num_video_frames=num_frames, + action_horizon=action_horizon, + action=action, + proprio=proprio, + context=context, + context_mask=context_mask, + negative_prompt=negative_prompt, + text_cfg_scale=text_cfg_scale, + num_inference_steps=num_inference_steps, + sigma_shift=sigma_shift, + seed=seed, + rand_device=rand_device, + tiled=tiled, + ) + + def forward(self, *args, **kwargs): + return self.training_loss(*args, **kwargs) diff --git a/src/lerobot/policies/fastwam/wan/video_dit.py b/src/lerobot/policies/fastwam/wan/video_dit.py new file mode 100644 index 000000000..a98f06cde --- /dev/null +++ b/src/lerobot/policies/fastwam/wan/video_dit.py @@ -0,0 +1,800 @@ +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import logging +from typing import Any + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as functional +from einops import rearrange + +from .model import ( + WanAttentionBlock, + WanLayerNorm, + WanModel, + WanRMSNorm, + rope_apply, + rope_params, + sinusoidal_embedding_1d, +) + +logger = logging.getLogger(__name__) + + +def get_sampling_sigmas(sampling_steps, shift): + # Vendored from Wan2.2 (formerly wan/utils/fm_solvers.py); computes the + # noise-level (sigma) schedule for Wan-compatible flow-matching inference. + sigma = np.linspace(1, 0, sampling_steps + 1)[:sampling_steps] + sigma = shift * sigma / (1 + (shift - 1) * sigma) + return sigma + + +def create_custom_forward(module): + def custom_forward(*inputs, **kwargs): + return module(*inputs, **kwargs) + + return custom_forward + + +def gradient_checkpoint_forward( + model, + use_gradient_checkpointing, + *args, + **kwargs, +): + if use_gradient_checkpointing: + model_output = torch.utils.checkpoint.checkpoint( + create_custom_forward(model), + *args, + **kwargs, + use_reentrant=False, + ) + else: + model_output = model(*args, **kwargs) + return model_output + + +def fastwam_masked_attention( + q: torch.Tensor, + k: torch.Tensor, + v: torch.Tensor, + num_heads: int, + ctx_mask: torch.Tensor | None = None, + fp32_attention: bool = True, +) -> torch.Tensor: + """FastWAM masked attention wrapper for MoT masks and CPU test coverage. + + The official Wan attention implementation is still used as the source of + the projection/norm modules. This wrapper only replaces the final attention + kernel because FastWAM needs explicit boolean masks for video/action MoT + routing, while the upstream FlashAttention path accepts sequence lengths + but not arbitrary [query, key] masks. + """ + + q = rearrange(q, "b s (n d) -> b n s d", n=num_heads) + k = rearrange(k, "b s (n d) -> b n s d", n=num_heads) + v = rearrange(v, "b s (n d) -> b n s d", n=num_heads) + if fp32_attention: + q = q.float() + k = k.float() + v = v.float() + else: + q = q.to(dtype=v.dtype) + k = k.to(dtype=v.dtype) + x = functional.scaled_dot_product_attention(q, k, v, attn_mask=ctx_mask) + return rearrange(x, "b n s d -> b s (n d)", n=num_heads) + + +def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor): + return x * (1 + scale) + shift + + +class WanContinuousFlowMatchScheduler: + """Continuous-time Flow-Matching scheduler with shift-based Wan sampling.""" + + def __init__(self, num_train_timesteps: int = 1000, shift: float = 5.0, eps: float = 1e-10): + if num_train_timesteps <= 0: + raise ValueError(f"`num_train_timesteps` must be positive, got {num_train_timesteps}") + if shift <= 0: + raise ValueError(f"`shift` must be positive, got {shift}") + self.num_train_timesteps = int(num_train_timesteps) + self.shift = float(shift) + self.eps = float(eps) + self._y_min, self._weight_norm_const = self._precompute_training_weight_stats() + + @staticmethod + def _phi(u: torch.Tensor, shift: float) -> torch.Tensor: + return shift * u / (1.0 + (shift - 1.0) * u) + + def _precompute_training_weight_stats(self) -> tuple[float, float]: + steps = self.num_train_timesteps + u_grid = torch.linspace(1.0, 0.0, steps + 1, dtype=torch.float64)[:-1] + t_grid = self._phi(u_grid, self.shift) * float(steps) + y_grid = torch.exp(-2.0 * ((t_grid - (steps / 2.0)) / steps) ** 2) + y_min = float(y_grid.min().item()) + y_shifted_grid = y_grid - y_min + norm_const = float(y_shifted_grid.mean().item()) + return y_min, norm_const + + def sample_training_t(self, batch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor: + if batch_size <= 0: + raise ValueError(f"`batch_size` must be positive, got {batch_size}") + u = torch.rand((batch_size,), device=device, dtype=torch.float32) + sigma = self._phi(u, self.shift) + timestep = sigma * float(self.num_train_timesteps) + return timestep.to(dtype=dtype) + + def training_weight(self, timestep: torch.Tensor) -> torch.Tensor: + t = timestep.to(dtype=torch.float32) + steps = float(self.num_train_timesteps) + y = torch.exp(-2.0 * ((t - (steps / 2.0)) / steps) ** 2) + y_shifted = y - self._y_min + weight = y_shifted / (self._weight_norm_const + self.eps) + if weight.numel() == 1: + return weight.reshape(()) + return weight + + def add_noise( + self, original_samples: torch.Tensor, noise: torch.Tensor, timestep: torch.Tensor + ) -> torch.Tensor: + sigma = (timestep / float(self.num_train_timesteps)).to( + original_samples.device, dtype=original_samples.dtype + ) + if sigma.ndim == 0: + return (1 - sigma) * original_samples + sigma * noise + sigma = sigma.view(-1, *([1] * (original_samples.ndim - 1))) + return (1 - sigma) * original_samples + sigma * noise + + @staticmethod + def training_target(sample: torch.Tensor, noise: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor: + del timestep + return noise - sample + + def build_inference_schedule( + self, + num_inference_steps: int, + device: torch.device, + dtype: torch.dtype, + shift_override: float | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be positive, got {num_inference_steps}") + shift = self.shift if shift_override is None else float(shift_override) + if shift <= 0: + raise ValueError(f"`shift` must be positive, got {shift}") + + sigma_steps = torch.as_tensor( + get_sampling_sigmas(num_inference_steps, shift), + device=device, + dtype=torch.float32, + ) + timesteps = sigma_steps * float(self.num_train_timesteps) + sigma_next = torch.cat([sigma_steps[1:], sigma_steps.new_zeros(1)]) + deltas = sigma_next - sigma_steps + return timesteps.to(dtype=dtype), deltas.to(dtype=dtype) + + @staticmethod + def step(model_output: torch.Tensor, delta: torch.Tensor, sample: torch.Tensor) -> torch.Tensor: + delta = delta.to(sample.device, dtype=sample.dtype) + if delta.ndim == 0: + return sample + model_output * delta + delta = delta.view(-1, *([1] * (sample.ndim - 1))) + return sample + model_output * delta + + +def precompute_freqs_cis(dim: int, end: int = 1024, theta: float = 10000.0): + return rope_params(end, dim, theta) + + +def apply_dense_rope(x: torch.Tensor, freqs: torch.Tensor, num_heads: int) -> torch.Tensor: + x = rearrange(x, "b s (n d) -> b s n d", n=num_heads) + x_out = torch.view_as_complex(x.to(torch.float32).reshape(x.shape[0], x.shape[1], x.shape[2], -1, 2)) + freqs = freqs.to(torch.complex64) if freqs.device.type == "npu" else freqs + x_out = torch.view_as_real(x_out * freqs).flatten(2) + return x_out.to(x.dtype) + + +def _linear_input(linear: nn.Linear, x: torch.Tensor) -> torch.Tensor: + return x.to(dtype=linear.weight.dtype) + + +def _wan_layer_norm(norm: nn.Module, x: torch.Tensor) -> torch.Tensor: + if isinstance(norm, WanLayerNorm) and norm.weight is not None: + weight = norm.weight.float() + bias = norm.bias.float() if norm.bias is not None else None + return functional.layer_norm(x.float(), norm.normalized_shape, weight, bias, norm.eps).to( + dtype=x.dtype + ) + return norm(x) + + +def create_group_causal_attn_mask( + num_temporal_groups: int, num_query_per_group: int, num_key_per_group: int, mode: str = "causal" +) -> torch.Tensor: + if mode not in ["causal", "group_diagonal"]: + raise ValueError(f"`mode` must be 'causal' or 'group_diagonal', got {mode}.") + if num_temporal_groups <= 0: + raise ValueError(f"`num_temporal_groups` must be positive, got {num_temporal_groups}.") + if num_query_per_group <= 0: + raise ValueError(f"`num_query_per_group` must be positive, got {num_query_per_group}.") + if num_key_per_group <= 0: + raise ValueError(f"`num_key_per_group` must be positive, got {num_key_per_group}.") + + total_num_query_tokens = num_temporal_groups * num_query_per_group + total_num_key_tokens = num_temporal_groups * num_key_per_group + query_time_indices = torch.arange(num_temporal_groups).repeat_interleave(num_query_per_group).unsqueeze(1) + key_time_indices = torch.arange(num_temporal_groups).repeat_interleave(num_key_per_group).unsqueeze(0) + + if mode == "causal": + attn_mask = query_time_indices >= key_time_indices + else: + attn_mask = query_time_indices == key_time_indices + + if attn_mask.shape != (total_num_query_tokens, total_num_key_tokens): + raise RuntimeError("Attention mask shape mismatch.") + return attn_mask + + +class FastWAMAttentionBlock(WanAttentionBlock): + """Wan attention block with FastWAM's arbitrary boolean mask support.""" + + def __init__( + self, + hidden_dim: int, + attn_head_dim: int, + num_heads: int, + ffn_dim: int, + eps: float = 1e-6, + fp32_attention: bool = True, + ): + attention_dim = attn_head_dim * num_heads + if hidden_dim == attention_dim: + super().__init__( + dim=hidden_dim, + ffn_dim=ffn_dim, + num_heads=num_heads, + qk_norm=True, + cross_attn_norm=True, + eps=eps, + ) + else: + nn.Module.__init__(self) + self.dim = hidden_dim + self.ffn_dim = ffn_dim + self.num_heads = num_heads + self.qk_norm = True + self.cross_attn_norm = True + self.eps = eps + self.norm1 = WanLayerNorm(hidden_dim, eps) + self.self_attn = _FastWAMProjectedAttention(hidden_dim, attention_dim, num_heads, eps) + self.norm3 = WanLayerNorm(hidden_dim, eps, elementwise_affine=True) + self.cross_attn = _FastWAMProjectedAttention(hidden_dim, attention_dim, num_heads, eps) + self.norm2 = WanLayerNorm(hidden_dim, eps) + self.ffn = nn.Sequential( + nn.Linear(hidden_dim, ffn_dim), + nn.GELU(approximate="tanh"), + nn.Linear(ffn_dim, hidden_dim), + ) + self.modulation = nn.Parameter(torch.randn(1, 6, hidden_dim) / hidden_dim**0.5) + self.attn_head_dim = attn_head_dim + self.fp32_attention = bool(fp32_attention) + + @staticmethod + def split_modulation(block, t_mod: torch.Tensor): + has_seq = len(t_mod.shape) == 4 + chunk_dim = 2 if has_seq else 1 + + base_mod = block.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (base_mod + t_mod).chunk( + 6, dim=chunk_dim + ) + if has_seq: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + shift_msa.squeeze(2), + scale_msa.squeeze(2), + gate_msa.squeeze(2), + shift_mlp.squeeze(2), + scale_mlp.squeeze(2), + gate_mlp.squeeze(2), + ) + return shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp + + def project_self_attention( + self, x: torch.Tensor, freqs: torch.Tensor | dict[str, torch.Tensor] + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + q = self.self_attn.norm_q(self.self_attn.q(x)) + k = self.self_attn.norm_k(self.self_attn.k(x)) + v = self.self_attn.v(x) + if isinstance(freqs, dict): + b, s = x.shape[:2] + q = rope_apply( + q.view(b, s, self.num_heads, self.attn_head_dim), + freqs["grid_sizes"], + freqs["freqs"], + ).flatten(2) + k = rope_apply( + k.view(b, s, self.num_heads, self.attn_head_dim), + freqs["grid_sizes"], + freqs["freqs"], + ).flatten(2) + else: + q = apply_dense_rope(q, freqs, self.num_heads) + k = apply_dense_rope(k, freqs, self.num_heads) + return q, k, v + + def apply_cross_attention( + self, x: torch.Tensor, context: torch.Tensor, context_mask: torch.Tensor | None = None + ) -> torch.Tensor: + if context_mask is not None and context_mask.dim() == 3: + context_mask = context_mask.unsqueeze(1) + attn = self.cross_attn + b, n, d = x.size(0), attn.num_heads, attn.head_dim + q = attn.norm_q(attn.q(x)).view(b, -1, n * d) + k = attn.norm_k(attn.k(context)).view(b, -1, n * d) + v = attn.v(context).view(b, -1, n * d) + x = fastwam_masked_attention( + q=q, + k=k, + v=v, + num_heads=n, + ctx_mask=context_mask, + fp32_attention=self.fp32_attention, + ) + return attn.o(_linear_input(attn.o, x)) + + def project_self_attention_output(self, x: torch.Tensor) -> torch.Tensor: + return self.self_attn.o(_linear_input(self.self_attn.o, x)) + + def apply_norm1(self, x: torch.Tensor) -> torch.Tensor: + return _wan_layer_norm(self.norm1, x) + + def apply_norm2(self, x: torch.Tensor) -> torch.Tensor: + return _wan_layer_norm(self.norm2, x) + + def apply_norm3(self, x: torch.Tensor) -> torch.Tensor: + return _wan_layer_norm(self.norm3, x) + + def forward( + self, + x: torch.Tensor, + context: torch.Tensor, + t_mod: torch.Tensor, + freqs: torch.Tensor, + context_mask: torch.Tensor | None = None, + self_attn_mask: torch.Tensor | None = None, + ) -> torch.Tensor: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.split_modulation(self, t_mod) + residual_x = x + attn_input = modulate(self.apply_norm1(x), shift_msa, scale_msa) + q, k, v = self.project_self_attention(attn_input, freqs) + y = fastwam_masked_attention( + q=q, + k=k, + v=v, + num_heads=self.num_heads, + ctx_mask=self_attn_mask, + fp32_attention=self.fp32_attention, + ) + x = residual_x + gate_msa * self.project_self_attention_output(y) + x = x + self.apply_cross_attention(self.apply_norm3(x), context, context_mask=context_mask) + mlp_input = modulate(self.apply_norm2(x), shift_mlp, scale_mlp) + return x + gate_mlp * self.ffn(mlp_input) + + +class _FastWAMProjectedAttention(nn.Module): + def __init__(self, hidden_dim: int, attention_dim: int, num_heads: int, eps: float): + super().