mirror of
https://github.com/huggingface/lerobot.git
synced 2026-06-16 15:57:03 +00:00
Compare commits
1 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 AdilZouitine
|
ab94626b92 |
@@ -30,7 +30,7 @@ pytest -sx tests/test_stuff.py::test_something
|
||||
```
|
||||
|
||||
```bash
|
||||
lerobot-train --some.option=true
|
||||
python -m lerobot.scripts.train --some.option=true
|
||||
```
|
||||
|
||||
## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
|
||||
|
||||
@@ -29,8 +29,8 @@ on:
|
||||
env:
|
||||
UV_VERSION: "0.8.0"
|
||||
PYTHON_VERSION: "3.10"
|
||||
DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
|
||||
DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
|
||||
DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-gpu:latest
|
||||
DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-cpu:latest
|
||||
|
||||
# Ensures that only the latest commit is built, canceling older runs.
|
||||
concurrency:
|
||||
|
||||
@@ -44,7 +44,7 @@ test-end-to-end:
|
||||
${MAKE} DEVICE=$(DEVICE) test-smolvla-ete-eval
|
||||
|
||||
test-act-ete-train:
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=act \
|
||||
--policy.dim_model=64 \
|
||||
--policy.n_action_steps=20 \
|
||||
@@ -68,12 +68,12 @@ test-act-ete-train:
|
||||
--output_dir=tests/outputs/act/
|
||||
|
||||
test-act-ete-train-resume:
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
|
||||
--resume=true
|
||||
|
||||
test-act-ete-eval:
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
|
||||
--policy.device=$(DEVICE) \
|
||||
--env.type=aloha \
|
||||
@@ -82,7 +82,7 @@ test-act-ete-eval:
|
||||
--eval.batch_size=1
|
||||
|
||||
test-diffusion-ete-train:
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=diffusion \
|
||||
--policy.down_dims='[64,128,256]' \
|
||||
--policy.diffusion_step_embed_dim=32 \
|
||||
@@ -106,7 +106,7 @@ test-diffusion-ete-train:
|
||||
--output_dir=tests/outputs/diffusion/
|
||||
|
||||
test-diffusion-ete-eval:
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
|
||||
--policy.device=$(DEVICE) \
|
||||
--env.type=pusht \
|
||||
@@ -115,7 +115,7 @@ test-diffusion-ete-eval:
|
||||
--eval.batch_size=1
|
||||
|
||||
test-tdmpc-ete-train:
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=tdmpc \
|
||||
--policy.device=$(DEVICE) \
|
||||
--policy.push_to_hub=false \
|
||||
@@ -137,7 +137,7 @@ test-tdmpc-ete-train:
|
||||
--output_dir=tests/outputs/tdmpc/
|
||||
|
||||
test-tdmpc-ete-eval:
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
|
||||
--policy.device=$(DEVICE) \
|
||||
--env.type=xarm \
|
||||
@@ -148,7 +148,7 @@ test-tdmpc-ete-eval:
|
||||
|
||||
|
||||
test-smolvla-ete-train:
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=smolvla \
|
||||
--policy.n_action_steps=20 \
|
||||
--policy.chunk_size=20 \
|
||||
@@ -171,7 +171,7 @@ test-smolvla-ete-train:
|
||||
--output_dir=tests/outputs/smolvla/
|
||||
|
||||
test-smolvla-ete-eval:
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=tests/outputs/smolvla/checkpoints/000004/pretrained_model \
|
||||
--policy.device=$(DEVICE) \
|
||||
--env.type=aloha \
|
||||
|
||||
@@ -6,7 +6,7 @@
|
||||
|
||||
<div align="center">
|
||||
|
||||
[](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml?query=branch%3Amain)
|
||||
[](https://github.com/huggingface/lerobot/actions/workflows/nighty.yml?query=branch%3Amain)
|
||||
[](https://www.python.org/downloads/)
|
||||
[](https://github.com/huggingface/lerobot/blob/main/LICENSE)
|
||||
[](https://pypi.org/project/lerobot/)
|
||||
@@ -101,9 +101,6 @@
|
||||
## Installation
|
||||
|
||||
LeRobot works with Python 3.10+ and PyTorch 2.2+.
|
||||
|
||||
### Environment Setup
|
||||
|
||||
Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
|
||||
|
||||
```bash
|
||||
@@ -127,21 +124,10 @@ conda install ffmpeg -c conda-forge
|
||||
>
|
||||
> - _[On Linux only]_ Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
|
||||
|
||||
### Install LeRobot 🤗
|
||||
|
||||
#### From Source
|
||||
|
||||
First, clone the repository and navigate into the directory:
|
||||
Install 🤗 LeRobot:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/lerobot.git
|
||||
cd lerobot
|
||||
```
|
||||
|
||||
Then, install the library in editable mode. This is useful if you plan to contribute to the code.
|
||||
|
||||
```bash
|
||||
pip install -e .
|
||||
pip install lerobot
|
||||
```
|
||||
|
||||
> **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
|
||||
@@ -159,34 +145,6 @@ For instance, to install 🤗 LeRobot with aloha and pusht, use:
|
||||
pip install -e ".[aloha, pusht]"
|
||||
```
|
||||
|
||||
### Installation from PyPI
|
||||
|
||||
**Core Library:**
|
||||
Install the base package with:
|
||||
|
||||
```bash
|
||||
pip install lerobot
|
||||
```
|
||||
|
||||
_This installs only the default dependencies._
|
||||
|
||||
**Extra Features:**
|
||||
To install additional functionality, use one of the following:
|
||||
|
||||
```bash
|
||||
pip install 'lerobot[all]' # All available features
|
||||
pip install 'lerobot[aloha,pusht]' # Specific features (Aloha & Pusht)
|
||||
pip install 'lerobot[feetech]' # Feetech motor support
|
||||
```
|
||||
|
||||
_Replace `[...]` with your desired features._
|
||||
|
||||
**Available Tags:**
|
||||
For a full list of optional dependencies, see:
|
||||
https://pypi.org/project/lerobot/
|
||||
|
||||
### Weights & Biases
|
||||
|
||||
To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
|
||||
|
||||
```bash
|
||||
@@ -276,7 +234,7 @@ Check out [example 2](https://github.com/huggingface/lerobot/blob/main/examples/
|
||||
We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
|
||||
|
||||
```bash
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=lerobot/diffusion_pusht \
|
||||
--env.type=pusht \
|
||||
--eval.batch_size=10 \
|
||||
@@ -288,10 +246,10 @@ lerobot-eval \
|
||||
Note: After training your own policy, you can re-evaluate the checkpoints with:
|
||||
|
||||
```bash
|
||||
lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
|
||||
python -m lerobot.scripts.eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
|
||||
```
|
||||
|
||||
See `lerobot-eval --help` for more instructions.
|
||||
See `python -m lerobot.scripts.eval --help` for more instructions.
|
||||
|
||||
### Train your own policy
|
||||
|
||||
@@ -303,7 +261,7 @@ A link to the wandb logs for the run will also show up in yellow in your termina
|
||||
|
||||
\<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>
|
||||
|
||||
Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.
|
||||
Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python -m lerobot.scripts.eval --help` for more instructions.
|
||||
|
||||
#### Reproduce state-of-the-art (SOTA)
|
||||
|
||||
@@ -311,7 +269,7 @@ We provide some pretrained policies on our [hub page](https://huggingface.co/ler
|
||||
You can reproduce their training by loading the config from their run. Simply running:
|
||||
|
||||
```bash
|
||||
lerobot-train --config_path=lerobot/diffusion_pusht
|
||||
python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
|
||||
```
|
||||
|
||||
reproduces SOTA results for Diffusion Policy on the PushT task.