__init__() + self.dim = hidden_dim + self.num_heads = num_heads + self.head_dim = attention_dim // num_heads + self.q = nn.Linear(hidden_dim, attention_dim) + self.k = nn.Linear(hidden_dim, attention_dim) + self.v = nn.Linear(hidden_dim, attention_dim) + self.o = nn.Linear(attention_dim, hidden_dim) + self.norm_q = WanRMSNorm(attention_dim, eps=eps) + self.norm_k = WanRMSNorm(attention_dim, eps=eps) + + +class WanVideoDiT(WanModel): + def __init__( + self, + hidden_dim: int, + in_dim: int, + ffn_dim: int, + out_dim: int, + text_dim: int, + freq_dim: int, + eps: float, + patch_size: tuple[int, int, int], + num_heads: int, + attn_head_dim: int, + num_layers: int, + has_image_input: bool = False, + has_image_pos_emb: bool = False, + has_ref_conv: bool = False, + add_control_adapter: bool = False, + in_dim_control_adapter: int = 24, + seperated_timestep: bool = False, + require_vae_embedding: bool = False, + require_clip_embedding: bool = False, + fuse_vae_embedding_in_latents: bool = True, + action_conditioned: bool = False, + action_dim: int = 7, + action_group_causal_mask_mode="causal", + video_attention_mask_mode: str = "bidirectional", + use_gradient_checkpointing: bool = False, + fp32_attention: bool = True, + ): + del in_dim_control_adapter + if has_image_input: + raise ValueError("FastWAM currently expects Wan2.2 TI2V latents with fused image conditioning.") + if has_image_pos_emb: + raise ValueError("FastWAM does not support extra image positional embeddings in WanVideoDiT.") + if has_ref_conv: + raise ValueError("FastWAM does not support reference convolutions in WanVideoDiT.") + if add_control_adapter: + raise ValueError("FastWAM does not support control adapters in WanVideoDiT.") + if require_clip_embedding: + raise ValueError("FastWAM does not support CLIP embedding conditioning in WanVideoDiT.") + if require_vae_embedding or not fuse_vae_embedding_in_latents: + raise ValueError("FastWAM expects VAE conditioning to be fused in latents.") + if attn_head_dim != hidden_dim // num_heads: + raise ValueError( + "`attn_head_dim` must match the upstream Wan head dimension `hidden_dim // num_heads`; " + f"got {attn_head_dim} vs {hidden_dim // num_heads}." + ) + + super().__init__( + model_type="ti2v", + patch_size=patch_size, + text_len=512, + in_dim=in_dim, + dim=hidden_dim, + ffn_dim=ffn_dim, + freq_dim=freq_dim, + text_dim=text_dim, + out_dim=out_dim, + num_heads=num_heads, + num_layers=num_layers, + qk_norm=True, + cross_attn_norm=True, + eps=eps, + ) + self.blocks = torch.nn.ModuleList( + [ + FastWAMAttentionBlock( + hidden_dim=hidden_dim, + attn_head_dim=attn_head_dim, + num_heads=num_heads, + ffn_dim=ffn_dim, + eps=eps, + fp32_attention=fp32_attention, + ) + for _ in range(num_layers) + ] + ) + self.init_weights() + + self.hidden_dim = hidden_dim + self.attn_head_dim = attn_head_dim + self.seperated_timestep = seperated_timestep + self.fuse_vae_embedding_in_latents = fuse_vae_embedding_in_latents + self.video_attention_mask_mode = str(video_attention_mask_mode) + self.action_conditioned = action_conditioned + self.action_dim = action_dim + self.fp32_attention = bool(fp32_attention) + + if self.action_conditioned: + self.action_embedding = torch.nn.Linear(action_dim, hidden_dim) + self.action_group_causal_mask_mode = action_group_causal_mask_mode + + self.use_gradient_checkpointing = use_gradient_checkpointing + if self.use_gradient_checkpointing: + logger.info( + "Using gradient checkpointing for DiT blocks. This will save memory but use more computation." + ) + + def patchify(self, x: torch.Tensor): + return self.patch_embedding(x) + + def _validate_forward_inputs( + self, + x: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor | None, + action: torch.Tensor | None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + if x.ndim != 5: + raise ValueError(f"`latents` must be 5D [B, C, T, H, W], got shape {tuple(x.shape)}") + num_latent_frames = x.shape[2] + if context.ndim != 3: + raise ValueError(f"`context` must be 3D [B, L, D], got shape {tuple(context.shape)}") + if timestep.ndim != 1: + raise ValueError(f"`timestep` must be 1D [B] or [1], got shape {tuple(timestep.shape)}") + if self.action_conditioned: + allow_text_only_single_frame = num_latent_frames == 1 and action is None + if not allow_text_only_single_frame: + if action is None: + raise ValueError("Action input is required for action-conditioned model.") + if action.ndim != 3: + raise ValueError( + f"`action` must be 3D [B, action_horizon, action_dim], got shape {tuple(action.shape)}" + ) + if action.shape[2] != self.action_dim: + raise ValueError( + f"`action` last dimension must be {self.action_dim}, got {action.shape[2]}" + ) + if num_latent_frames <= 1: + raise ValueError( + f"video length must be > 1 for action-conditioned model, got {num_latent_frames}" + ) + if action.shape[1] % (num_latent_frames - 1) != 0: + raise ValueError( + "action horizon must be divisible by (num_latent_frames - 1), " + f"got action_horizon={action.shape[1]}" + ) + if context_mask is None: + context_mask = torch.ones( + (context.shape[0], context.shape[1]), dtype=torch.bool, device=context.device + ) + else: + if context_mask.ndim != 2: + raise ValueError(f"`context_mask` must be 2D [B, L], got shape {tuple(context_mask.shape)}") + if context_mask.shape[0] != context.shape[0] or context_mask.shape[1] != context.shape[1]: + raise ValueError( + "`context_mask` shape must match `context` shape [B, L], " + f"got {tuple(context_mask.shape)} vs {tuple(context.shape)}" + ) + + batch_size = x.shape[0] + if batch_size != context.shape[0]: + if not self.training and batch_size == 1: + x = x.expand(context.shape[0], -1, -1, -1, -1) + batch_size = context.shape[0] + else: + raise ValueError( + f"Batch mismatch between latents and context: {batch_size} vs {context.shape[0]}." + ) + + if timestep.shape[0] not in (1, batch_size): + raise ValueError( + f"`timestep` length must be 1 or batch_size({batch_size}), got {timestep.shape[0]}" + ) + if timestep.shape[0] == 1 and batch_size > 1: + if self.training: + raise ValueError("During training, timestep length must match batch_size.") + timestep = timestep.expand(batch_size) + return x, timestep, context_mask + + def build_video_to_video_mask( + self, + video_seq_len: int, + video_tokens_per_frame: int, + device: torch.device, + ) -> torch.Tensor: + if video_seq_len <= 0: + raise ValueError(f"`video_seq_len` must be positive, got {video_seq_len}") + if video_tokens_per_frame <= 0: + raise ValueError(f"`video_tokens_per_frame` must be positive, got {video_tokens_per_frame}") + + if self.video_attention_mask_mode == "bidirectional": + return torch.ones((video_seq_len, video_seq_len), dtype=torch.bool, device=device) + + if self.video_attention_mask_mode == "per_frame_causal": + if video_seq_len % video_tokens_per_frame != 0: + raise ValueError( + "`video_seq_len` must be divisible by `video_tokens_per_frame` in `per_frame_causal` mode, " + f"got {video_seq_len} and {video_tokens_per_frame}" + ) + num_video_frames = video_seq_len // video_tokens_per_frame + frame_causal = torch.tril( + torch.ones((num_video_frames, num_video_frames), dtype=torch.bool, device=device) + ) + return frame_causal.repeat_interleave(video_tokens_per_frame, dim=0).repeat_interleave( + video_tokens_per_frame, dim=1 + ) + + if self.video_attention_mask_mode == "first_frame_causal": + video_mask = torch.ones((video_seq_len, video_seq_len), dtype=torch.bool, device=device) + first_frame_tokens = min(video_tokens_per_frame, video_seq_len) + video_mask[:first_frame_tokens, first_frame_tokens:] = False + return video_mask + + raise ValueError(f"Unsupported video attention mask mode: {self.video_attention_mask_mode}") + + def pre_dit( + self, + x: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor | None = None, + action: torch.Tensor | None = None, + fuse_vae_embedding_in_latents: bool = False, + ) -> dict[str, Any]: + x, timestep, context_mask = self._validate_forward_inputs( + x=x, + timestep=timestep, + context=context, + context_mask=context_mask, + action=action, + ) + model_dtype = self.patch_embedding.weight.dtype + x = x.to(dtype=model_dtype) + context = context.to(dtype=model_dtype) + if action is not None: + action = action.to(dtype=model_dtype) + + batch_size = x.shape[0] + patch_h = int(self.patch_size[1]) + patch_w = int(self.patch_size[2]) + if x.shape[3] % patch_h != 0 or x.shape[4] % patch_w != 0: + raise ValueError( + "Latent spatial shape must be divisible by DiT patch size, " + f"got HxW=({x.shape[3]}, {x.shape[4]}), patch=({patch_h}, {patch_w})" + ) + tokens_per_frame = (x.shape[3] // patch_h) * (x.shape[4] // patch_w) + + if not (self.seperated_timestep and fuse_vae_embedding_in_latents): + raise NotImplementedError( + "FastWAM currently requires separated timesteps with fused VAE latents." + ) + + token_timesteps = torch.ones( + (batch_size, x.shape[2], tokens_per_frame), + dtype=model_dtype, + device=timestep.device, + ) * timestep.to(dtype=model_dtype).view(batch_size, 1, 1) + token_timesteps[:, 0, :] = 0 + token_timesteps = token_timesteps.reshape(batch_size, -1) + # Wan keeps the time embedding in fp32: the AdaLN modulation in the vendored + # Head/Block asserts e.dtype == float32 (numerical stability of the scale/shift). + # Upstream guarantees this via an fp32 autocast region, so it holds even when the + # model runs in bf16. Mirror that here, then cast the per-block modulation back to + # model_dtype so the bf16 attention blocks are not upcast to fp32. + with torch.amp.autocast("cuda", dtype=torch.float32): + token_t_emb = sinusoidal_embedding_1d(self.freq_dim, token_timesteps.reshape(-1)).float() + t = self.time_embedding(token_t_emb).reshape(batch_size, -1, self.hidden_dim) + t_mod = self.time_projection(t).unflatten(2, (6, self.hidden_dim)) + t_mod = t_mod.to(dtype=model_dtype) + + x = self.patchify(x) + f, h, w = x.shape[2:] + + context = self.text_embedding(context) + context_len = context.shape[1] + if self.action_conditioned and action is not None: + action_len = action.shape[1] + action_emb = self.action_embedding(action) + action_pos_embed = sinusoidal_embedding_1d( + self.hidden_dim, torch.arange(action_len, device=action_emb.device) + ).to(dtype=action_emb.dtype) + action_emb = action_emb + action_pos_embed.unsqueeze(0) + context = torch.cat([context, action_emb], dim=1) + + num_temporal_groups = f - 1 + if num_temporal_groups <= 0: + raise ValueError( + "Action-conditioned context mask requires at least 2 latent frames when `action` is provided." + ) + if action_emb.shape[1] % num_temporal_groups != 0: + raise ValueError( + f"Action embedding length {action_emb.shape[1]} must be divisible by " + f"number of temporal groups {num_temporal_groups}" + ) + action_group_mask = create_group_causal_attn_mask( + num_temporal_groups=num_temporal_groups, + num_query_per_group=tokens_per_frame, + num_key_per_group=action_len // num_temporal_groups, + mode=self.action_group_causal_mask_mode, + ).to(context.device) + + seq_len = f * h * w + final_context_mask = torch.zeros( + (batch_size, seq_len, context.shape[1]), dtype=torch.bool, device=context.device + ) + final_context_mask[:, :, :context_len] = context_mask.unsqueeze(1).expand(-1, seq_len, -1) + final_context_mask[:, tokens_per_frame:, context_len:] = action_group_mask.unsqueeze(0).expand( + batch_size, -1, -1 + ) + context_mask = final_context_mask + elif self.action_conditioned and action is None: + if f != 1: + raise ValueError( + "Action-conditioned model requires `action` unless running single-frame text-only mode " + "with num_latent_frames=1." + ) + context_mask = context_mask.unsqueeze(1).expand(-1, f * h * w, -1) + else: + context_mask = context_mask.unsqueeze(1).expand(-1, f * h * w, -1) + + x_tokens = rearrange(x, "b c f h w -> b (f h w) c").contiguous() + grid_sizes = torch.tensor([[f, h, w]] * batch_size, dtype=torch.long, device=x_tokens.device) + freqs = {"grid_sizes": grid_sizes, "freqs": self.freqs.to(x_tokens.device)} + + return { + "tokens": x_tokens, + "freqs": freqs, + "t": t, + "t_mod": t_mod, + "context": context, + "context_mask": context_mask, + "meta": { + "grid_sizes": grid_sizes, + "tokens_per_frame": tokens_per_frame, + "batch_size": batch_size, + }, + } + + def post_dit(self, x_tokens: torch.Tensor, pre_state: dict[str, Any]) -> torch.Tensor: + x = self.head(x_tokens, pre_state["t"]) + return torch.stack(super().unpatchify(x, pre_state["meta"]["grid_sizes"])) + + def forward( + self, + x: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + context_mask: torch.Tensor | None = None, + action: torch.Tensor | None = None, + fuse_vae_embedding_in_latents: bool = False, + ): + pre_state = self.pre_dit( + x=x, + timestep=timestep, + context=context, + context_mask=context_mask, + action=action, + fuse_vae_embedding_in_latents=fuse_vae_embedding_in_latents, + ) + x_tokens = pre_state["tokens"] + context_emb = pre_state["context"] + t_mod = pre_state["t_mod"] + freqs = pre_state["freqs"] + context_attn_mask = pre_state["context_mask"] + self_attn_mask = ( + self.build_video_to_video_mask( + video_seq_len=x_tokens.shape[1], + video_tokens_per_frame=int(pre_state["meta"]["tokens_per_frame"]), + device=x_tokens.device, + ) + if self.video_attention_mask_mode != "bidirectional" + else None + ) + + for block in self.blocks: + if self.use_gradient_checkpointing: + x_tokens = gradient_checkpoint_forward( + block, + self.use_gradient_checkpointing, + x_tokens, + context_emb, + t_mod, + freqs, + context_mask=context_attn_mask, + self_attn_mask=self_attn_mask, + ) + else: + x_tokens = block( + x_tokens, + context_emb, + t_mod, + freqs, + context_mask=context_attn_mask, + self_attn_mask=self_attn_mask, + ) + + return self.post_dit(x_tokens, pre_state) diff --git a/src/lerobot/policies/groot/action_head/action_encoder.py b/src/lerobot/policies/groot/action_head/action_encoder.py deleted file mode 100644 index c6fa0a779..000000000 --- a/src/lerobot/policies/groot/action_head/action_encoder.py +++ /dev/null @@ -1,54 +0,0 @@ -# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import torch -import torch.nn as nn - - -def swish(x): - return x * torch.sigmoid(x) - - -class SinusoidalPositionalEncoding(nn.Module): - """ - Produces a sinusoidal encoding of shape (B, T, w) - given timesteps of shape (B, T). - """ - - def __init__(self, embedding_dim): - super().__init__() - self.embedding_dim = embedding_dim - - def forward(self, timesteps): - # timesteps: shape (B, T) - # We'll compute sin/cos frequencies across dim T - timesteps = timesteps.float() # ensure float - - b, t = timesteps.shape - device = timesteps.device - - half_dim = self.embedding_dim // 2 - # typical log space frequencies for sinusoidal encoding - exponent = -torch.arange(half_dim, dtype=torch.float, device=device) * ( - torch.log(torch.tensor(10000.0)) / half_dim - ) - # Expand timesteps to (B, T, 1) then multiply - freqs = timesteps.unsqueeze(-1) * exponent.exp() # (B, T, half_dim) - - sin = torch.sin(freqs) - cos = torch.cos(freqs) - enc = torch.cat([sin, cos], dim=-1) # (B, T, w) - - return enc diff --git a/src/lerobot/policies/groot/action_head/cross_attention_dit.py b/src/lerobot/policies/groot/action_head/cross_attention_dit.py index a4cd1a0b7..697459240 100755 --- a/src/lerobot/policies/groot/action_head/cross_attention_dit.py +++ b/src/lerobot/policies/groot/action_head/cross_attention_dit.py @@ -1,11 +1,12 @@ -# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 +#!/usr/bin/env python + +# Copyright 2025 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # -# http://www.apache.org/licenses/LICENSE-2.0 +# http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, @@ -14,6 +15,7 @@ # limitations under the License. +import logging from typing import TYPE_CHECKING import torch @@ -42,6 +44,9 @@ else: Timesteps = None +logger = logging.getLogger(__name__) + + class TimestepEncoder(nn.Module): def __init__(self, embedding_dim, compute_dtype=torch.float32): require_package("diffusers", extra="groot") @@ -181,8 +186,7 @@ class BasicTransformerBlock(nn.Module): attn_output = self.attn1( norm_hidden_states, encoder_hidden_states=encoder_hidden_states, - attention_mask=attention_mask, - # encoder_attention_mask=encoder_attention_mask, + attention_mask=encoder_attention_mask if encoder_hidden_states is not None else attention_mask, ) if self.final_dropout: attn_output = self.final_dropout(attn_output) @@ -266,8 +270,8 @@ class DiT(ModelMixin, ConfigMixin): self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6) self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim) self.proj_out_2 = nn.Linear(self.inner_dim, self.config.output_dim) - print( - "Total number of DiT parameters: ", + logger.debug( + "Total number of DiT parameters: %d", sum(p.numel() for p in self.parameters() if p.requires_grad), ) @@ -318,6 +322,71 @@ class DiT(ModelMixin, ConfigMixin): return self.proj_out_2(hidden_states) +class AlternateVLDiT(DiT): + """N1.7 DiT variant that alternates cross-attention over image and text tokens.""" + + def __init__(self, *args, attend_text_every_n_blocks: int = 2, **kwargs): + super().__init__(*args, **kwargs) + self.attend_text_every_n_blocks = attend_text_every_n_blocks + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + timestep: torch.LongTensor | None = None, + encoder_attention_mask: torch.Tensor | None = None, + return_all_hidden_states: bool = False, + image_mask: torch.Tensor | None = None, + backbone_attention_mask: torch.Tensor | None = None, + ): + if image_mask is None: + raise ValueError("image_mask is required for AlternateVLDiT.") + if backbone_attention_mask is None: + raise ValueError("backbone_attention_mask is required for AlternateVLDiT.") + + temb = self.timestep_encoder(timestep) + hidden_states = hidden_states.contiguous() + encoder_hidden_states = encoder_hidden_states.contiguous() + + image_attention_mask = image_mask & backbone_attention_mask + non_image_attention_mask = (~image_mask) & backbone_attention_mask + + all_hidden_states = [hidden_states] + if not self.config.interleave_self_attention: + raise ValueError("AlternateVLDiT requires interleave_self_attention=True.") + + for idx, block in enumerate(self.transformer_blocks): + if idx % 2 == 1: + hidden_states = block( + hidden_states, + attention_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + temb=temb, + ) + else: + curr_encoder_attention_mask = ( + non_image_attention_mask + if idx % (2 * self.attend_text_every_n_blocks) == 0 + else image_attention_mask + ) + hidden_states = block( + hidden_states, + attention_mask=None, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=curr_encoder_attention_mask, + temb=temb, + ) + all_hidden_states.append(hidden_states) + + conditioning = temb + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + if return_all_hidden_states: + return self.proj_out_2(hidden_states), all_hidden_states + return self.proj_out_2(hidden_states) + + class SelfAttentionTransformer(ModelMixin, ConfigMixin): _supports_gradient_checkpointing = True @@ -362,8 +431,8 @@ class SelfAttentionTransformer(ModelMixin, ConfigMixin): for _ in range(self.config.num_layers) ] ) - print( - "Total number of SelfAttentionTransformer parameters: ", + logger.debug( + "Total number of SelfAttentionTransformer parameters: %d", sum(p.numel() for p in self.parameters() if p.requires_grad), ) diff --git a/src/lerobot/policies/groot/action_head/flow_matching_action_head.py b/src/lerobot/policies/groot/action_head/flow_matching_action_head.py deleted file mode 100644 index 986820670..000000000 --- a/src/lerobot/policies/groot/action_head/flow_matching_action_head.py +++ /dev/null @@ -1,408 +0,0 @@ -# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from dataclasses import field -from typing import TYPE_CHECKING - -import torch -import torch.nn.functional as F # noqa: N812 -from torch import nn -from torch.distributions import Beta - -from lerobot.utils.import_utils import _transformers_available - -# Conditional import for type checking and lazy loading -if TYPE_CHECKING or _transformers_available: - from transformers import PretrainedConfig - from transformers.feature_extraction_utils import BatchFeature -else: - PretrainedConfig = object - BatchFeature = None - -from .action_encoder import ( - SinusoidalPositionalEncoding, - swish, -) -from .cross_attention_dit import DiT, SelfAttentionTransformer - - -class CategorySpecificLinear(nn.Module): - def __init__(self, num_categories, input_dim, hidden_dim): - super().__init__() - self.num_categories = num_categories - # For each category, we have separate weights and biases. - self.W = nn.Parameter(0.02 * torch.randn(num_categories, input_dim, hidden_dim)) - self.b = nn.Parameter(torch.zeros(num_categories, hidden_dim)) - - def forward(self, x, cat_ids): - selected_w = self.W[cat_ids] - selected_b = self.b[cat_ids] - return torch.bmm(x, selected_w) + selected_b.unsqueeze(1) - - -class CategorySpecificMLP(nn.Module): - def __init__(self, num_categories, input_dim, hidden_dim, output_dim): - super().__init__() - self.num_categories = num_categories - self.layer1 = CategorySpecificLinear(num_categories, input_dim, hidden_dim) - self.layer2 = CategorySpecificLinear(num_categories, hidden_dim, output_dim) - - def forward(self, x, cat_ids): - hidden = F.relu(self.layer1(x, cat_ids)) - return self.layer2(hidden, cat_ids) - - -class MultiEmbodimentActionEncoder(nn.Module): - def __init__(self, action_dim, hidden_size, num_embodiments): - super().__init__() - self.hidden_size = hidden_size - self.num_embodiments = num_embodiments - - # W1: R^{w x d}, W2: R^{w x 2w}, W3: R^{w x w} - self.W1 = CategorySpecificLinear(num_embodiments, action_dim, hidden_size) # (d -> w) - self.W2 = CategorySpecificLinear(num_embodiments, 2 * hidden_size, hidden_size) # (2w -> w) - self.W3 = CategorySpecificLinear(num_embodiments, hidden_size, hidden_size) # (w -> w) - self.pos_encoding = SinusoidalPositionalEncoding(hidden_size) - - def forward(self, actions, timesteps, cat_ids): - """ - actions: shape (B, T, action_dim) - timesteps: shape (B,) -- a single scalar per batch item - cat_ids: shape (B,) - returns: shape (B, T, hidden_size) - """ - b, t, _ = actions.shape - - # 1) Expand each batch's single scalar time 'tau' across all T steps - # so that shape => (B, T) - # e.g. if timesteps is (B,), replicate across T - if timesteps.dim() == 1 and timesteps.shape[0] == b: - # shape (B,) => (B,T) - timesteps = timesteps.unsqueeze(1).expand(-1, t) - else: - raise ValueError("Expected `timesteps` to have shape (B,) so we can replicate across T.") - - # 2) Standard action MLP step for shape => (B, T, w) - a_emb = self.W1(actions, cat_ids) - - # 3) Get the sinusoidal encoding (B, T, w) - tau_emb = self.pos_encoding(timesteps).to(dtype=a_emb.dtype) - - # 4) Concat along last dim => (B, T, 2w), then W2 => (B, T, w), swish - x = torch.cat([a_emb, tau_emb], dim=-1) - x = swish(self.W2(x, cat_ids)) - - # 5) Finally W3 => (B, T, w) - x = self.W3(x, cat_ids) - return x - - -class FlowmatchingActionHeadConfig(PretrainedConfig): - """NOTE: N1.5 uses XEmbFlowmatchingPolicyHeadConfig as action head""" - - add_pos_embed: bool = field(default=True, metadata={"help": "Whether to add positional embedding"}) - model_dtype: str = field(default="float32", metadata={"help": "Model data type."}) - diffusion_model_cfg: dict = field(default=None, metadata={"help": "Diffusion model configuration."}) - input_embedding_dim: int = field(default=1536, metadata={"help": "Input embedding channel dimension."}) - backbone_embedding_dim: int = field( - default=1536, metadata={"help": "Backbone embedding channel dimension."} - ) - - hidden_size: int = field(default=1024, metadata={"help": "Input embedding dimension."}) - max_seq_len: int = field(default=1024, metadata={"help": "Maximum Sequence Length"}) - action_dim: int = field(default=None, metadata={"help": "Action dimension."}) - action_horizon: int = field(default=None, metadata={"help": "Action horizon."}) - noise_beta_alpha: float = field(default=1.5, metadata={"help": ""}) - noise_beta_beta: float = field(default=1.0, metadata={"help": ""}) - noise_s: float = field(default=0.999, metadata={"help": "Flow matching noise Beta distribution s."}) - num_timestep_buckets: int = field( - default=1000, metadata={"help": "Number of timestep discretization buckets."} - ) - num_inference_timesteps: int = field( - default=None, - metadata={"help": "Number of inference steps for noise diffusion."}, - ) - max_num_embodiments: int = field(default=32, metadata={"help": "Number of embodiments."}) - tune_projector: bool = field(default=True, metadata={"help": "Whether to tune the projector."}) - tune_diffusion_model: bool = field( - default=True, metadata={"help": "Whether to tune the diffusion model."} - ) - load_pretrained_det_decode_layer_path: str = field( - default=None, metadata={"help": "Path to pretrained detection model."} - ) - detection_coeff: float = field(default=1.0, metadata={"help": "Detection coefficient."}) - - freeze_decode_layer: bool = field(default=False) - expand_batch: int = field(default=None) - use_vlln: bool = field(default=True) - - vl_self_attention_cfg: dict = field(default=None) - num_target_vision_tokens: int = field(default=32, metadata={"help": "Number of target vision tokens."}) - - def __init__(self, **kwargs): - super().__init__(**kwargs) - for key, value in kwargs.items(): - setattr(self, key, value) - - -class FlowmatchingActionHead(nn.Module): - config_class = FlowmatchingActionHeadConfig - supports_gradient_checkpointing = True - - def __init__( - self, - config: FlowmatchingActionHeadConfig, - ): - super().