|
||||
@@ -353,7 +311,7 @@ If you want, you can cite this work with:
|
||||
|
||||
```bibtex
|
||||
@misc{cadene2024lerobot,
|
||||
author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
|
||||
author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascale, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
|
||||
title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
|
||||
howpublished = "\url{https://github.com/huggingface/lerobot}",
|
||||
year = {2024}
|
||||
|
||||
@@ -1,11 +0,0 @@
|
||||
compute_environment: LOCAL_MACHINE
|
||||
debug: false
|
||||
distributed_type: NO
|
||||
downcast_bf16: 'no'
|
||||
enable_cpu_affinity: false
|
||||
machine_rank: 0
|
||||
main_training_function: main
|
||||
mixed_precision: 'no'
|
||||
num_machines: 1
|
||||
num_processes: 1
|
||||
use_cpu: false
|
||||
@@ -1,18 +0,0 @@
|
||||
compute_environment: LOCAL_MACHINE
|
||||
debug: false
|
||||
distributed_type: MULTI_GPU
|
||||
downcast_bf16: 'no'
|
||||
enable_cpu_affinity: false
|
||||
gpu_ids: all
|
||||
machine_rank: 0
|
||||
main_training_function: main
|
||||
mixed_precision: 'no'
|
||||
num_machines: 1
|
||||
num_processes: 2
|
||||
rdzv_backend: static
|
||||
same_network: true
|
||||
tpu_env: []
|
||||
tpu_use_cluster: false
|
||||
tpu_use_sudo: false
|
||||
use_cpu: false
|
||||
dynamo_backend: "no"
|
||||
@@ -29,7 +29,7 @@ ENV DEBIAN_FRONTEND=noninteractive \
|
||||
|
||||
# Install system dependencies and uv (as root)
|
||||
RUN apt-get update && apt-get install -y --no-install-recommends \
|
||||
build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
|
||||
build-essential git curl libglib2.0-0 libegl1-mesa ffmpeg \
|
||||
libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
|
||||
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
|
||||
&& mv /root/.local/bin/uv /usr/local/bin/uv \
|
||||
|
||||
@@ -39,8 +39,6 @@
|
||||
- sections:
|
||||
- local: notebooks
|
||||
title: Notebooks
|
||||
- local: feetech
|
||||
title: Updating Feetech Firmware
|
||||
title: "Resources"
|
||||
- sections:
|
||||
- local: contributing
|
||||
|
||||
@@ -9,7 +9,7 @@ To instantiate a camera, you need a camera identifier. This identifier might cha
|
||||
To find the camera indices of the cameras plugged into your system, run the following script:
|
||||
|
||||
```bash
|
||||
lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
|
||||
python -m lerobot.find_cameras opencv # or realsense for Intel Realsense cameras
|
||||
```
|
||||
|
||||
The output will look something like this if you have two cameras connected:
|
||||
|
||||
@@ -1,71 +0,0 @@
|
||||
# Feetech Motor Firmware Update
|
||||
|
||||
This tutorial guides you through updating the firmware of Feetech motors using the official Feetech software.
|
||||
|
||||
## Prerequisites
|
||||
|
||||
- Windows computer (Feetech software is only available for Windows)
|
||||
- Feetech motor control board
|
||||
- USB cable to connect the control board to your computer
|
||||
- Feetech motors connected to the control board
|
||||
|
||||
## Step 1: Download Feetech Software
|
||||
|
||||
1. Visit the official Feetech software download page: [https://www.feetechrc.com/software.html](https://www.feetechrc.com/software.html)
|
||||
2. Download the latest version of the Feetech debugging software (FD)
|
||||
3. Install the software on your Windows computer
|
||||
|
||||
## Step 2: Hardware Setup
|
||||
|
||||
1. Connect your Feetech motors to the motor control board
|
||||
2. Connect the motor control board to your Windows computer via USB cable
|
||||
3. Ensure power is supplied to the motors
|
||||
|
||||
## Step 3: Configure Connection
|
||||
|
||||
1. Launch the Feetech debugging software
|
||||
2. Select the correct COM port from the port dropdown menu
|
||||
- If unsure which port to use, check Windows Device Manager under "Ports (COM & LPT)"
|
||||
3. Set the appropriate baud rate (typically 1000000 for most Feetech motors)
|
||||
4. Click "Open" to establish communication with the control board
|
||||
|
||||
## Step 4: Scan for Motors
|
||||
|
||||
1. Once connected, click the "Search" button to detect all connected motors
|
||||
2. The software will automatically discover and list all motors on the bus
|
||||
3. Each motor will appear with its ID number
|
||||
|
||||
## Step 5: Update Firmware
|
||||
|
||||
For each motor you want to update:
|
||||
|
||||
1. **Select the motor** from the list by clicking on it
|
||||
2. **Click on Upgrade tab**:
|
||||
3. **Click on Online button**:
|
||||
- If an potential firmware update is found, it will be displayed in the box
|
||||
4. **Click on Upgrade button**:
|
||||
- The update progress will be displayed
|
||||
|
||||
## Step 6: Verify Update
|
||||
|
||||
1. After the update completes, the software should automatically refresh the motor information
|
||||
2. Verify that the firmware version has been updated to the expected version
|
||||
|
||||
## Important Notes
|
||||
|
||||
⚠️ **Warning**: Do not disconnect power or USB during firmware updates, it will potentially brick the motor.
|
||||
|
||||
## Bonus: Motor Debugging on Linux/macOS
|
||||
|
||||
For debugging purposes only, you can use the open-source Feetech Debug Tool:
|
||||
|
||||
- **Repository**: [FT_SCServo_Debug_Qt](https://github.com/CarolinePascal/FT_SCServo_Debug_Qt/tree/fix/port-search-timer)
|
||||
|
||||
### Installation Instructions
|
||||
|
||||
Follow the instructions in the repository to install the tool, for Ubuntu you can directly install it, for MacOS you need to build it from source.
|
||||
|
||||
**Limitations:**
|
||||
|
||||
- This tool is for debugging and parameter adjustment only
|
||||
- Firmware updates must still be done on Windows with official Feetech software
|
||||
@@ -412,7 +412,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
|
||||
To train the classifier, use the `train.py` script with your configuration:
|
||||
|
||||
```bash
|
||||
lerobot-train --config_path path/to/reward_classifier_train_config.json
|
||||
python -m lerobot.scripts.train --config_path path/to/reward_classifier_train_config.json
|
||||
```
|
||||
|
||||
**Deploying and Testing the Model**
|
||||
@@ -458,7 +458,7 @@ The reward classifier will automatically provide rewards based on the visual inp
|
||||
3. **Train the classifier**:
|
||||
|
||||
```bash
|
||||
lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
|
||||
python -m lerobot.scripts.train --config_path src/lerobot/configs/reward_classifier_train_config.json
|
||||
```
|
||||
|
||||
4. **Test the classifier**:
|
||||
|
||||
+11
-11
@@ -19,7 +19,7 @@ pip install -e ".[hopejr]"
|
||||
Before starting calibration and operation, you need to identify the USB ports for each HopeJR component. Run this script to find the USB ports for the arm, hand, glove, and exoskeleton:
|
||||
|
||||
```bash
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
|
||||
This will display the available USB ports and their associated devices. Make note of the port paths (e.g., `/dev/tty.usbmodem58760433331`, `/dev/tty.usbmodem11301`) as you'll need to specify them in the `--robot.port` and `--teleop.port` parameters when recording data, replaying episodes, or running teleoperation scripts.
|
||||
@@ -31,7 +31,7 @@ Before performing teleoperation, HopeJR's limbs need to be calibrated. Calibrati
|
||||
### 1.1 Calibrate Robot Hand
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=hope_jr_hand \
|
||||
--robot.port=/dev/tty.usbmodem58760432281 \
|
||||
--robot.id=blue \
|
||||
@@ -81,7 +81,7 @@ Once you have set the appropriate boundaries for all joints, click "Save" to sav
|
||||
### 1.2 Calibrate Teleoperator Glove
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=homunculus_glove \
|
||||
--teleop.port=/dev/tty.usbmodem11201 \
|
||||
--teleop.id=red \
|
||||
@@ -120,7 +120,7 @@ Once calibration is complete, the system will save the calibration to `/Users/yo
|
||||
### 1.3 Calibrate Robot Arm
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=hope_jr_arm \
|
||||
--robot.port=/dev/tty.usbserial-1110 \
|
||||
--robot.id=white
|
||||
@@ -146,7 +146,7 @@ Use the calibration interface to set the range boundaries for each joint. Move e
|
||||
### 1.4 Calibrate Teleoperator Exoskeleton
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=homunculus_arm \
|
||||
--teleop.port=/dev/tty.usbmodem11201 \
|
||||
--teleop.id=black
|
||||
@@ -178,7 +178,7 @@ Due to global variable conflicts in the Feetech middleware, teleoperation for ar
|
||||
### Hand
|
||||
|
||||
```bash
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=hope_jr_hand \
|
||||
--robot.port=/dev/tty.usbmodem58760432281 \
|
||||
--robot.id=blue \
|
||||
@@ -194,7 +194,7 @@ lerobot-teleoperate \
|
||||
### Arm
|
||||
|
||||
```bash
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=hope_jr_arm \
|
||||
--robot.port=/dev/tty.usbserial-1110 \
|
||||
--robot.id=white \
|
||||
@@ -214,7 +214,7 @@ Record, Replay and Train with Hope-JR is still experimental.
|
||||
This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).
|
||||
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=hope_jr_hand \
|
||||
--robot.port=/dev/tty.usbmodem58760432281 \
|
||||
--robot.id=right \
|
||||
@@ -236,7 +236,7 @@ lerobot-record \
|
||||
### Replay
|
||||
|
||||
```bash
|
||||
lerobot-replay \
|
||||
python -m lerobot.replay \
|
||||
--robot.type=hope_jr_hand \
|
||||
--robot.port=/dev/tty.usbmodem58760432281 \
|
||||
--robot.id=right \
|
||||
@@ -248,7 +248,7 @@ lerobot-replay \
|
||||
### Train
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=nepyope/hand_record_test_with_video_data \
|
||||
--policy.type=act \
|
||||
--output_dir=outputs/train/hopejr_hand \
|
||||
@@ -263,7 +263,7 @@ lerobot-train \
|
||||
This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).
|
||||
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=hope_jr_hand \
|
||||
--robot.port=/dev/tty.usbmodem58760432281 \
|
||||
--robot.id=right \
|
||||
|
||||
+10
-10
@@ -45,7 +45,7 @@ Note that the `id` associated with a robot is used to store the calibration file
|
||||
<hfoptions id="teleoperate_so101">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.id=my_awesome_follower_arm \
|
||||
@@ -101,7 +101,7 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
|
||||
<hfoptions id="teleoperate_koch_camera">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=koch_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.id=my_awesome_follower_arm \
|
||||
@@ -174,7 +174,7 @@ Now you can record a dataset. To record 5 episodes and upload your dataset to th
|
||||
<hfoptions id="record">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem585A0076841 \
|
||||
--robot.id=my_awesome_follower_arm \
|
||||
@@ -294,7 +294,7 @@ dataset.push_to_hub()
|
||||
|
||||
#### Dataset upload
|
||||
|
||||
Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. `https://huggingface.co/datasets/${HF_USER}/so101_test`) that you can obtain by running:
|
||||
Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
|
||||
|
||||
```bash
|
||||
echo https://huggingface.co/datasets/${HF_USER}/so101_test
|
||||
@@ -376,7 +376,7 @@ You can replay the first episode on your robot with either the command below or
|
||||
<hfoptions id="replay">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-replay \
|
||||
python -m lerobot.replay \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.id=my_awesome_follower_arm \
|
||||
@@ -428,10 +428,10 @@ Your robot should replicate movements similar to those you recorded. For example
|
||||
|
||||
## Train a policy
|
||||
|
||||
To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=${HF_USER}/so101_test \
|
||||
--policy.type=act \
|
||||
--output_dir=outputs/train/act_so101_test \
|
||||
@@ -444,7 +444,7 @@ lerobot-train \
|
||||
Let's explain the command:
|
||||
|
||||
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
|
||||
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
|
||||
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
|
||||
3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
|
||||
4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
|
||||
|
||||
@@ -453,7 +453,7 @@ Training should take several hours. You will find checkpoints in `outputs/train/
|
||||
To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
|
||||
--resume=true
|
||||
```
|
||||
@@ -490,7 +490,7 @@ You can use the `record` script from [`lerobot/record.py`](https://github.com/hu
|
||||
<hfoptions id="eval">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/ttyACM1 \
|
||||
--robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
|
||||
|
||||
@@ -96,10 +96,10 @@ If you uploaded your dataset to the hub you can [visualize your dataset online](
|
||||
|
||||
## Train a policy
|
||||
|
||||
To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=${HF_USER}/il_gym \
|
||||
--policy.type=act \
|
||||
--output_dir=outputs/train/il_sim_test \
|
||||
@@ -111,7 +111,7 @@ lerobot-train \
|
||||
Let's explain the command:
|
||||
|
||||
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/il_gym`.
|
||||
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
|
||||
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
|
||||
3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
|
||||
4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
|
||||
|
||||
|
||||
@@ -1,6 +1,15 @@
|
||||
# Installation
|
||||
|
||||
## Environment Setup
|
||||
## Install LeRobot
|
||||
|
||||
Currently only available from source.
|
||||
|
||||
Download our source code:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/lerobot.git
|
||||
cd lerobot
|
||||
```
|
||||
|
||||
Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
|
||||
|
||||
@@ -31,49 +40,12 @@ conda install ffmpeg -c conda-forge
|
||||
>
|
||||
> - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
|
||||
|
||||
## Install LeRobot 🤗
|
||||
|
||||
### From Source
|
||||
|
||||
First, clone the repository and navigate into the directory:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/lerobot.git
|
||||
cd lerobot
|
||||
```
|
||||
|
||||
Then, install the library in editable mode. This is useful if you plan to contribute to the code.
|
||||
Install 🤗 LeRobot:
|
||||
|
||||
```bash
|
||||
pip install -e .
|
||||
```
|
||||
|
||||
### Installation from PyPI
|
||||
|
||||
**Core Library:**
|
||||
Install the base package with:
|
||||
|
||||
```bash
|
||||
pip install lerobot
|
||||
```
|
||||
|
||||
_This installs only the default dependencies._
|
||||
|
||||
**Extra Features:**
|
||||
To install additional functionality, use one of the following:
|
||||
|
||||
```bash
|
||||
pip install 'lerobot[all]' # All available features
|
||||
pip install 'lerobot[aloha,pusht]' # Specific features (Aloha & Pusht)
|
||||
pip install 'lerobot[feetech]' # Feetech motor support
|
||||
```
|
||||
|
||||
_Replace `[...]` with your desired features._
|
||||
|
||||
**Available Tags:**
|
||||
For a full list of optional dependencies, see:
|
||||
https://pypi.org/project/lerobot/
|
||||
|
||||
### Troubleshooting
|
||||
|
||||
If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
|
||||
|
||||
@@ -31,7 +31,7 @@ pip install -e ".[dynamixel]"
|
||||
To find the port for each bus servo adapter, run this script:
|
||||
|
||||
```bash
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
|
||||
<hfoptions id="example">
|
||||
@@ -98,7 +98,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--robot.type=koch_follower \
|
||||
--robot.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -174,7 +174,7 @@ Do the same steps for the leader arm but modify the command or script accordingl
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--teleop.type=koch_leader \
|
||||
--teleop.port=/dev/tty.usbmodem575E0031751 \ # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -211,7 +211,7 @@ Run the following command or API example to calibrate the follower arm:
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=koch_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
|
||||
@@ -249,7 +249,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=koch_leader \
|
||||
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
|
||||
|
||||
@@ -60,7 +60,7 @@ First, we will assemble the two SO100/SO101 arms. One to attach to the mobile ba
|
||||
To find the port for each bus servo adapter, run this script:
|
||||
|
||||
```bash
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
|
||||
<hfoptions id="example">
|
||||
@@ -116,7 +116,7 @@ The instructions for configuring the motors can be found in the SO101 [docs](./s
|
||||
You can run this command to setup motors for LeKiwi. It will first setup the motors for arm (id 6..1) and then setup motors for wheels (9,8,7)
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--robot.type=lekiwi \
|
||||
--robot.port=/dev/tty.usbmodem58760431551 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -174,7 +174,7 @@ The calibration process is very important because it allows a neural network tra
|
||||
Make sure the arm is connected to the Raspberry Pi and run this script or API example (on the Raspberry Pi via SSH) to launch calibration of the follower arm:
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=lekiwi \
|
||||
--robot.id=my_awesome_kiwi # <- Give the robot a unique name
|
||||
```
|
||||
@@ -193,7 +193,7 @@ Then, to calibrate the leader arm (which is attached to the laptop/pc). Run the
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=so100_leader \
|
||||
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
|
||||
|
||||
@@ -54,7 +54,7 @@ If you don't have a gpu device, you can train using our notebook on [.