__init__() - self.hidden_size = config.hidden_size - self.input_embedding_dim = config.input_embedding_dim - - self.model = DiT(**config.diffusion_model_cfg) - self.action_dim = config.action_dim - self.action_horizon = config.action_horizon - self.num_inference_timesteps = config.num_inference_timesteps - - self.state_encoder = CategorySpecificMLP( - num_categories=config.max_num_embodiments, - input_dim=config.max_state_dim, - hidden_dim=self.hidden_size, - output_dim=self.input_embedding_dim, - ) - self.action_encoder = MultiEmbodimentActionEncoder( - action_dim=config.action_dim, - hidden_size=self.input_embedding_dim, - num_embodiments=config.max_num_embodiments, - ) - self.action_decoder = CategorySpecificMLP( - num_categories=config.max_num_embodiments, - input_dim=self.hidden_size, - hidden_dim=self.hidden_size, - output_dim=self.action_dim, - ) - self.future_tokens = nn.Embedding(config.num_target_vision_tokens, self.input_embedding_dim) - nn.init.normal_(self.future_tokens.weight, mean=0.0, std=0.02) - - self.vlln = nn.LayerNorm(config.backbone_embedding_dim) if config.use_vlln else nn.Identity() - self.vl_self_attention = ( - SelfAttentionTransformer(**config.vl_self_attention_cfg) if config.use_vlln else nn.Identity() - ) - - if config.add_pos_embed: - self.position_embedding = nn.Embedding(config.max_seq_len, self.input_embedding_dim) - nn.init.normal_(self.position_embedding.weight, mean=0.0, std=0.02) - - self._noise_beta_alpha = config.noise_beta_alpha - self._noise_beta_beta = config.noise_beta_beta - self._beta_dist = None - self.num_timestep_buckets = config.num_timestep_buckets - self.config = config - self.set_trainable_parameters(config.tune_projector, config.tune_diffusion_model) - - def set_trainable_parameters(self, tune_projector: bool, tune_diffusion_model: bool): - self.tune_projector = tune_projector - self.tune_diffusion_model = tune_diffusion_model - for p in self.parameters(): - p.requires_grad = True - if not tune_projector: - self.state_encoder.requires_grad_(False) - self.action_encoder.requires_grad_(False) - self.action_decoder.requires_grad_(False) - if self.config.add_pos_embed: - self.position_embedding.requires_grad_(False) - if not tune_diffusion_model: - self.model.requires_grad_(False) - print(f"Tune action head projector: {self.tune_projector}") - print(f"Tune action head diffusion model: {self.tune_diffusion_model}") - # Check if any parameters are still trainable. If not, print a warning. - if not tune_projector and not tune_diffusion_model: - for name, p in self.named_parameters(): - if p.requires_grad: - print(f"Action head trainable parameter: {name}") - if not any(p.requires_grad for p in self.parameters()): - print("Warning: No action head trainable parameters found.") - - def set_frozen_modules_to_eval_mode(self): - """ - Huggingface will call model.train() at each training_step. To ensure - the expected behaviors for modules like dropout, batchnorm, etc., we - need to call model.eval() for the frozen modules. - """ - if self.training: - if not self.tune_projector: - self.state_encoder.eval() - self.action_encoder.eval() - self.action_decoder.eval() - if self.config.add_pos_embed: - self.position_embedding.eval() - if not self.tune_diffusion_model: - self.model.eval() - - def sample_time(self, batch_size, device, dtype): - if self._beta_dist is None: - self._beta_dist = Beta(self._noise_beta_alpha, self._noise_beta_beta, validate_args=False) - sample = self._beta_dist.sample([batch_size]).to(device, dtype=dtype) - return (self.config.noise_s - sample) / self.config.noise_s - - def prepare_input(self, batch: dict) -> BatchFeature: - return BatchFeature(data=batch) - - def process_backbone_output(self, backbone_output: BatchFeature) -> BatchFeature: - backbone_features = backbone_output["backbone_features"] - backbone_features = self.vlln(backbone_features) - backbone_features = self.vl_self_attention(backbone_features) - backbone_output["backbone_features"] = backbone_features - return backbone_output - - def forward(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature: - # Set frozen modules to eval - self.set_frozen_modules_to_eval_mode() - - backbone_output = self.process_backbone_output(backbone_output) - - if self.config.expand_batch is not None: - for k, v in backbone_output.items(): - ndim = len(v.shape) - factors = [self.config.expand_batch] - while len(factors) < ndim: - factors.append(1) - factors = tuple(factors) - expanded = v.repeat(*factors) - backbone_output[k] = expanded - - for k, v in action_input.items(): - ndim = len(v.shape) - factors = [self.config.expand_batch] - while len(factors) < ndim: - factors.append(1) - factors = tuple(factors) - expanded = v.repeat(*factors) - action_input[k] = expanded - - # Get vision and language embeddings. - vl_embs = backbone_output.backbone_features - device = vl_embs.device - - # Get embodiment ID. - embodiment_id = action_input.embodiment_id - - # Embed state. - state_features = self.state_encoder(action_input.state, embodiment_id) - - # Embed noised action trajectory. - actions = action_input.action - noise = torch.randn(actions.shape, device=actions.device, dtype=actions.dtype) - t = self.sample_time(actions.shape[0], device=actions.device, dtype=actions.dtype) - t = t[:, None, None] # shape (B,1,1) for broadcast - - noisy_trajectory = (1 - t) * noise + t * actions - velocity = actions - noise - - # Convert (continuous) t -> discrete if needed - t_discretized = (t[:, 0, 0] * self.num_timestep_buckets).long() - action_features = self.action_encoder(noisy_trajectory, t_discretized, embodiment_id) - - # Maybe add position embedding. - if self.config.add_pos_embed: - pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device) - pos_embs = self.position_embedding(pos_ids).unsqueeze(0) - action_features = action_features + pos_embs - - # Join vision, language, state and action embedding along sequence dimension. - future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1) - sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1) - - vl_attn_mask = backbone_output.backbone_attention_mask - - model_output = self.model( - hidden_states=sa_embs, - encoder_hidden_states=vl_embs, - encoder_attention_mask=vl_attn_mask, - timestep=t_discretized, - return_all_hidden_states=False, # NOTE (YL): not using flare now - ) - pred = self.action_decoder(model_output, embodiment_id) - pred_actions = pred[:, -actions.shape[1] :] - - # Slice out only the action portion of pred and target. - action_mask = action_input.action_mask - loss = F.mse_loss(pred_actions, velocity, reduction="none") * action_mask - loss = loss.sum() / action_mask.sum() - output_dict = { - "loss": loss, - } - return BatchFeature(data=output_dict) - - @torch.no_grad() - def get_action(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature: - backbone_output = self.process_backbone_output(backbone_output) - - # Get vision and language embeddings. - vl_embs = backbone_output.backbone_features - embodiment_id = action_input.embodiment_id - - # Embed state. - state_features = self.state_encoder(action_input.state, embodiment_id) - - # Set initial actions as the sampled noise. - batch_size = vl_embs.shape[0] - device = vl_embs.device - actions = torch.randn( - size=(batch_size, self.config.action_horizon, self.config.action_dim), - dtype=vl_embs.dtype, - device=device, - ) - - num_steps = self.num_inference_timesteps - dt = 1.0 / num_steps - - # Run denoising steps. - for t in range(num_steps): - t_cont = t / float(num_steps) # e.g. goes 0, 1/N, 2/N, ... - t_discretized = int(t_cont * self.num_timestep_buckets) - - # Embed noised action trajectory. - timesteps_tensor = torch.full(size=(batch_size,), fill_value=t_discretized, device=device) - action_features = self.action_encoder(actions, timesteps_tensor, embodiment_id) - # Maybe add position embedding. - if self.config.add_pos_embed: - pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device) - pos_embs = self.position_embedding(pos_ids).unsqueeze(0) - action_features = action_features + pos_embs - - # Join vision, language, state and action embedding along sequence dimension. - future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1) - sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1) - - # Run model forward. - model_output = self.model( - hidden_states=sa_embs, - encoder_hidden_states=vl_embs, - timestep=timesteps_tensor, - ) - pred = self.action_decoder(model_output, embodiment_id) - - pred_velocity = pred[:, -self.action_horizon :] - - # Update actions using euler integration. - actions = actions + dt * pred_velocity - return BatchFeature(data={"action_pred": actions}) - - @property - def device(self): - return next(iter(self.parameters())).device - - @property - def dtype(self): - return next(iter(self.parameters())).dtype diff --git a/src/lerobot/policies/groot/configuration_groot.py b/src/lerobot/policies/groot/configuration_groot.py index 17cb631d7..97e08bb76 100644 --- a/src/lerobot/policies/groot/configuration_groot.py +++ b/src/lerobot/policies/groot/configuration_groot.py @@ -14,12 +14,229 @@ # See the License for the specific language governing permissions and # limitations under the License. +import logging +import math from dataclasses import dataclass, field +from pathlib import Path from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature, PreTrainedConfig -from lerobot.optim import AdamWConfig, CosineDecayWithWarmupSchedulerConfig +from lerobot.optim import AdamWConfig, DiffuserSchedulerConfig from lerobot.utils.constants import ACTION, OBS_STATE +from .utils import read_json + +logger = logging.getLogger(__name__) + +GROOT_N1_7 = "n1.7" +# Legacy GR00T N1.5 identifier. N1.5 is NOT a supported model_version (it is +# intentionally absent from _GROOT_MODEL_VERSION_ALIASES so normalize_groot_model_version +# still rejects it). It is retained only so that infer_groot_model_version can recognise +# an N1.5 base path/checkpoint and the N1.7 config/loader can reject the mismatch. +GROOT_N1_5 = "n1.5" +# Canonical guidance appended to every error raised when an N1.5 checkpoint, config, +# or processor pipeline is detected. Keep this message in sync with docs/source/groot.mdx. +GROOT_N1_5_REMOVAL_GUIDANCE = ( + "GR00T N1.5 support was removed from LeRobot. " + "To keep using an N1.5 checkpoint, pin the last release that supports it: " + "`pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 " + "(model_version='n1.7', base model nvidia/GR00T-N1.7-3B)." +) +GROOT_N1_7_BASE_MODEL = "nvidia/GR00T-N1.7-3B" +GROOT_N1_7_BACKBONE_MODEL = "nvidia/Cosmos-Reason2-2B" +# Default GR00T N1.7 training resolution. Fallback if processor_config lacks sizing. Prevents mismatched +# full-res patchification by forcing a resize. Mirrored by GR00T_N1_7_DEFAULTS in groot_n1_7.py. +N1_7_DEFAULT_IMAGE_TARGET_SIZE = (256, 256) +N1_7_DEFAULT_IMAGE_CROP_SIZE = (230, 230) +GROOT_ACTION_DECODE_TRANSFORM_LIBERO = "libero" +# Sentinel meaning "the user did not pick an action decode transform": __post_init__ resolves it +# to the embodiment default ('libero' for 'libero_sim', otherwise None). It is distinct from an +# explicit 'none' (resolved to None) so an opt-out survives a draccus save/load round-trip. +GROOT_ACTION_DECODE_TRANSFORM_AUTO = "auto" + +_GROOT_MODEL_VERSION_ALIASES = { + "n1.7": GROOT_N1_7, + "n1_7": GROOT_N1_7, + "n1d7": GROOT_N1_7, + "n17": GROOT_N1_7, + "1.7": GROOT_N1_7, +} + +# Legacy N1.5 spellings, kept ONLY so they can be detected and rejected with +# GROOT_N1_5_REMOVAL_GUIDANCE (see GROOT_N1_5 above). Never map these to a supported version. +_GROOT_N1_5_VERSION_ALIASES = {"n1.5", "n1_5", "n1d5", "n15", "1.5"} + +_GROOT_ACTION_DECODE_TRANSFORM_ALIASES = { + GROOT_ACTION_DECODE_TRANSFORM_AUTO: GROOT_ACTION_DECODE_TRANSFORM_AUTO, + "none": None, + "": None, + GROOT_ACTION_DECODE_TRANSFORM_LIBERO: GROOT_ACTION_DECODE_TRANSFORM_LIBERO, +} + + +def normalize_groot_model_version(model_version: str) -> str: + normalized = _GROOT_MODEL_VERSION_ALIASES.get(model_version.lower()) + if normalized is None: + supported = GROOT_N1_7 + message = f"Unsupported GR00T model_version '{model_version}'. Supported versions: {supported}." + if model_version.lower() in _GROOT_N1_5_VERSION_ALIASES: + message = f"{message} {GROOT_N1_5_REMOVAL_GUIDANCE}" + raise ValueError(message) + return normalized + + +def normalize_groot_action_decode_transform(transform: str | None) -> str | None: + if transform is None: + return None + normalized = _GROOT_ACTION_DECODE_TRANSFORM_ALIASES.get(transform.lower()) + if normalized is None and transform.lower() not in _GROOT_ACTION_DECODE_TRANSFORM_ALIASES: + supported = ", ".join( + sorted(key for key, value in _GROOT_ACTION_DECODE_TRANSFORM_ALIASES.items() if value is not None) + ) + raise ValueError( + f"Unsupported GR00T N1.7 action decode transform '{transform}'. " + f"Supported transforms: none, {supported}." + ) + return normalized + + +def infer_groot_model_version(model_path: str | None) -> str | None: + if not model_path: + return None + model_path_lower = model_path.lower() + if "gr00t-n1.7" in model_path_lower or "gr00t_n1.7" in model_path_lower: + return GROOT_N1_7 + # Detect legacy N1.5 paths so the N1.7 config/loader can reject the mismatch. + # N1.5 is unsupported, but it must still be recognised here to fail loudly + # rather than silently treating an N1.5 checkpoint as N1.7. + if "gr00t-n1.5" in model_path_lower or "gr00t_n1.5" in model_path_lower: + return GROOT_N1_5 + config_version = _infer_groot_model_version_from_local_config(model_path) + if config_version is not None: + return config_version + return None + + +def is_raw_groot_n1_7_checkpoint(model_path: str | Path | None) -> bool: + if model_path is None: + return False + + path = Path(model_path).expanduser() + if path.is_dir(): + config_path = path / "config.json" + elif path.name == "config.json": + config_path = path + else: + return False + + config = read_json(config_path) + return "type" not in config and _infer_groot_model_version_from_config(config) == GROOT_N1_7 + + +def infer_groot_n1_7_embodiment_tag(model_path: str | Path | None) -> str | None: + if model_path is None: + return None + + processor_config_path = Path(model_path).expanduser() / "processor_config.json" + processor_config = read_json(processor_config_path) + + modality_configs = processor_config.get("processor_kwargs", {}).get("modality_configs", {}) + if not isinstance(modality_configs, dict): + return None + if "libero_sim" in modality_configs: + return "libero_sim" + if len(modality_configs) == 1: + return next(iter(modality_configs)) + return None + + +def infer_groot_n1_7_action_horizon( + model_path: str | Path | None, embodiment_tag: str | None = None +) -> int | None: + if model_path is None: + return None + + processor_config_path = Path(model_path).expanduser() / "processor_config.json" + processor_config = read_json(processor_config_path) + + processor_kwargs = processor_config.get("processor_kwargs", {}) + if not isinstance(processor_kwargs, dict): + return None + modality_configs = processor_kwargs.get("modality_configs", {}) + if not isinstance(modality_configs, dict): + return None + + if embodiment_tag is None: + embodiment_tag = infer_groot_n1_7_embodiment_tag(model_path) + if embodiment_tag is None: + return None + + embodiment_config = modality_configs.get(embodiment_tag, {}) + if not isinstance(embodiment_config, dict): + return None + action_config = embodiment_config.get("action", {}) + if not isinstance(action_config, dict): + return None + delta_indices = action_config.get("delta_indices", []) + if not isinstance(delta_indices, list): + return None + return len(delta_indices) or None + + +def infer_groot_n1_7_action_execution_horizon( + model_path: str | Path | None, embodiment_tag: str | None = None +) -> int | None: + action_horizon = infer_groot_n1_7_action_horizon(model_path, embodiment_tag) + if action_horizon is None: + return None + + if embodiment_tag is None: + embodiment_tag = infer_groot_n1_7_embodiment_tag(model_path) + if embodiment_tag == "libero_sim": + # NVIDIA's N1.7 LIBERO rollout wrapper replans after 8 of the 16 decoded + # actions. Keeping that execution cadence avoids stale open-loop chunks. + return min(action_horizon, 8) + return action_horizon + + +def _infer_groot_model_version_from_local_config(model_path: str) -> str | None: + path = Path(model_path).expanduser() + if path.is_dir(): + config_path = path / "config.json" + elif path.name == "config.json": + config_path = path + else: + return None + + return _infer_groot_model_version_from_config(read_json(config_path)) + + +def _infer_groot_model_version_from_config(config: dict) -> str | None: + model_version = config.get("model_version") + if isinstance(model_version, str): + if model_version.lower() in _GROOT_N1_5_VERSION_ALIASES: + return GROOT_N1_5 + try: + return normalize_groot_model_version(model_version) + except ValueError: + return None + + candidates = [config.get("model_type"), *(config.get("architectures") or [])] + for candidate in candidates: + if not isinstance(candidate, str): + continue + normalized = candidate.lower().replace("-", "_") + if normalized in {"gr00tn1d7", "gr00t_n1d7", "gr00t_n1_7"}: + return GROOT_N1_7 + if normalized in {"gr00t_n1_5", "gr00tn1_5", "gr00t_n15", "gr00t_n1d5", "gr00tn1d5"}: + return GROOT_N1_5 + if config.get("model_name") == GROOT_N1_7_BACKBONE_MODEL: + return GROOT_N1_7 + # The Eagle VLM backbone is specific to pre-N1.7 GR00T checkpoints (N1.7 uses Cosmos/Qwen3-VL). + backbone_cfg = config.get("backbone_cfg") + if isinstance(backbone_cfg, dict) and "eagle_path" in backbone_cfg: + return GROOT_N1_5 + return None + @PreTrainedConfig.register_subclass("groot") @dataclass @@ -28,35 +245,44 @@ class GrootConfig(PreTrainedConfig): # Basic policy settings n_obs_steps: int = 1 - chunk_size: int = 50 - n_action_steps: int = 50 + chunk_size: int = 40 + n_action_steps: int = 40 # Dimension settings (must match pretrained GR00T model expectations) # Maximum state dimension. Shorter states will be zero-padded. - max_state_dim: int = 64 + max_state_dim: int = 132 # Maximum action dimension. Shorter actions will be zero-padded. - max_action_dim: int = 32 + max_action_dim: int = 132 - # Normalization (start with identity, adjust as needed) + # GR00T normalizes state/action internally in its processor steps (min/max with + # q01/q99 percentiles, per embodiment), and the Qwen3-VL backbone's image processor + # handles image normalization. The policy therefore does NOT use LeRobot's + # NormalizerProcessorStep/UnnormalizerProcessorStep, so this mapping is intentionally + # IDENTITY for every feature and is not consulted by make_groot_pre_post_processors. normalization_mapping: dict[str, NormalizationMode] = field( default_factory=lambda: { "VISUAL": NormalizationMode.IDENTITY, - "STATE": NormalizationMode.MEAN_STD, - "ACTION": NormalizationMode.MEAN_STD, + "STATE": NormalizationMode.IDENTITY, + "ACTION": NormalizationMode.IDENTITY, } ) - # Image preprocessing (adjust to match Groot's expected input) - image_size: tuple[int, int] = (224, 224) + # Groot-specific model parameters - # Groot-specific model parameters (from groot_finetune_script.py) + # Path or HuggingFace model ID for the base GR00T N1.7 model whose backbone weights and + # checkpoint sidecars (statistics.json, processor_config.json, ...) are loaded. This is the + # model *source*, and is intentionally distinct from the inherited `pretrained_path`: + # `pretrained_path` (`--policy.path`) points at a saved LeRobot checkpoint directory whose + # `config.json` carries a `type` field, whereas a raw NVIDIA GR00T checkpoint has no such + # field and so can only be loaded through `base_model_path` (`--policy.base_model_path`). + # Defaults to GROOT_N1_7_BASE_MODEL when unset (resolved in __post_init__). + base_model_path: str | None = None - # Path or HuggingFace model ID for the base Groot model - base_model_path: str = "nvidia/GR00T-N1.5-3B" - - # HF repo ID (or local path) that hosts vocab.json and merges.txt for Eagle tokenizer. - tokenizer_assets_repo: str = "lerobot/eagle2hg-processor-groot-n1p5" + # Optional named action transform applied after raw N1.7 checkpoint decoding and before env.step(). + # 'auto' (default) resolves to the embodiment default ('libero' for 'libero_sim', otherwise no + # transform). Pass 'none' to explicitly disable the transform, including for 'libero_sim'. + action_decode_transform: str | None = GROOT_ACTION_DECODE_TRANSFORM_AUTO # Embodiment tag to use for training (e.g. 'new_embodiment', 'gr1') embodiment_tag: str = "new_embodiment" @@ -75,38 +301,67 @@ class GrootConfig(PreTrainedConfig): # Whether to fine-tune the diffusion model tune_diffusion_model: bool = True - # LoRA parameters (from groot_finetune_script.py) - # Rank for the LORA model. If 0, no LORA will be used. - lora_rank: int = 0 + # Whether to fine-tune the VL LayerNorm + VL self-attention projector in the action head. + tune_vlln: bool = True - # Alpha value for the LORA model - lora_alpha: int = 16 + # Number of top LLM backbone layers to fine-tune (0 = none). Lets you adapt just the final + # language layers without unfreezing the whole backbone; independent of `tune_llm`, which tunes + # the entire LLM. + tune_top_llm_layers: int = 0 - # Dropout rate for the LORA model - lora_dropout: float = 0.1 + # Inference-time knob: Number of flow-matching denoising steps used to decode an action chunk. + # Trades inference latency for action quality. + # None keeps the checkpoint value (GR00T N1.7 default: 4). + num_inference_timesteps: int | None = None - # Whether to use the full model for LORA - lora_full_model: bool = False + # Inference-time knob: Real-Time Chunking (RTC) overlap-blend ramp rate, used when the RTC engine + # supplies a previous-chunk prefix. Higher values blend the overlapping prefix more aggressively. + # None keeps the checkpoint value (GR00T N1.7 default: 6.0). + rtc_ramp_rate: float | None = None - # Training parameters (matching groot_finetune_script.py) + # Inference-time knob: Whether to request the flash-attention-2 kernel for the Qwen3-VL backbone. + # flash-attn is an optional, user-managed optimization; when it is absent (the default), + # the backbone transparently falls back to SDPA, which is numerically equivalent. + # Set to True only after installing a flash-attn build matching your torch/CUDA env. + use_flash_attention: bool = False + + # Enable GR00T-style state-relative action chunks (action chunk expressed relative to the current + # observation state). + use_relative_actions: bool = False + + # relative_exclude_joints names the action dimensions that stay absolute; the + # match is substring/case-insensitive against the dataset action feature names. With the empty + # default every dimension is treated as relative, including the gripper -- set e.g. ["gripper"] to + # keep the gripper absolute, matching the Isaac-GR00T single-arm + absolute-gripper convention. + relative_exclude_joints: list[str] = field(default_factory=list) + + # Training parameters optimizer_lr: float = 1e-4 - optimizer_betas: tuple[float, float] = (0.95, 0.999) + # Isaac-GR00T N1.7 fine-tunes with AdamW betas (0.9, 0.999). + optimizer_betas: tuple[float, float] = (0.9, 0.999) optimizer_eps: float = 1e-8 optimizer_weight_decay: float = 1e-5 warmup_ratio: float = 0.05 use_bf16: bool = True + # The native N1.7 fine-tuning recipe keeps model parameters in FP32 and computes under BF16 autocast. + model_params_fp32: bool = True - # Dataset parameters - # Video backend to use for training ('decord' or 'torchvision_av') + # TODO(Steven): Remove these deprecated fields in a future release. + # Deprecated Isaac-GR00T runner / GR00T N1.5 fields, plus the (never-wired) LoRA fields — all + # unused by the LeRobot N1.7 implementation except the `tokenizer_assets_repo` N1.5 tripwire and + # the `image_size` legacy remap in __post_init__. They are kept ONLY so a config.json saved by an + # earlier lerobot release (notably a GR00T N1.5 checkpoint) still parses under draccus — which + # rejects unknown fields — and is then rejected with a clear N1.5 removal message rather than an + # opaque draccus decoding error. + image_size: tuple[int, int] = (256, 256) # image sizing is handled by the backbone's image processor. + tokenizer_assets_repo: str | None = None + lora_rank: int = 0 + lora_alpha: int = 16 + lora_dropout: float = 0.1 + lora_full_model: bool = False video_backend: str = "decord" - - # Whether to balance dataset weights in mixture datasets balance_dataset_weights: bool = True - - # Whether to sample trajectories weighted by their length balance_trajectory_weights: bool = True - - # Optional dataset paths for delegating training to Isaac-GR00T runner dataset_paths: list[str] | None = None output_dir: str = "./tmp/gr00t" save_steps: int = 1000 @@ -117,6 +372,65 @@ class GrootConfig(PreTrainedConfig): resume: bool = False def __post_init__(self): + if self.tokenizer_assets_repo is not None: + raise ValueError( + "Config sets 'tokenizer_assets_repo', which only existed for GR00T N1.5; this looks " + f"like a legacy GR00T N1.5 checkpoint or config. {GROOT_N1_5_REMOVAL_GUIDANCE}" + ) + + self.action_decode_transform = normalize_groot_action_decode_transform(self.action_decode_transform) + if self.base_model_path is None: + self.base_model_path = GROOT_N1_7_BASE_MODEL + + # The N1.7 LIBERO checkpoints emit a [0, 1] gripper action, but the LIBERO + # simulator expects the OpenVLA/[-1, 1] sign convention. NVIDIA's rollout + # wrapper applies this conversion; mirror it here so eval on the + # 'libero_sim' embodiment grasps correctly instead of scoring 0% success. + # This matches the embodiment-specific handling already done for the + # action execution horizon (see infer_groot_n1_7_action_execution_horizon). + # Only the 'auto' sentinel resolves to the embodiment default; an explicit + # 'none' (normalized to None above) keeps the transform disabled. + if self.action_decode_transform == GROOT_ACTION_DECODE_TRANSFORM_AUTO: + self.action_decode_transform = ( + GROOT_ACTION_DECODE_TRANSFORM_LIBERO if self.embodiment_tag == "libero_sim" else None + ) + + # GR00T N1.5-era default values (e.g. --policy.chunk_size=50 from old commands or + # stale configs) are migrated to the values the N1.7 checkpoints expect, with a + # warning. The dataclass defaults are already the N1.7 values, so a plain + # GrootConfig() never triggers this. + legacy_default_remaps = ( + ("max_state_dim", 64, 132), + ("max_action_dim", 32, 132), + ("chunk_size", 50, 40), + ("n_action_steps", 50, 40), + ("image_size", (224, 224), (256, 256)), + ) + for field_name, legacy_value, n1_7_value in legacy_default_remaps: + current_value = getattr(self, field_name) + if isinstance(legacy_value, tuple): + current_value = tuple(current_value) + if current_value == legacy_value: + logger.warning( + "GrootConfig.%s=%s matches a legacy GR00T N1.5-era default; remapping it to %s, " + "the value expected by GR00T N1.7 checkpoints. Set a different value explicitly " + "if this is not what you want.", + field_name, + legacy_value, + n1_7_value, + ) + setattr(self, field_name, n1_7_value) + + inferred_version = infer_groot_model_version(self.base_model_path) + if inferred_version is not None and inferred_version != GROOT_N1_7: + message = ( + f"GR00T model_version '{GROOT_N1_7}' does not match base_model_path " + f"'{self.base_model_path}', which looks like '{inferred_version}'." + ) + if inferred_version == GROOT_N1_5: + message = f"{message} {GROOT_N1_5_REMOVAL_GUIDANCE}" + raise ValueError(message) + super().