|
||||
|
||||
```bash
|
||||
cd lerobot && lerobot-train \
|
||||
cd lerobot && python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/smolvla_base \
|
||||
--dataset.repo_id=${HF_USER}/mydataset \
|
||||
--batch_size=64 \
|
||||
@@ -73,7 +73,7 @@ cd lerobot && lerobot-train \
|
||||
Fine-tuning is an art. For a complete overview of the options for finetuning, run
|
||||
|
||||
```bash
|
||||
lerobot-train --help
|
||||
python -m lerobot.scripts.train --help
|
||||
```
|
||||
|
||||
<p align="center">
|
||||
@@ -97,7 +97,7 @@ Similarly for when recording an episode, it is recommended that you are logged i
|
||||
Once you are logged in, you can run inference in your setup by doing:
|
||||
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/ttyACM0 \ # <- Use your port
|
||||
--robot.id=my_blue_follower_arm \ # <- Use your robot id
|
||||
|
||||
@@ -26,7 +26,7 @@ Unlike the SO-101, the motor connectors are not easily accessible once the arm i
|
||||
To find the port for each bus servo adapter, run this script:
|
||||
|
||||
```bash
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
|
||||
<hfoptions id="example">
|
||||
@@ -93,7 +93,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/tty.usbmodem585A0076841 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -168,7 +168,7 @@ Do the same steps for the leader arm.
|
||||
<hfoptions id="setup_motors">
|
||||
<hfoption id="Command">
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--teleop.type=so100_leader \
|
||||
--teleop.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -568,7 +568,7 @@ Run the following command or API example to calibrate the follower arm:
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
|
||||
@@ -606,7 +606,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=so100_leader \
|
||||
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
|
||||
|
||||
@@ -162,7 +162,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
|
||||
To find the port for each bus servo adapter, connect MotorBus to your computer via USB and power. Run the following script and disconnect the MotorBus when prompted:
|
||||
|
||||
```bash
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
|
||||
<hfoptions id="example">
|
||||
@@ -240,7 +240,7 @@ Connect the usb cable from your computer and the power supply to the follower ar
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem585A0076841 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -316,7 +316,7 @@ Do the same steps for the leader arm.
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--teleop.type=so101_leader \
|
||||
--teleop.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
|
||||
```
|
||||
@@ -353,7 +353,7 @@ Run the following command or API example to calibrate the follower arm:
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
|
||||
@@ -402,7 +402,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
|
||||
<hfoption id="Command">
|
||||
|
||||
```bash
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=so101_leader \
|
||||
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
|
||||
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
|
||||
|
||||
@@ -62,7 +62,7 @@ By default, every field takes its default value specified in the dataclass. If a
|
||||
Let's say that we want to train [Diffusion Policy](../src/lerobot/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=lerobot/pusht \
|
||||
--policy.type=diffusion \
|
||||
--env.type=pusht
|
||||
@@ -77,7 +77,7 @@ Let's break this down:
|
||||
Let's see another example. Let's say you've been training [ACT](../src/lerobot/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=act \
|
||||
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
|
||||
--env.type=aloha \
|
||||
@@ -90,7 +90,7 @@ We now want to train a different policy for aloha on another task. We'll change
|
||||
Looking at the [`AlohaEnv`](../src/lerobot/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=act \
|
||||
--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
|
||||
--env.type=aloha \
|
||||
@@ -127,7 +127,7 @@ Now, let's assume that we want to reproduce the run just above. That run has pro
|
||||
We can then simply load the config values from this file using:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
|
||||
--output_dir=outputs/train/act_aloha_transfer_2
|
||||
```
|
||||
@@ -137,7 +137,7 @@ lerobot-train \
|
||||
Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
|
||||
--output_dir=outputs/train/act_aloha_transfer_2
|
||||
--policy.n_action_steps=80
|
||||
@@ -148,7 +148,7 @@ lerobot-train \
|
||||
`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
|
||||
|
||||
```bash
|
||||
lerobot-train --config_path=lerobot/diffusion_pusht
|
||||
python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
|
||||
```
|
||||
|
||||
will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
|
||||
@@ -160,7 +160,7 @@ Being able to resume a training run is important in case it crashed or aborted f
|
||||
Let's reuse the command from the previous run and add a few more options:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=act \
|
||||
--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
|
||||
--env.type=aloha \
|
||||
@@ -179,7 +179,7 @@ INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
|
||||
Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
|
||||
--resume=true
|
||||
```
|
||||
@@ -190,7 +190,7 @@ Another reason for which you might want to resume a run is simply to extend trai
|
||||
You could double the number of steps of the previous run with:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
|
||||
--resume=true \
|
||||
--steps=200000
|
||||
@@ -224,7 +224,7 @@ In addition to the features currently in Draccus, we've added a special `.path`
|
||||
For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/act_aloha_sim_transfer_cube_human \
|
||||
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
|
||||
--env.type=aloha \
|
||||
@@ -270,7 +270,7 @@ We'll summarize here the main use cases to remember from this tutorial.
|
||||
#### Train a policy from scratch – CLI
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=act \ # <- select 'act' policy
|
||||
--env.type=pusht \ # <- select 'pusht' environment
|
||||
--dataset.repo_id=lerobot/pusht # <- train on this dataset
|
||||
@@ -279,7 +279,7 @@ lerobot-train \
|
||||
#### Train a policy from scratch - config file + CLI
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=path/to/pretrained_model \ # <- can also be a repo_id
|
||||
--policy.n_action_steps=80 # <- you may still override values
|
||||
```
|
||||
@@ -287,7 +287,7 @@ lerobot-train \
|
||||
#### Resume/continue a training run
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--config_path=checkpoint/pretrained_model/ \
|
||||
--resume=true \
|
||||
--steps=200000 # <- you can change some training parameters
|
||||
@@ -296,7 +296,7 @@ lerobot-train \
|
||||
#### Fine-tuning
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/act_aloha_sim_transfer_cube_human \ # <- can also be a local path to a checkpoint
|
||||
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
|
||||
--env.type=aloha \
|
||||
|
||||
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-replay \
|
||||
python -m lerobot.replay \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.id=black \
|
||||
|
||||
+4
-5
@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
|
||||
|
||||
[project]
|
||||
name = "lerobot"
|
||||
version = "0.3.4"
|
||||
version = "0.3.2"
|
||||
description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
|
||||
readme = "README.md"
|
||||
license = { text = "Apache-2.0" }
|
||||
@@ -68,16 +68,15 @@ dependencies = [
|
||||
"einops>=0.8.0",
|
||||
"opencv-python-headless>=4.9.0",
|
||||
"av>=14.2.0",
|
||||
"torch>=2.2.1",
|
||||
"torchcodec>=0.2.1; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
|
||||
"torchvision>=0.21.0",
|
||||
"jsonlines>=4.0.0",
|
||||
"packaging>=24.2",
|
||||
"pynput>=1.7.7",
|
||||
"pyserial>=3.5",
|
||||
"wandb>=0.20.0",
|
||||
|
||||
"torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
|
||||
"torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
|
||||
"torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
|
||||
|
||||
"draccus==0.10.0", # TODO: Remove ==
|
||||
"gymnasium>=0.29.1,<1.0.0", # TODO: Bumb dependency
|
||||
"rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency
|
||||
|
||||
@@ -18,7 +18,7 @@ Helper to recalibrate your device (robot or teleoperator).
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-calibrate \
|
||||
python -m lerobot.calibrate \
|
||||
--teleop.type=so100_leader \
|
||||
--teleop.port=/dev/tty.usbmodem58760431551 \
|
||||
--teleop.id=blue
|
||||
|
||||
@@ -60,7 +60,7 @@ class OpenCVCamera(Camera):
|
||||
or port changes, especially on Linux. Use the provided utility script to find
|
||||
available camera indices or paths:
|
||||
```bash
|
||||
lerobot-find-cameras opencv
|
||||
python -m lerobot.find_cameras opencv
|
||||
```
|
||||
|
||||
The camera's default settings (FPS, resolution, color mode) are used unless
|
||||
@@ -165,7 +165,8 @@ class OpenCVCamera(Camera):
|
||||
self.videocapture.release()
|
||||
self.videocapture = None
|
||||
raise ConnectionError(
|
||||
f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
|
||||
f"Failed to open {self}."
|
||||
f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
|
||||
)
|
||||
|
||||
self._configure_capture_settings()
|
||||
|
||||
@@ -51,7 +51,7 @@ class RealSenseCamera(Camera):
|
||||
|
||||
Use the provided utility script to find available camera indices and default profiles:
|
||||
```bash
|
||||
lerobot-find-cameras realsense
|
||||
python -m lerobot.find_cameras realsense
|
||||
```
|
||||
|
||||
A `RealSenseCamera` instance requires a configuration object specifying the
|
||||
@@ -176,7 +176,8 @@ class RealSenseCamera(Camera):
|
||||
self.rs_profile = None
|
||||
self.rs_pipeline = None
|
||||
raise ConnectionError(
|
||||
f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
|
||||
f"Failed to open {self}."
|
||||
"Run `python -m lerobot.find_cameras realsense` to find available cameras."
|
||||
) from e
|
||||
|
||||
self._configure_capture_settings()
|
||||
|
||||
@@ -27,7 +27,6 @@ from huggingface_hub.constants import CONFIG_NAME
|
||||
from huggingface_hub.errors import HfHubHTTPError
|
||||
|
||||
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
|
||||
from lerobot.constants import ACTION, OBS_STATE
|
||||
from lerobot.optim.optimizers import OptimizerConfig
|
||||
from lerobot.optim.schedulers import LRSchedulerConfig
|
||||
from lerobot.utils.hub import HubMixin
|
||||
@@ -120,8 +119,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
|
||||
|
||||
@property
|
||||
def robot_state_feature(self) -> PolicyFeature | None:
|
||||
for ft_name, ft in self.input_features.items():
|
||||
if ft.type is FeatureType.STATE and ft_name == OBS_STATE:
|
||||
for _, ft in self.input_features.items():
|
||||
if ft.type is FeatureType.STATE:
|
||||
return ft
|
||||
return None
|
||||
|
||||
@@ -138,8 +137,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
|
||||
|
||||
@property
|
||||
def action_feature(self) -> PolicyFeature | None:
|
||||
for ft_name, ft in self.output_features.items():
|
||||
if ft.type is FeatureType.ACTION and ft_name == ACTION:
|
||||
for _, ft in self.output_features.items():
|
||||
if ft.type is FeatureType.ACTION:
|
||||
return ft
|
||||
return None
|
||||
|
||||
|
||||
@@ -825,8 +825,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
|
||||
"""
|
||||
if not episode_data:
|
||||
episode_buffer = self.episode_buffer
|
||||
else:
|
||||
episode_buffer = episode_data
|
||||
|
||||
validate_episode_buffer(episode_buffer, self.meta.total_episodes, self.features)
|
||||
|
||||
|
||||
@@ -20,7 +20,7 @@ Helper to find the camera devices available in your system.
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-find-cameras
|
||||
python -m lerobot.find_cameras
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
@@ -18,7 +18,7 @@ Helper to find the USB port associated with your MotorsBus.
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-find-port
|
||||
python -m lerobot.find_port
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
@@ -107,8 +107,6 @@ X_SERIES_ENCODINGS_TABLE = {
|
||||
"Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
|
||||
"Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
|
||||
"Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
|
||||
"Goal_Position": X_SERIES_CONTROL_TABLE["Goal_Position"][1],
|
||||
"Present_Position": X_SERIES_CONTROL_TABLE["Present_Position"][1],
|
||||
"Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
|
||||
"Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
|
||||
"Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
|
||||
|
||||
@@ -222,7 +222,7 @@ class MotorsBus(abc.ABC):
|
||||
A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
|
||||
To find the port, you can run our utility script:
|
||||
```bash
|
||||
lerobot-find-port.py
|
||||
python -m lerobot.find_port.py
|
||||
>>> Finding all available ports for the MotorsBus.
|
||||
>>> ["/dev/tty.usbmodem575E0032081", "/dev/tty.usbmodem575E0031751"]
|
||||
>>> Remove the usb cable from your MotorsBus and press Enter when done.
|
||||
@@ -446,7 +446,7 @@ class MotorsBus(abc.ABC):
|
||||
except (FileNotFoundError, OSError, serial.SerialException) as e:
|
||||
raise ConnectionError(
|
||||
f"\nCould not connect on port '{self.port}'. Make sure you are using the correct port."
|
||||
"\nTry running `lerobot-find-port`\n"
|
||||
"\nTry running `python -m lerobot.find_port`\n"
|
||||
) from e
|
||||
|
||||
@abc.abstractmethod
|
||||
|
||||
@@ -24,6 +24,7 @@ def create_stats_buffers(
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
) -> dict[str, dict[str, nn.ParameterDict]]:
|
||||
"""
|
||||
Create buffers per modality (e.g. "observation.image", "action") containing their mean, std, min, max
|
||||
@@ -60,8 +61,8 @@ def create_stats_buffers(
|
||||
|
||||
buffer = {}
|
||||
if norm_mode is NormalizationMode.MEAN_STD:
|
||||
mean = torch.ones(shape, dtype=torch.float32) * torch.inf
|
||||
std = torch.ones(shape, dtype=torch.float32) * torch.inf
|
||||
mean = torch.ones(shape, dtype=dtype) * torch.inf
|
||||
std = torch.ones(shape, dtype=dtype) * torch.inf
|
||||
buffer = nn.ParameterDict(
|
||||
{
|
||||
"mean": nn.Parameter(mean, requires_grad=False),
|
||||
@@ -69,8 +70,8 @@ def create_stats_buffers(
|
||||
}
|
||||
)
|
||||
elif norm_mode is NormalizationMode.MIN_MAX:
|
||||
min = torch.ones(shape, dtype=torch.float32) * torch.inf
|
||||
max = torch.ones(shape, dtype=torch.float32) * torch.inf
|
||||
min = torch.ones(shape, dtype=dtype) * torch.inf
|
||||
max = torch.ones(shape, dtype=dtype) * torch.inf
|
||||
buffer = nn.ParameterDict(
|
||||
{
|
||||
"min": nn.Parameter(min, requires_grad=False),
|
||||
@@ -82,22 +83,22 @@ def create_stats_buffers(
|
||||
if stats:
|
||||
if isinstance(stats[key]["mean"], np.ndarray):
|
||||
if norm_mode is NormalizationMode.MEAN_STD:
|
||||
buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
|
||||
buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
|
||||
buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=dtype)
|
||||
buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=dtype)
|
||||
elif norm_mode is NormalizationMode.MIN_MAX:
|
||||
buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
|
||||
buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
|
||||
buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=dtype)
|
||||
buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=dtype)
|
||||
elif isinstance(stats[key]["mean"], torch.Tensor):
|
||||
# Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
|
||||
# tensors anywhere (for example, when we use the same stats for normalization and
|
||||
# unnormalization). See the logic here
|
||||
# https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
|
||||
if norm_mode is NormalizationMode.MEAN_STD:
|
||||
buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
|
||||
buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
|
||||
buffer["mean"].data = stats[key]["mean"].clone().to(dtype=dtype)
|
||||
buffer["std"].data = stats[key]["std"].clone().to(dtype=dtype)
|
||||
elif norm_mode is NormalizationMode.MIN_MAX:
|
||||
buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
|
||||
buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
|
||||
buffer["min"].data = stats[key]["min"].clone().to(dtype=dtype)
|
||||
buffer["max"].data = stats[key]["max"].clone().to(dtype=dtype)
|
||||
else:
|
||||
type_ = type(stats[key]["mean"])
|
||||
raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
|
||||
@@ -121,6 +122,7 @@ class Normalize(nn.Module):
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
):
|
||||
"""
|
||||
Args:
|
||||
@@ -144,7 +146,7 @@ class Normalize(nn.Module):
|
||||
self.features = features
|
||||
self.norm_map = norm_map
|
||||
self.stats = stats
|
||||
stats_buffers = create_stats_buffers(features, norm_map, stats)
|
||||
stats_buffers = create_stats_buffers(features, norm_map, stats, dtype)
|
||||
for key, buffer in stats_buffers.items():
|
||||
setattr(self, "buffer_" + key.replace(".", "_"), buffer)
|
||||
|
||||
@@ -195,6 +197,7 @@ class Unnormalize(nn.Module):
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
):
|
||||
"""
|
||||
Args:
|
||||
@@ -219,7 +222,7 @@ class Unnormalize(nn.Module):
|
||||
self.norm_map = norm_map
|
||||
self.stats = stats
|
||||
# `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
|
||||
stats_buffers = create_stats_buffers(features, norm_map, stats)
|
||||
stats_buffers = create_stats_buffers(features, norm_map, stats, dtype)
|
||||
for key, buffer in stats_buffers.items():
|
||||
setattr(self, "buffer_" + key.replace(".", "_"), buffer)
|
||||
|
||||
@@ -262,6 +265,7 @@ def _initialize_stats_buffers(
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
) -> None:
|
||||
"""Register statistics buffers (mean/std or min/max) on the given *module*.