__post_init__() if self.n_action_steps > self.chunk_size: @@ -124,9 +438,6 @@ class GrootConfig(PreTrainedConfig): f"n_action_steps ({self.n_action_steps}) cannot exceed chunk_size ({self.chunk_size})" ) - # groot_repo_path is now optional since we ported the components - # No validation needed - def validate_features(self) -> None: """Validate and set up input/output features for Groot.""" image_features = [key for key, feat in self.input_features.items() if feat.type == FeatureType.VISUAL] @@ -173,15 +484,20 @@ class GrootConfig(PreTrainedConfig): betas=self.optimizer_betas, eps=self.optimizer_eps, weight_decay=self.optimizer_weight_decay, + grad_clip_norm=1.0, ) - def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig: - """Return scheduler configuration.""" - return CosineDecayWithWarmupSchedulerConfig( - num_warmup_steps=int(10000 * self.warmup_ratio), # 5% warmup by default - num_decay_steps=10000, # Adjust based on training steps - peak_lr=self.optimizer_lr, - decay_lr=self.optimizer_lr * 0.1, + def get_scheduler_preset(self) -> DiffuserSchedulerConfig: + """Return scheduler configuration. + + Isaac-GR00T uses the HF Trainer cosine schedule with ~5% warmup over the + actual training update count; DiffuserSchedulerConfig wraps the same + diffusers/transformers `get_scheduler("cosine")` implementation and + derives num_training_steps from the outer --steps value at runtime. + """ + return DiffuserSchedulerConfig( + name="cosine", + num_warmup_steps=math.ceil(self.max_steps * self.warmup_ratio), ) @property @@ -192,7 +508,15 @@ class GrootConfig(PreTrainedConfig): @property def action_delta_indices(self) -> list[int]: """Return indices for delta actions.""" - return list(range(min(self.chunk_size, 16))) + model_action_horizon = ( + infer_groot_n1_7_action_horizon(self.base_model_path, self.embodiment_tag) or 40 + ) + return list(range(min(self.chunk_size, model_action_horizon))) + + @property + def drop_n_last_frames(self) -> int: + """Exclude episode tails that cannot supply a complete N1.7 action chunk.""" + return max(0, len(self.action_delta_indices) - 1) @property def reward_delta_indices(self) -> None: diff --git a/src/lerobot/policies/groot/eagle2_hg_model/configuration_eagle2_5_vl.py b/src/lerobot/policies/groot/eagle2_hg_model/configuration_eagle2_5_vl.py deleted file mode 100755 index 526b4f7a2..000000000 --- a/src/lerobot/policies/groot/eagle2_hg_model/configuration_eagle2_5_vl.py +++ /dev/null @@ -1,135 +0,0 @@ -# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -import copy - -from transformers.configuration_utils import PretrainedConfig -from transformers.models.llama.configuration_llama import LlamaConfig -from transformers.models.qwen2.configuration_qwen2 import Qwen2Config -from transformers.models.qwen3.configuration_qwen3 import Qwen3Config -from transformers.models.siglip.configuration_siglip import SiglipVisionConfig -from transformers.utils import logging - -logger = logging.get_logger(__name__) - - -class Eagle25VLConfig(PretrainedConfig): - model_type = "eagle_2_5_vl" - is_composition = True - sub_configs = {"vision_config": SiglipVisionConfig, "text_config": Qwen2Config} - - def __init__( - self, - vision_config=None, - text_config=None, - use_backbone_lora=0, - use_llm_lora=0, - pad2square=False, - select_layer=-4, - force_image_size=None, - downsample_ratio=0.5, - template=None, - dynamic_image_size=False, - use_thumbnail=False, - loss_version="v1", - min_dynamic_tiles=1, - max_dynamic_tiles=6, - mlp_checkpoint=False, - initializer_range=0.02, - _attn_implementation="flash_attention_2", - _attn_implementation_autoset=False, - llm_config=None, - image_token_index=None, - use_pixel_shuffle=True, - mlp_connector_layers=2, - **kwargs, - ): - super().__init__(**kwargs) - - if vision_config is None: - vision_config = {"model_type": "siglip_vision_model"} - logger.info("vision_config is None. Initializing the InternVisionConfig with default values.") - - if text_config is None: - text_config = {"architectures": ["Qwen2ForCausalLM"]} - logger.info( - "text_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`)." - ) - - if vision_config["model_type"] == "siglip_vision_model": - self.vision_config = SiglipVisionConfig(**vision_config) - else: - raise ValueError("Unsupported model_type: {}".format(vision_config["model_type"])) - - if text_config["architectures"][0] == "LlamaForCausalLM": - self.text_config = LlamaConfig(**text_config) - elif text_config["architectures"][0] == "Qwen2ForCausalLM": - self.text_config = Qwen2Config(**text_config) - elif text_config["architectures"][0] == "Qwen3ForCausalLM": - self.text_config = Qwen3Config(**text_config) - else: - raise ValueError("Unsupported architecture: {}".format(text_config["architectures"][0])) - self.use_backbone_lora = use_backbone_lora - self.use_llm_lora = use_llm_lora - self.mlp_checkpoint = mlp_checkpoint - self.pad2square = pad2square - self.select_layer = select_layer - self.force_image_size = force_image_size - self.downsample_ratio = downsample_ratio - self.template = template - self.dynamic_image_size = dynamic_image_size - self.use_thumbnail = use_thumbnail - self.loss_version = loss_version - self.initializer_range = initializer_range - self.min_dynamic_tiles = min_dynamic_tiles - self.max_dynamic_tiles = max_dynamic_tiles - self.tie_word_embeddings = self.text_config.tie_word_embeddings - self._attn_implementation = _attn_implementation - self._attn_implementation_autoset = _attn_implementation_autoset - self.image_token_index = image_token_index - self.use_pixel_shuffle = use_pixel_shuffle - self.mlp_connector_layers = mlp_connector_layers - logger.info(f"min_dynamic_tiles: {self.min_dynamic_tiles}") - logger.info(f"max_dynamic_tiles: {self.max_dynamic_tiles}") - - def to_dict(self): - """ - Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`]. - - Returns: - `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance, - """ - output = copy.deepcopy(self.__dict__) - output["vision_config"] = self.vision_config.to_dict() - output["text_config"] = self.text_config.to_dict() - output["model_type"] = self.__class__.model_type - output["use_backbone_lora"] = self.use_backbone_lora - output["use_llm_lora"] = self.use_llm_lora - output["pad2square"] = self.pad2square - output["select_layer"] = self.select_layer - output["force_image_size"] = self.force_image_size - output["downsample_ratio"] = self.downsample_ratio - output["template"] = self.template - output["dynamic_image_size"] = self.dynamic_image_size - output["use_thumbnail"] = self.use_thumbnail - output["min_dynamic_tiles"] = self.min_dynamic_tiles - output["max_dynamic_tiles"] = self.max_dynamic_tiles - output["tie_word_embeddings"] = self.tie_word_embeddings - output["_attn_implementation"] = self._attn_implementation - output["_attn_implementation_autoset"] = self._attn_implementation_autoset - output["use_pixel_shuffle"] = self.use_pixel_shuffle - output["mlp_connector_layers"] = self.mlp_connector_layers - return output diff --git a/src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py b/src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py deleted file mode 100644 index 90e9dcecc..000000000 --- a/src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py +++ /dev/null @@ -1,503 +0,0 @@ -# -------------------------------------------------------- -# NVIDIA -# Copyright (c) 2025 NVIDIA -# Licensed under The MIT License [see LICENSE for details] -# -------------------------------------------------------- - -from __future__ import annotations - -# copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py -from transformers.image_processing_utils import ( - BatchFeature, - get_patch_output_size, -) -from transformers.image_processing_utils_fast import ( - BaseImageProcessorFast, - ImagesKwargs, - group_images_by_shape, - reorder_images, -) -from transformers.image_utils import ( - IMAGENET_STANDARD_MEAN, # 0.5, 0.5, 0.5 - IMAGENET_STANDARD_STD, # 0.5, 0.5, 0.5 - ChannelDimension, - ImageInput, - PILImageResampling, - SizeDict, - get_image_size, - make_flat_list_of_images, - validate_kwargs, -) -from transformers.processing_utils import Unpack -from transformers.utils import ( - TensorType, - add_start_docstrings, - is_torch_available, - is_torchvision_v2_available, -) -from transformers.video_utils import VideoInput - -if is_torch_available(): - import torch -if is_torchvision_v2_available(): - from torchvision.transforms.v2 import functional as F # noqa: N812 - from transformers.image_utils import pil_torch_interpolation_mapping -else: - from torchvision.transforms import functional as F # noqa: N812 - - -def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> torch.Tensor: - """Crop the given numpy array. - - Args: - img (torch.Tensor): Image to be cropped. Format should be (C, H, W). - left (int): The left coordinate of the crop box. - top (int): The top coordinate of the crop box. - right (int): The right coordinate of the crop box. - bottom (int): The bottom coordinate of the crop box. - - Returns: - torch.Tensor: Cropped image. - """ - if not isinstance(img, torch.Tensor): - raise TypeError(f"img should be torch.Tensor. Got {type(img)}") - - if img.ndim not in [2, 3]: - raise ValueError(f"Image should have 2 or 3 dimensions. Got {img.ndim}") - - img_height = img.shape[1] - img_width = img.shape[2] - if top < 0 or left < 0 or bottom > img_height or right > img_width: - raise ValueError("Crop coordinates out of bounds") - - if top >= bottom or left >= right: - raise ValueError("Invalid crop coordinates") - - return img[:, top:bottom, left:right] - - -class Eagle25VLFastImageProcessorKwargs(ImagesKwargs): - max_dynamic_tiles: int | None - min_dynamic_tiles: int | None - use_thumbnail: bool | None - pad_during_tiling: bool | None - do_pad: bool | None - - -@add_start_docstrings( - "Constructs a fast ConvNeXT image processor. Based on [`SiglipImageProcessor`] with incorporation of processing each video frame.", - # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING, TODO: this was depreciated from transformers remove! - """ - image_grid_pinpoints (`List[List[int]]`, *optional*): - A list of possible resolutions to use for processing high resolution images. The best resolution is selected - based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess` - method. Not used for processing videos. - do_pad (`bool`, *optional*): - Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest - number of patches in the batch. Padding will be applied to the bottom and right with zeros. - """, -) -class Eagle25VLImageProcessorFast(BaseImageProcessorFast): - resample = PILImageResampling.BICUBIC - image_mean = IMAGENET_STANDARD_MEAN - image_std = IMAGENET_STANDARD_STD - size = {"height": 448, "width": 448} - default_to_square = False - crop_size = None - do_resize = True - do_center_crop = None - do_rescale = True - do_normalize = True - do_convert_rgb = True - do_pad = True - max_dynamic_tiles = 12 - min_dynamic_tiles = 1 - use_thumbnail = True - pad_during_tiling = False - valid_kwargs = Eagle25VLFastImageProcessorKwargs - model_input_names = ["pixel_values_videos"] - - def __init__(self, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs]): - super().__init__(**kwargs) - - @add_start_docstrings( - # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING_PREPROCESS, TODO: this was depreciated from transformers remove! - """ - max_dynamic_tiles (`int`, *optional*): - The maximum number of dynamic tiles to use for processing high resolution images. - min_dynamic_tiles (`int`, *optional*): - The minimum number of dynamic tiles to use for processing high resolution images. - use_thumbnail (`bool`, *optional*): - Whether to use a thumbnail for processing high resolution images. - pad_during_tiling (`bool`, *optional*): - Whether to pad the image during tiling. - do_pad (`bool`, *optional*): - Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest - number of patches in the batch. Padding will be applied to the bottom and right with zeros. - """, - ) - - # NOTE(YL): we will overload the preprocess method to add the image_flags - # def preprocess( - # self, images: ImageInput, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs] - # ) -> BatchFeature: - # return super().preprocess(images, **kwargs) - - def _prepare_images_structure( - self, - images: ImageInput, - expected_ndims: int = 3, - ) -> ImageInput: - """ - Prepare the images structure for processing. - - Args: - images (`ImageInput`): - The input images to process. - expected_ndims (`int`, *optional*, defaults to 3): - Expected number of dimensions for the images (added for transformers >=4.53.0 compatibility). - - Returns: - `ImageInput`: The images with a valid nesting. - """ - return make_flat_list_of_images(images) - - def _resize_for_patching( - self, - image: torch.Tensor, - target_resolution: tuple, - interpolation: F.InterpolationMode, - input_data_format: ChannelDimension, - ) -> torch.Tensor: - """ - Resizes an image to a target resolution while maintaining aspect ratio. - - Args: - image ("torch.Tensor"): - The input image. - target_resolution (tuple): - The target resolution (height, width) of the image. - interpolation (`InterpolationMode`): - Resampling filter to use if resizing the image. - input_data_format (`ChannelDimension` or `str`): - The channel dimension format of the input image. - - Returns: - "torch.Tensor": The resized and padded image. - """ - new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format) - - # Resize the image - resized_image = F.resize(image, (new_height, new_width), interpolation=interpolation) - - return resized_image - - def find_closest_aspect_ratio(self, aspect_ratio, target_ratios, width, height, image_size): - """ - previous version mainly focus on ratio. - We also consider area ratio here. - """ - best_factor = float("-inf") - best_ratio = (1, 1) - area = width * height - for ratio in target_ratios: - target_aspect_ratio = ratio[0] / ratio[1] - # ratio_diff = abs(aspect_ratio - target_aspect_ratio) - # area_ratio = (ratio[0] * ratio[1] * image_size * image_size) / area - """ - new area > 60% of original image area is enough. - """ - factor_based_on_area_n_ratio = min( - (ratio[0] * ratio[1] * image_size * image_size) / area, 0.6 - ) * min(target_aspect_ratio / aspect_ratio, aspect_ratio / target_aspect_ratio) - - if factor_based_on_area_n_ratio > best_factor: - best_factor = factor_based_on_area_n_ratio - best_ratio = ratio - - return best_ratio - - def _pad_for_patching( - self, image: torch.Tensor, target_resolution: tuple, input_data_format: ChannelDimension - ) -> torch.Tensor: - """ - Pad an image to a target resolution while maintaining aspect ratio. - """ - target_height, target_width = target_resolution - new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format) - - paste_x = (target_width - new_width) // 2 - paste_y = (target_height - new_height) // 2 - - padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x, paste_y]) - - return padded_image - - def _get_image_patches( - self, - image: torch.Tensor, - min_num: int, - max_num: int, - size: tuple, - tile_size: int, - use_thumbnail: bool, - interpolation: F.InterpolationMode, - pad_during_tiling: bool, - ) -> list[torch.Tensor]: - image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST) - orig_height, orig_width = image_size - aspect_ratio = orig_width / orig_height - - # calculate the existing image aspect ratio - target_ratios = { - (i, j) - for n in range(min_num, max_num + 1) - for i in range(1, n + 1) - for j in range(1, n + 1) - if i * j <= max_num and i * j >= min_num - } - target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) - - # find the closest aspect ratio to the target - target_aspect_ratio = self.find_closest_aspect_ratio( - aspect_ratio, target_ratios, orig_width, orig_height, tile_size - ) - - # calculate the target width and height - target_width = tile_size * target_aspect_ratio[0] - target_height = tile_size * target_aspect_ratio[1] - blocks = target_aspect_ratio[0] * target_aspect_ratio[1] - if pad_during_tiling: - resized_image = self._resize_for_patching( - image, - (target_height, target_width), - interpolation=interpolation, - input_data_format=ChannelDimension.FIRST, - ) - padded_image = self._pad_for_patching( - resized_image, - (target_height, target_width), - input_data_format=ChannelDimension.FIRST, - ) - image_used_to_split = padded_image - else: - image_used_to_split = F.resize(image, (target_height, target_width), interpolation=interpolation) - - processed_tiles = [] - for i in range(blocks): - box = ( - (i % (target_width // tile_size)) * tile_size, - (i // (target_width // tile_size)) * tile_size, - ((i % (target_width // tile_size)) + 1) * tile_size, - ((i // (target_width // tile_size)) + 1) * tile_size, - ) - # split the image - split_img = crop(image_used_to_split, box[0], box[1], box[2], box[3]) - processed_tiles.append(split_img) - assert len(processed_tiles) == blocks - - if use_thumbnail and len(processed_tiles) != 1: - thumbnail_img = F.resize(image, (tile_size, tile_size), interpolation=interpolation) - processed_tiles.append(thumbnail_img) - - return processed_tiles - - def _pad_for_batching( - self, - pixel_values: list[torch.Tensor], - ) -> list[torch.Tensor]: - """ - Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches. - - Args: - pixel_values (`List[torch.Tensor]`): - An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`) - - Returns: - List[`torch.Tensor`]: The padded images. - """ - max_patch = max(len(x) for x in pixel_values) - pixel_values = [ - torch.nn.functional.pad(image, pad=[0, 0, 0, 0, 0, 0, 0, max_patch - image.shape[0]]) - for image in pixel_values - ] - - return pixel_values - - def _preprocess( - self, - images: list[torch.Tensor], - do_resize: bool, - size: SizeDict, - max_dynamic_tiles: int, - min_dynamic_tiles: int, - use_thumbnail: bool, - pad_during_tiling: bool, - interpolation: F.InterpolationMode | None, - do_center_crop: bool, - crop_size: SizeDict, - do_rescale: bool, - rescale_factor: float, - do_normalize: bool, - image_mean: float | list[float] | None, - image_std: float | list[float] | None, - do_pad: bool, - return_tensors: str | TensorType | None, - pad_size: SizeDict | None = None, # Added for transformers >=4.53.0 compatibility - disable_grouping: bool | None = None, # Added for transformers >=4.53.0 compatibility - ) -> BatchFeature: - processed_images = [] - image_sizes = [] - # Determine the size tuple - if size and size.height and size.width: - size_tuple = (size.height, size.width) - else: - size_tuple = (size.shortest_edge, size.shortest_edge) - - # Determine the patch size - if crop_size and crop_size.height: - tile_size = crop_size.height - elif size and size.height: - tile_size = size.height - else: - tile_size = size.shortest_edge - - for image in images: - image_patches = self._get_image_patches( - image, - min_num=min_dynamic_tiles, - max_num=max_dynamic_tiles, - size=size_tuple, - tile_size=tile_size, - use_thumbnail=use_thumbnail, - interpolation=interpolation, - pad_during_tiling=pad_during_tiling, - ) - - # Group images by size for batched processing - processed_image_patches_grouped = {} - # Added for transformers >=4.53.0 compatibility - grouped_image_patches, grouped_image_patches_index = group_images_by_shape( - image_patches, - disable_grouping=disable_grouping, - ) - - for shape, stacked_image_patches in grouped_image_patches.items(): - if do_resize: - stacked_image_patches = self.resize( - image=stacked_image_patches, - size=size, - interpolation=interpolation, - ) - if do_center_crop: - stacked_image_patches = self.center_crop(stacked_image_patches, crop_size) - # Fused rescale and normalize - stacked_image_patches = self.rescale_and_normalize( - stacked_image_patches, - do_rescale, - rescale_factor, - do_normalize, - image_mean, - image_std, - ) - processed_image_patches_grouped[shape] = stacked_image_patches - processed_image_patches = reorder_images( - processed_image_patches_grouped, grouped_image_patches_index - ) - processed_image_patches = ( - torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches - ) - processed_images.append(processed_image_patches) - image_sizes.append(get_image_size(image, ChannelDimension.FIRST)) - - if do_pad: - processed_images = self._pad_for_batching(processed_images) - - # processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images - processed_images = torch.cat(processed_images, dim=0) if return_tensors else processed_images - return BatchFeature( - data={"pixel_values": processed_images, "image_sizes": image_sizes}, - tensor_type=return_tensors, - ) - - def preprocess( - self, - images: ImageInput, - videos: VideoInput = None, - **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs], - ) -> BatchFeature: - validate_kwargs( - captured_kwargs=kwargs.keys(), - valid_processor_keys=self.valid_kwargs.__annotations__.keys(), - ) - # Set default kwargs from self. This ensures that if a kwarg is not provided - # by the user, it gets its default value from the instance, or is set to None. - for kwarg_name in self.valid_kwargs.__annotations__: - kwargs.setdefault(kwarg_name, getattr(self, kwarg_name, None)) - - # Extract parameters that are only used for preparing the input images - do_convert_rgb = kwargs.pop("do_convert_rgb") - input_data_format = kwargs.pop("input_data_format") - device = kwargs.pop("device") - # Prepare input images - # transformers >= 4.53.0: uses _prepare_image_like_inputs instead of _prepare_input_images - if images is not None: - images = self._prepare_image_like_inputs( - images=images, - do_convert_rgb=do_convert_rgb, - input_data_format=input_data_format, - device=device, - ) - - if videos is not None: - videos = self._prepare_image_like_inputs( - images=videos, - do_convert_rgb=do_convert_rgb, - input_data_format=input_data_format, - device=device, - ) - - # Update kwargs that need further processing before being validated - kwargs = self._further_process_kwargs(**kwargs) - - # Validate kwargs - self._validate_preprocess_kwargs(**kwargs) - - # torch resize uses interpolation instead of resample - # Added for transformers >=4.53.0 compatibility - resample = kwargs.pop("resample", self.resample) - kwargs["interpolation"] = ( - pil_torch_interpolation_mapping[resample] - if isinstance(resample, PILImageResampling | int) - else resample - ) - - # Filter kwargs to only include those accepted by _preprocess - valid_preprocess_kwargs = { - "do_resize", - "size", - "max_dynamic_tiles", - "min_dynamic_tiles", - "use_thumbnail", - "pad_during_tiling", - "interpolation", - "do_center_crop", - "crop_size", - "do_rescale", - "rescale_factor", - "do_normalize", - "image_mean", - "image_std", - "do_pad", - "return_tensors", - "pad_size", - "disable_grouping", - } - filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_preprocess_kwargs} - if images is not None: - return self._preprocess(images, **filtered_kwargs) - elif videos is not None: - return self._preprocess(videos, **filtered_kwargs) - - -__all__ = ["Eagle25VLImageProcessorFast"] diff --git a/src/lerobot/policies/groot/eagle2_hg_model/modeling_eagle2_5_vl.py b/src/lerobot/policies/groot/eagle2_hg_model/modeling_eagle2_5_vl.py deleted file mode 100755 index 6e5532ea4..000000000 --- a/src/lerobot/policies/groot/eagle2_hg_model/modeling_eagle2_5_vl.py +++ /dev/null @@ -1,396 +0,0 @@ -# -------------------------------------------------------- -# NVIDIA -# Copyright (c) 2025 NVIDIA -# Licensed under The MIT License [see LICENSE for details] -# -------------------------------------------------------- - -import inspect - -import torch -import torch.utils.checkpoint as cp -from peft import LoraConfig, get_peft_model -from torch import nn -from torch.nn import CrossEntropyLoss -from transformers import GenerationConfig -from transformers.generation import GenerationMixin -from transformers.modeling_outputs import CausalLMOutputWithPast -from transformers.modeling_utils import PreTrainedModel -from transformers.models.llama.modeling_llama import LlamaForCausalLM -from transformers.models.qwen2.modeling_qwen2 import Qwen2ForCausalLM -from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM -from transformers.models.siglip.modeling_siglip import SiglipVisionModel -from transformers.utils import add_start_docstrings, logging - -from .configuration_eagle2_5_vl import Eagle25VLConfig - -logger = logging.get_logger(__name__) - - -# copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/modeling_llava_onevision.py#L241C1-L280C1 -EAGLE2_5_VL_START_DOCSTRING = r""" - This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the - library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads - etc.) - - This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. - Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage - and behavior. - - Parameters: - config ([`Eagle25VLConfig`]): - Model configuration class with all the parameters of the model. Initializing with a config file does not - load the weights associated with the model, only the configuration. Check out the - [`~PreTrainedModel.from_pretrained`] method to load the model weights. -""" - - -@add_start_docstrings( - "The bare Eagle2_5_VL Model outputting raw hidden-states without any specific head on top.", - EAGLE2_5_VL_START_DOCSTRING, -) -class Eagle25VLPreTrainedModel(PreTrainedModel): - config_class = Eagle25VLConfig - base_model_prefix = "model" - main_input_name = "input_ids" - supports_gradient_checkpointing = True - _no_split_modules = [ - "Qwen2DecoderLayer", - "LlamaDecoderLayer", - "Siglip2EncoderLayer", - "SiglipEncoderLayer", - ] - _skip_keys_device_placement = "past_key_values" - _supports_flash_attn = True - _supports_flash_attn_2 = True - _supports_cache_class = True - _supports_static_cache = True - _supports_quantized_cache = True - _supports_sdpa = True - - def _init_weights(self, module): - std = self.config.initializer_range - if isinstance(module, nn.Linear | nn.Conv2d): - module.weight.data.normal_(mean=0.0, std=std) - if module.bias is not None: - module.bias.data.zero_() - elif isinstance(module, nn.Embedding): - module.weight.data.normal_(mean=0.0, std=std) - if module.padding_idx is not None: - module.weight.data[module.padding_idx].zero_() - - -class Eagle25VLForConditionalGeneration(Eagle25VLPreTrainedModel, GenerationMixin): - config_class = Eagle25VLConfig - - def __init__(self, config: Eagle25VLConfig, vision_model=None, language_model=None): - super().