|
||||
|
||||
@@ -282,8 +286,8 @@ def _initialize_stats_buffers(
|
||||
prefix = key.replace(".", "_")
|
||||
|
||||
if norm_mode is NormalizationMode.MEAN_STD:
|
||||
mean = torch.full(shape, torch.inf, dtype=torch.float32)
|
||||
std = torch.full(shape, torch.inf, dtype=torch.float32)
|
||||
mean = torch.full(shape, torch.inf, dtype=dtype)
|
||||
std = torch.full(shape, torch.inf, dtype=dtype)
|
||||
|
||||
if stats and key in stats and "mean" in stats[key] and "std" in stats[key]:
|
||||
mean_data = stats[key]["mean"]
|
||||
@@ -293,8 +297,8 @@ def _initialize_stats_buffers(
|
||||
# tensors anywhere (for example, when we use the same stats for normalization and
|
||||
# unnormalization). See the logic here
|
||||
# https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
|
||||
mean = mean_data.clone().to(dtype=torch.float32)
|
||||
std = std_data.clone().to(dtype=torch.float32)
|
||||
mean = mean_data.clone().to(dtype=dtype)
|
||||
std = std_data.clone().to(dtype=dtype)
|
||||
else:
|
||||
raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
|
||||
|
||||
@@ -303,15 +307,15 @@ def _initialize_stats_buffers(
|
||||
continue
|
||||
|
||||
if norm_mode is NormalizationMode.MIN_MAX:
|
||||
min_val = torch.full(shape, torch.inf, dtype=torch.float32)
|
||||
max_val = torch.full(shape, torch.inf, dtype=torch.float32)
|
||||
min_val = torch.full(shape, torch.inf, dtype=dtype)
|
||||
max_val = torch.full(shape, torch.inf, dtype=dtype)
|
||||
|
||||
if stats and key in stats and "min" in stats[key] and "max" in stats[key]:
|
||||
min_data = stats[key]["min"]
|
||||
max_data = stats[key]["max"]
|
||||
if isinstance(min_data, torch.Tensor):
|
||||
min_val = min_data.clone().to(dtype=torch.float32)
|
||||
max_val = max_data.clone().to(dtype=torch.float32)
|
||||
min_val = min_data.clone().to(dtype=dtype)
|
||||
max_val = max_data.clone().to(dtype=dtype)
|
||||
else:
|
||||
raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
|
||||
|
||||
@@ -330,12 +334,13 @@ class NormalizeBuffer(nn.Module):
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
):
|
||||
super().__init__()
|
||||
self.features = features
|
||||
self.norm_map = norm_map
|
||||
|
||||
_initialize_stats_buffers(self, features, norm_map, stats)
|
||||
_initialize_stats_buffers(self, features, norm_map, stats, dtype)
|
||||
|
||||
def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
|
||||
batch = dict(batch)
|
||||
@@ -379,12 +384,13 @@ class UnnormalizeBuffer(nn.Module):
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[str, NormalizationMode],
|
||||
stats: dict[str, dict[str, Tensor]] | None = None,
|
||||
dtype: torch.dtype = torch.float32,
|
||||
):
|
||||
super().__init__()
|
||||
self.features = features
|
||||
self.norm_map = norm_map
|
||||
|
||||
_initialize_stats_buffers(self, features, norm_map, stats)
|
||||
_initialize_stats_buffers(self, features, norm_map, stats, dtype)
|
||||
|
||||
def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
|
||||
# batch = dict(batch)
|
||||
|
||||
@@ -30,7 +30,7 @@ pip install -e ".[pi0]"
|
||||
|
||||
Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/pi0 \
|
||||
--dataset.repo_id=danaaubakirova/koch_test
|
||||
```
|
||||
@@ -38,7 +38,7 @@ lerobot-train \
|
||||
Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
|
||||
pretrained with VLM default parameters before pi0 finetuning:
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=pi0 \
|
||||
--dataset.repo_id=danaaubakirova/koch_test
|
||||
```
|
||||
|
||||
@@ -25,14 +25,14 @@ Disclaimer: It is not expected to perform as well as the original implementation
|
||||
|
||||
Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/pi0fast_base \
|
||||
--dataset.repo_id=danaaubakirova/koch_test
|
||||
```
|
||||
|
||||
Example of training the pi0+FAST neural network with from scratch:
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=pi0fast \
|
||||
--dataset.repo_id=danaaubakirova/koch_test
|
||||
```
|
||||
|
||||
@@ -28,7 +28,7 @@ pip install -e ".[smolvla]"
|
||||
|
||||
Example of finetuning the smolvla pretrained model (`smolvla_base`):
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/smolvla_base \
|
||||
--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
|
||||
--batch_size=64 \
|
||||
@@ -38,7 +38,7 @@ lerobot-train \
|
||||
Example of finetuning a smolVLA. SmolVLA is composed of a pretrained VLM,
|
||||
and an action expert.
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.type=smolvla \
|
||||
--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
|
||||
--batch_size=64 \
|
||||
@@ -673,19 +673,19 @@ class VLAFlowMatching(nn.Module):
|
||||
for params in self.state_proj.parameters():
|
||||
params.requires_grad = self.config.train_state_proj
|
||||
|
||||
def sample_noise(self, shape, device):
|
||||
def sample_noise(self, shape, device, dtype):
|
||||
noise = torch.normal(
|
||||
mean=0.0,
|
||||
std=1.0,
|
||||
size=shape,
|
||||
dtype=torch.float32,
|
||||
dtype=dtype,
|
||||
device=device,
|
||||
)
|
||||
return noise
|
||||
|
||||
def sample_time(self, bsize, device):
|
||||
def sample_time(self, bsize, device, dtype):
|
||||
beta_dist = torch.distributions.Beta(concentration1=1.5, concentration0=1.0)
|
||||
time_beta = beta_dist.sample((bsize,)).to(device=device, dtype=torch.float32)
|
||||
time_beta = beta_dist.sample((bsize,)).to(device=device, dtype=dtype)
|
||||
time = time_beta * 0.999 + 0.001
|
||||
return time
|
||||
|
||||
@@ -831,10 +831,10 @@ class VLAFlowMatching(nn.Module):
|
||||
) -> Tensor:
|
||||
"""Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
|
||||
if noise is None:
|
||||
noise = self.sample_noise(actions.shape, actions.device)
|
||||
noise = self.sample_noise(actions.shape, actions.device, actions.dtype)
|
||||
|
||||
if time is None:
|
||||
time = self.sample_time(actions.shape[0], actions.device)
|
||||
time = self.sample_time(actions.shape[0], actions.device, actions.dtype)
|
||||
|
||||
time_expanded = time[:, None, None]
|
||||
x_t = time_expanded * noise + (1 - time_expanded) * actions
|
||||
@@ -868,10 +868,11 @@ class VLAFlowMatching(nn.Module):
|
||||
"""Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
|
||||
bsize = state.shape[0]
|
||||
device = state.device
|
||||
dtype = state.dtype
|
||||
|
||||
if noise is None:
|
||||
actions_shape = (bsize, self.config.chunk_size, self.config.max_action_dim)
|
||||
noise = self.sample_noise(actions_shape, device)
|
||||
noise = self.sample_noise(actions_shape, device, dtype)
|
||||
|
||||
prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
|
||||
images, img_masks, lang_tokens, lang_masks, state=state
|
||||
@@ -888,18 +889,13 @@ class VLAFlowMatching(nn.Module):
|
||||
fill_kv_cache=True,
|
||||
)
|
||||
dt = -1.0 / self.config.num_steps
|
||||
dt = torch.tensor(dt, dtype=torch.float32, device=device)
|
||||
dt = torch.tensor(dt, dtype=dtype, device=device)
|
||||
|
||||
x_t = noise
|
||||
time = torch.tensor(1.0, dtype=torch.float32, device=device)
|
||||
time = torch.tensor(1.0, dtype=dtype, device=device)
|
||||
while time >= -dt / 2:
|
||||
expanded_time = time.expand(bsize)
|
||||
v_t = self.denoise_step(
|
||||
prefix_pad_masks,
|
||||
past_key_values,
|
||||
x_t,
|
||||
expanded_time,
|
||||
)
|
||||
v_t = self.denoise_step(prefix_pad_masks, past_key_values, x_t, expanded_time, dtype)
|
||||
# Euler step
|
||||
x_t += dt * v_t
|
||||
time += dt
|
||||
@@ -911,6 +907,7 @@ class VLAFlowMatching(nn.Module):
|
||||
past_key_values,
|
||||
x_t,
|
||||
timestep,
|
||||
dtype,
|
||||
):
|
||||
"""Apply one denoising step of the noise `x_t` at a given timestep."""
|
||||
suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(x_t, timestep)
|
||||
@@ -936,6 +933,6 @@ class VLAFlowMatching(nn.Module):
|
||||
)
|
||||
suffix_out = outputs_embeds[1]
|
||||
suffix_out = suffix_out[:, -self.config.chunk_size :]
|
||||
suffix_out = suffix_out.to(dtype=torch.float32)
|
||||
suffix_out = suffix_out.to(dtype=dtype)
|
||||
v_t = self.action_out_proj(suffix_out)
|
||||
return v_t
|
||||
|
||||
@@ -1,54 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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 .device_processor import DeviceProcessor
|
||||
from .normalize_processor import NormalizerProcessor, UnnormalizerProcessor
|
||||
from .observation_processor import VanillaObservationProcessor
|
||||
from .pipeline import (
|
||||
ActionProcessor,
|
||||
DoneProcessor,
|
||||
EnvTransition,
|
||||
IdentityProcessor,
|
||||
InfoProcessor,
|
||||
ObservationProcessor,
|
||||
ProcessorStep,
|
||||
ProcessorStepRegistry,
|
||||
RewardProcessor,
|
||||
RobotProcessor,
|
||||
TransitionKey,
|
||||
TruncatedProcessor,
|
||||
)
|
||||
from .rename_processor import RenameProcessor
|
||||
|
||||
__all__ = [
|
||||
"ActionProcessor",
|
||||
"DeviceProcessor",
|
||||
"DoneProcessor",
|
||||
"EnvTransition",
|
||||
"IdentityProcessor",
|
||||
"InfoProcessor",
|
||||
"NormalizerProcessor",
|
||||
"UnnormalizerProcessor",
|
||||
"ObservationProcessor",
|
||||
"ProcessorStep",
|
||||
"ProcessorStepRegistry",
|
||||
"RenameProcessor",
|
||||
"RewardProcessor",
|
||||
"RobotProcessor",
|
||||
"TransitionKey",
|
||||
"TruncatedProcessor",
|
||||
"VanillaObservationProcessor",
|
||||
]
|
||||
@@ -1,82 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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 dataclasses import dataclass
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import PolicyFeature
|
||||
from lerobot.processor.pipeline import EnvTransition, TransitionKey
|
||||
from lerobot.utils.utils import get_safe_torch_device
|
||||
|
||||
|
||||
@dataclass
|
||||
class DeviceProcessor:
|
||||
"""Processes transitions by moving tensors to the specified device.
|
||||
|
||||
This processor ensures that all tensors in the transition are moved to the
|
||||
specified device (CPU or GPU) before they are returned.
|
||||
"""
|
||||
|
||||
device: torch.device = "cpu"
|
||||
|
||||
def __post_init__(self):
|
||||
self.device = get_safe_torch_device(self.device)
|
||||
self.non_blocking = "cuda" in str(self.device)
|
||||
|
||||
def __call__(self, transition: EnvTransition) -> EnvTransition:
|
||||
# Create a copy of the transition
|
||||
new_transition = transition.copy()
|
||||
|
||||
# Process observation tensors
|
||||
observation = transition.get(TransitionKey.OBSERVATION)
|
||||
if observation is not None:
|
||||
new_observation = {
|
||||
k: v.to(self.device, non_blocking=self.non_blocking) if isinstance(v, torch.Tensor) else v
|
||||
for k, v in observation.items()
|
||||
}
|
||||
new_transition[TransitionKey.OBSERVATION] = new_observation
|
||||
|
||||
# Process action tensor
|
||||
action = transition.get(TransitionKey.ACTION)
|
||||
if action is not None and isinstance(action, torch.Tensor):
|
||||
new_transition[TransitionKey.ACTION] = action.to(self.device, non_blocking=self.non_blocking)
|
||||
|
||||
# Process reward tensor
|
||||
reward = transition.get(TransitionKey.REWARD)
|
||||
if reward is not None and isinstance(reward, torch.Tensor):
|
||||
new_transition[TransitionKey.REWARD] = reward.to(self.device, non_blocking=self.non_blocking)
|
||||
|
||||
# Process done tensor
|
||||
done = transition.get(TransitionKey.DONE)
|
||||
if done is not None and isinstance(done, torch.Tensor):
|
||||
new_transition[TransitionKey.DONE] = done.to(self.device, non_blocking=self.non_blocking)
|
||||
|
||||
# Process truncated tensor
|
||||
truncated = transition.get(TransitionKey.TRUNCATED)
|
||||
if truncated is not None and isinstance(truncated, torch.Tensor):
|
||||
new_transition[TransitionKey.TRUNCATED] = truncated.to(
|
||||
self.device, non_blocking=self.non_blocking
|
||||
)
|
||||
|
||||
return new_transition
|
||||
|
||||
def get_config(self) -> dict[str, Any]:
|
||||
"""Return configuration for serialization."""
|
||||
return {"device": self.device}
|
||||
|
||||
def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
|
||||
return features
|
||||
@@ -1,331 +0,0 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from collections.abc import Mapping
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Any
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch import Tensor
|
||||
|
||||
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDataset
|
||||
from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
|
||||
|
||||
|
||||
def _convert_stats_to_tensors(stats: dict[str, dict[str, Any]]) -> dict[str, dict[str, Tensor]]:
|
||||
"""Convert numpy arrays and other types to torch tensors."""
|
||||
tensor_stats: dict[str, dict[str, Tensor]] = {}
|
||||
for key, sub in stats.items():
|
||||
tensor_stats[key] = {}
|
||||
for stat_name, value in sub.items():
|
||||
if isinstance(value, np.ndarray):
|
||||
tensor_val = torch.from_numpy(value.astype(np.float32))
|
||||
elif isinstance(value, torch.Tensor):
|
||||
tensor_val = value.to(dtype=torch.float32)
|
||||
elif isinstance(value, (int, float, list, tuple)):
|
||||
tensor_val = torch.tensor(value, dtype=torch.float32)
|
||||
else:
|
||||
raise TypeError(f"Unsupported type for stats['{key}']['{stat_name}']: {type(value)}")
|
||||
tensor_stats[key][stat_name] = tensor_val
|
||||
return tensor_stats
|
||||
|
||||
|
||||
@dataclass
|
||||
@ProcessorStepRegistry.register(name="normalizer_processor")
|
||||
class NormalizerProcessor:
|
||||
"""Normalizes observations and actions in a single processor step.