__init__(config) - - image_size = config.force_image_size or config.vision_config.image_size - patch_size = config.vision_config.patch_size - self.patch_size = patch_size - if config.use_pixel_shuffle: - self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio**2)) - else: - self.num_image_token = int((image_size // patch_size) ** 2) - - self.select_layer = config.select_layer - self.downsample_ratio = config.downsample_ratio - self.loss_version = config.loss_version - self.mlp_checkpoint = config.mlp_checkpoint - self.use_pixel_shuffle = config.use_pixel_shuffle - self.mlp_connector_layers = config.mlp_connector_layers - logger.info(f"num_image_token: {self.num_image_token}") - logger.info(f"mlp_checkpoint: {self.mlp_checkpoint}") - if vision_model is not None: - self.vision_model = vision_model - else: - if config.vision_config.model_type == "siglip_vision_model": - config.vision_config._attn_implementation = "flash_attention_2" - self.vision_model = SiglipVisionModel(config.vision_config) - else: - raise NotImplementedError(f"{config.vision_config.model_type} is not implemented.") - - if language_model is not None: - self.language_model = language_model - else: - if config.text_config.architectures[0] == "LlamaForCausalLM": - self.language_model = LlamaForCausalLM(config.text_config) - elif config.text_config.architectures[0] == "Phi3ForCausalLM": - raise NotImplementedError("Phi3 is not implemented.") - # self.language_model = Phi3ForCausalLM(config.text_config) - elif config.text_config.architectures[0] == "Qwen2ForCausalLM": - assert config.text_config._attn_implementation == "flash_attention_2", ( - f"Qwen2 must use flash_attention_2 but got {config.text_config._attn_implementation}" - ) - self.language_model = Qwen2ForCausalLM(config.text_config) - elif config.text_config.architectures[0] == "Qwen3ForCausalLM": - self.language_model = Qwen3ForCausalLM(config.text_config) - else: - raise NotImplementedError(f"{config.text_config.architectures[0]} is not implemented.") - - vit_hidden_size = config.vision_config.hidden_size - llm_hidden_size = config.text_config.hidden_size - - if config.mlp_connector_layers == 2: - self.mlp1 = nn.Sequential( - nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2), - nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size), - nn.GELU(), - nn.Linear(llm_hidden_size, llm_hidden_size), - ) - elif config.mlp_connector_layers == 1 and config.use_pixel_shuffle: - self.mlp1 = nn.Sequential( - nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size), - ) - elif config.mlp_connector_layers == 1 and not config.use_pixel_shuffle: - self.mlp1 = nn.Sequential( - nn.Linear(vit_hidden_size, llm_hidden_size), - ) - else: - raise NotImplementedError(f"{config.mlp_connector_layers} is not implemented.") - - self.image_token_index = config.image_token_index - self.neftune_alpha = None - - if config.use_backbone_lora: - self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora) - - self.use_llm_lora = config.use_llm_lora - if config.use_llm_lora: - self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora) - - self.check_forward_kwargs() - - def check_forward_kwargs(self): - # We intentionally avoid using **kwargs in forward because Hugging Face Transformers - # has special handling for functions with **kwargs parameters that would affect - # how our model is processed during training and inference. - forward_params = inspect.signature(self.forward).parameters - assert not any(k.kind == inspect.Parameter.VAR_KEYWORD for k in forward_params.values()) - - def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05): - lora_config = LoraConfig( - r=r, - target_modules=[ - "self_attn.q_proj", - "self_attn.k_proj", - "self_attn.v_proj", - "self_attn.out_proj", - "mlp.fc1", - "mlp.fc2", - ], - lora_alpha=lora_alpha, - lora_dropout=lora_dropout, - ) - self.vision_model = get_peft_model(self.vision_model, lora_config) - self.vision_model.print_trainable_parameters() - - def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05): - lora_config = LoraConfig( - r=r, - target_modules=[ - "self_attn.q_proj", - "self_attn.k_proj", - "self_attn.v_proj", - "self_attn.o_proj", - "mlp.gate_proj", - "mlp.down_proj", - "mlp.up_proj", - ], - lora_alpha=lora_alpha, - lora_dropout=lora_dropout, - task_type="CAUSAL_LM", - ) - self.language_model = get_peft_model(self.language_model, lora_config) - self.language_model.enable_input_require_grads() - self.language_model.print_trainable_parameters() - self.use_llm_lora = True - - def forward( - self, - pixel_values: torch.FloatTensor, - input_ids: torch.LongTensor = None, - attention_mask: torch.Tensor | None = None, - position_ids: torch.LongTensor | None = None, - image_flags: torch.LongTensor | None = None, - past_key_values: list[torch.FloatTensor] | None = None, - labels: torch.LongTensor | None = None, - use_cache: bool | None = None, - output_attentions: bool | None = None, - output_hidden_states: bool | None = None, - return_dict: bool | None = None, - num_tiles_list: list[torch.Tensor] | None = None, - ) -> tuple | CausalLMOutputWithPast: - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - input_embeds = self.language_model.get_input_embeddings()(input_ids) - - vit_embeds = self.extract_feature(pixel_values) - - if image_flags is not None: - image_flags = image_flags.view(-1) - vit_embeds = vit_embeds[image_flags == 1] - - b, n, c = input_embeds.shape - input_embeds = input_embeds.reshape(b * n, c) - - input_ids = input_ids.reshape(b * n) - selected = input_ids == self.image_token_index - try: - input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, c) - except Exception as e: - vit_embeds = vit_embeds.reshape(-1, c) - print( - f"warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, " - f"vit_embeds.shape={vit_embeds.shape}" - ) - n_token = selected.sum() - input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token] - - input_embeds = input_embeds.reshape(b, n, c) - - outputs = self.language_model( - inputs_embeds=input_embeds, - attention_mask=attention_mask, - position_ids=position_ids, - past_key_values=past_key_values, - use_cache=use_cache, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - ) - logits = outputs.logits - - loss = None - if labels is not None: - # Shift so that tokens < n predict n - shift_logits = logits[..., :-1, :].contiguous() - shift_labels = labels[..., 1:].contiguous() - # Flatten the tokens - loss_fct = CrossEntropyLoss() - shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size) - shift_labels = shift_labels.view(-1) - # Enable model parallelism - shift_labels = shift_labels.to(shift_logits.device) - loss = loss_fct(shift_logits, shift_labels) - - if not return_dict: - output = (logits,) + outputs[1:] - return (loss,) + output if loss is not None else output - - return CausalLMOutputWithPast( - loss=loss, - logits=logits, - past_key_values=outputs.past_key_values, - hidden_states=outputs.hidden_states, - attentions=outputs.attentions, - ) - - def pixel_shuffle(self, x, scale_factor=0.5): - n, w, h, c = x.size() - # N, W, H, C --> N, W, H * scale, C // scale - x = x.view(n, w, int(h * scale_factor), int(c / scale_factor)) - # N, W, H * scale, C // scale --> N, H * scale, W, C // scale - x = x.permute(0, 2, 1, 3).contiguous() - # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2) - x = x.view(n, int(h * scale_factor), int(w * scale_factor), int(c / (scale_factor * scale_factor))) - - x = x.permute(0, 2, 1, 3).contiguous() - return x - - def extract_feature(self, pixel_values): - if self.select_layer == -1: - vit_embeds = self.vision_model( - pixel_values=pixel_values, output_hidden_states=False, return_dict=True - ) - if hasattr(vit_embeds, "last_hidden_state"): - vit_embeds = vit_embeds.last_hidden_state - - else: - vit_embeds = self.vision_model( - pixel_values=pixel_values, output_hidden_states=True, return_dict=True - ).hidden_states[self.select_layer] - - if self.use_pixel_shuffle: - h = w = int(vit_embeds.shape[1] ** 0.5) - vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1) - vit_embeds = self.pixel_shuffle( - vit_embeds, scale_factor=self.downsample_ratio - ) # torch.Size([B, 1024, 1024]) -> torch.Size([B, 16, 16, 4096]) - vit_embeds = vit_embeds.reshape( - vit_embeds.shape[0], -1, vit_embeds.shape[-1] - ) # torch.Size([B, 16, 16, 4096]) -> torch.Size([B, 256, 4096]) - - if self.mlp_checkpoint and vit_embeds.requires_grad: - vit_embeds = cp.checkpoint(self.mlp1, vit_embeds) - else: - vit_embeds = self.mlp1(vit_embeds) - - return vit_embeds - - @torch.no_grad() - def generate( - self, - pixel_values: torch.FloatTensor | None = None, - input_ids: torch.FloatTensor | None = None, - attention_mask: torch.LongTensor | None = None, - visual_features: torch.FloatTensor | None = None, - generation_config: GenerationConfig | None = None, - output_hidden_states: bool | None = None, - image_sizes: list[tuple[int, int]] | None = None, - **generate_kwargs, - ) -> torch.LongTensor: - if pixel_values is not None: - if visual_features is not None: - vit_embeds = visual_features - else: - vit_embeds = self.extract_feature(pixel_values) - - input_embeds = self.language_model.get_input_embeddings()(input_ids) - b, n, c = input_embeds.shape - input_embeds = input_embeds.reshape(b * n, c) - - input_ids = input_ids.reshape(b * n) - selected = input_ids == self.config.image_token_index - assert selected.sum() != 0 - input_embeds[selected] = vit_embeds.reshape(-1, c).to(input_embeds.device) - - input_embeds = input_embeds.reshape(b, n, c) - else: - input_embeds = self.language_model.get_input_embeddings()(input_ids) - - if "use_cache" not in generate_kwargs: - generate_kwargs["use_cache"] = True - - outputs = self.language_model.generate( - inputs_embeds=input_embeds, - attention_mask=attention_mask, - generation_config=generation_config, - output_hidden_states=output_hidden_states, - **generate_kwargs, - ) - - return outputs - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_input_embeddings - def get_input_embeddings(self): - return self.language_model.get_input_embeddings() - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_input_embeddings - def set_input_embeddings(self, value): - self.language_model.set_input_embeddings(value) - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_output_embeddings - def get_output_embeddings(self): - return self.language_model.get_output_embeddings() - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_output_embeddings - def set_output_embeddings(self, new_embeddings): - self.language_model.set_output_embeddings(new_embeddings) - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_decoder - def set_decoder(self, decoder): - self.language_model.set_decoder(decoder) - - # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_decoder - def get_decoder(self): - return self.language_model.get_decoder() diff --git a/src/lerobot/policies/groot/eagle2_hg_model/processing_eagle2_5_vl.py b/src/lerobot/policies/groot/eagle2_hg_model/processing_eagle2_5_vl.py deleted file mode 100755 index b36e70c47..000000000 --- a/src/lerobot/policies/groot/eagle2_hg_model/processing_eagle2_5_vl.py +++ /dev/null @@ -1,541 +0,0 @@ -# Copyright 2024 The HuggingFace Inc. team. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Processor class for Eagle25VL. -copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/processing_llava_onevision.py -""" - -import base64 -import os -import re -from io import BytesIO - -import requests -import torch -from PIL import Image -from transformers.feature_extraction_utils import BatchFeature -from transformers.image_utils import ImageInput -from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack -from transformers.tokenization_utils_base import PreTokenizedInput, TextInput -from transformers.utils import logging -from transformers.video_utils import VideoInput - -logger = logging.get_logger(__name__) - - -FRAME_FACTOR = 2 -FPS = 2.0 -FPS_MIN_FRAMES = 4 -FPS_MAX_FRAMES = 256 - - -def to_rgb(pil_image: Image.Image) -> Image.Image: - if pil_image.mode == "RGBA": - white_background = Image.new("RGB", pil_image.size, (255, 255, 255)) - white_background.paste(pil_image, mask=pil_image.split()[3]) # Use alpha channel as mask - return white_background - else: - return pil_image.convert("RGB") - - -def fetch_image(ele: dict[str, str | Image.Image]) -> Image.Image: - image = ele["image"] if "image" in ele else ele["image_url"] - image_obj = None - if isinstance(image, Image.Image): - image_obj = image - elif image.startswith("http://") or image.startswith("https://"): - response = requests.get(image, stream=True, timeout=10) - image_obj = Image.open(BytesIO(response.content)) - elif image.startswith("file://"): - image_obj = Image.open(image[7:]) - elif image.startswith("data:image"): - if "base64," in image: - _, base64_data = image.split("base64,", 1) - data = base64.b64decode(base64_data) - image_obj = Image.open(BytesIO(data)) - else: - image_obj = Image.open(image) - if image_obj is None: - raise ValueError( - f"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}" - ) - image = to_rgb(image_obj) - if "scale_factor" in ele: - scale_factor = ele["scale_factor"] - image = image.resize((image.width * scale_factor, image.height * scale_factor), Image.BILINEAR) - return image - - -class Eagle25VLProcessorKwargs(ProcessingKwargs, total=False): - # see processing_utils.ProcessingKwargs documentation for usage. - _defaults = { - "text_kwargs": { - "padding": False, - }, - "images_kwargs": {}, - "videos_kwargs": {"max_dynamic_tiles": 1}, - } - - -class Eagle25VLProcessor(ProcessorMixin): - r""" - Constructs a Eagle25VL processor which wraps a Eagle25VL video processor, Eagle25VL image processor and a Eagle25VL tokenizer into a single processor. - - [`Eagle25VLProcessor`] offers all the functionalities of [`Eagle25VLVideoProcessor`], [`Eagle25VLImageProcessor`] and [`Eagle25VLTokenizer`]. See the - [`~Eagle25VLVideoProcessor.__call__`], [`~Eagle25VLProcessor.__call__`] and [`~Eagle25VLProcessor.decode`] for more information. - - Args: - image_processor ([`LlavaOnevisionImageProcessor`], *optional*): - The image processor is a required input. - tokenizer ([`LlamaTokenizerFast`], *optional*): - The tokenizer is a required input. - num_image_tokens (`int`, *optional*): - Number of image tokens for one imagethat will be returned by vision tower. - vision_feature_select_strategy (`str`, *optional*): - The feature selection strategy used to select the vision feature from the vision backbone. - Should be same as in model's config - chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages - in a chat into a tokenizable string. - image_token (`str`, *optional*, defaults to `""`): - Special token used to denote image location. - video_token (`str`, *optional*, defaults to `"