|
||||
|
||||
This processor handles normalization of both observation and action tensors
|
||||
using either mean/std normalization or min/max scaling to a [-1, 1] range.
|
||||
|
||||
For each tensor key in the stats dictionary, the processor will:
|
||||
- Use mean/std normalization if those statistics are provided: (x - mean) / std
|
||||
- Use min/max scaling if those statistics are provided: 2 * (x - min) / (max - min) - 1
|
||||
|
||||
The processor can be configured to normalize only specific keys by setting
|
||||
the normalize_keys parameter.
|
||||
"""
|
||||
|
||||
# Features and normalisation map are mandatory to match the design of normalize.py
|
||||
features: dict[str, PolicyFeature]
|
||||
norm_map: dict[FeatureType, NormalizationMode]
|
||||
|
||||
# Pre-computed statistics coming from dataset.meta.stats for instance.
|
||||
stats: dict[str, dict[str, Any]] | None = None
|
||||
|
||||
# Explicit subset of keys to normalise. If ``None`` every key (except
|
||||
# "action") found in ``stats`` will be normalised. Using a ``set`` makes
|
||||
# membership checks O(1).
|
||||
normalize_keys: set[str] | None = None
|
||||
|
||||
eps: float = 1e-8
|
||||
|
||||
_tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
|
||||
|
||||
@classmethod
|
||||
def from_lerobot_dataset(
|
||||
cls,
|
||||
dataset: LeRobotDataset,
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[FeatureType, NormalizationMode],
|
||||
*,
|
||||
normalize_keys: set[str] | None = None,
|
||||
eps: float = 1e-8,
|
||||
) -> NormalizerProcessor:
|
||||
"""Factory helper that pulls statistics from a :class:`LeRobotDataset`.
|
||||
|
||||
The features and norm_map parameters are mandatory to match the design
|
||||
pattern used in normalize.py.
|
||||
"""
|
||||
|
||||
return cls(
|
||||
features=features,
|
||||
norm_map=norm_map,
|
||||
stats=dataset.meta.stats,
|
||||
normalize_keys=normalize_keys,
|
||||
eps=eps,
|
||||
)
|
||||
|
||||
def __post_init__(self):
|
||||
# Handle deserialization from JSON config
|
||||
if self.features and isinstance(list(self.features.values())[0], dict):
|
||||
# Features came from JSON - need to reconstruct PolicyFeature objects
|
||||
reconstructed_features = {}
|
||||
for key, ft_dict in self.features.items():
|
||||
reconstructed_features[key] = PolicyFeature(
|
||||
type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
|
||||
)
|
||||
self.features = reconstructed_features
|
||||
|
||||
if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
|
||||
# norm_map came from JSON - need to reconstruct enum keys and values
|
||||
reconstructed_norm_map = {}
|
||||
for ft_type_str, norm_mode_str in self.norm_map.items():
|
||||
reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
|
||||
self.norm_map = reconstructed_norm_map
|
||||
|
||||
# Convert statistics once so we avoid repeated numpy→Tensor conversions
|
||||
# during runtime.
|
||||
self.stats = self.stats or {}
|
||||
self._tensor_stats = _convert_stats_to_tensors(self.stats)
|
||||
|
||||
# Ensure *normalize_keys* is a set for fast look-ups and compare by
|
||||
# value later when returning the configuration.
|
||||
if self.normalize_keys is not None and not isinstance(self.normalize_keys, set):
|
||||
self.normalize_keys = set(self.normalize_keys)
|
||||
|
||||
def _normalize_obs(self, observation):
|
||||
if observation is None:
|
||||
return None
|
||||
|
||||
# Decide which keys should be normalised for this call.
|
||||
if self.normalize_keys is not None:
|
||||
keys_to_norm = self.normalize_keys
|
||||
else:
|
||||
# Use feature map to skip action keys.
|
||||
keys_to_norm = {k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION}
|
||||
|
||||
processed = dict(observation)
|
||||
for key in keys_to_norm:
|
||||
if key not in processed or key not in self._tensor_stats:
|
||||
continue
|
||||
|
||||
orig_val = processed[key]
|
||||
tensor = (
|
||||
orig_val.to(dtype=torch.float32)
|
||||
if isinstance(orig_val, torch.Tensor)
|
||||
else torch.as_tensor(orig_val, dtype=torch.float32)
|
||||
)
|
||||
stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
|
||||
|
||||
if "mean" in stats and "std" in stats:
|
||||
mean, std = stats["mean"], stats["std"]
|
||||
processed[key] = (tensor - mean) / (std + self.eps)
|
||||
elif "min" in stats and "max" in stats:
|
||||
min_val, max_val = stats["min"], stats["max"]
|
||||
processed[key] = 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
|
||||
return processed
|
||||
|
||||
def _normalize_action(self, action):
|
||||
if action is None or "action" not in self._tensor_stats:
|
||||
return action
|
||||
|
||||
tensor = (
|
||||
action.to(dtype=torch.float32)
|
||||
if isinstance(action, torch.Tensor)
|
||||
else torch.as_tensor(action, dtype=torch.float32)
|
||||
)
|
||||
stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
|
||||
if "mean" in stats and "std" in stats:
|
||||
mean, std = stats["mean"], stats["std"]
|
||||
return (tensor - mean) / (std + self.eps)
|
||||
if "min" in stats and "max" in stats:
|
||||
min_val, max_val = stats["min"], stats["max"]
|
||||
return 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
|
||||
raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
|
||||
|
||||
def __call__(self, transition: EnvTransition) -> EnvTransition:
|
||||
observation = self._normalize_obs(transition.get(TransitionKey.OBSERVATION))
|
||||
action = self._normalize_action(transition.get(TransitionKey.ACTION))
|
||||
|
||||
# Create a new transition with normalized values
|
||||
new_transition = transition.copy()
|
||||
new_transition[TransitionKey.OBSERVATION] = observation
|
||||
new_transition[TransitionKey.ACTION] = action
|
||||
return new_transition
|
||||
|
||||
def get_config(self) -> dict[str, Any]:
|
||||
config = {
|
||||
"eps": self.eps,
|
||||
"features": {
|
||||
key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
|
||||
},
|
||||
"norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
|
||||
}
|
||||
if self.normalize_keys is not None:
|
||||
# Serialise as a list for YAML / JSON friendliness
|
||||
config["normalize_keys"] = sorted(self.normalize_keys)
|
||||
return config
|
||||
|
||||
def state_dict(self) -> dict[str, Tensor]:
|
||||
flat = {}
|
||||
for key, sub in self._tensor_stats.items():
|
||||
for stat_name, tensor in sub.items():
|
||||
flat[f"{key}.{stat_name}"] = tensor
|
||||
return flat
|
||||
|
||||
def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
|
||||
self._tensor_stats.clear()
|
||||
for flat_key, tensor in state.items():
|
||||
key, stat_name = flat_key.rsplit(".", 1)
|
||||
self._tensor_stats.setdefault(key, {})[stat_name] = tensor
|
||||
|
||||
def reset(self):
|
||||
pass
|
||||
|
||||
def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
|
||||
return features
|
||||
|
||||
|
||||
@dataclass
|
||||
@ProcessorStepRegistry.register(name="unnormalizer_processor")
|
||||
class UnnormalizerProcessor:
|
||||
"""Inverse normalisation for observations and actions.
|
||||
|
||||
Exactly mirrors :class:`NormalizerProcessor` but applies the inverse
|
||||
transform.
|
||||
"""
|
||||
|
||||
features: dict[str, PolicyFeature]
|
||||
norm_map: dict[FeatureType, NormalizationMode]
|
||||
stats: dict[str, dict[str, Any]] | None = None
|
||||
|
||||
_tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
|
||||
|
||||
@classmethod
|
||||
def from_lerobot_dataset(
|
||||
cls,
|
||||
dataset: LeRobotDataset,
|
||||
features: dict[str, PolicyFeature],
|
||||
norm_map: dict[FeatureType, NormalizationMode],
|
||||
) -> UnnormalizerProcessor:
|
||||
return cls(features=features, norm_map=norm_map, stats=dataset.meta.stats)
|
||||
|
||||
def __post_init__(self):
|
||||
# Handle deserialization from JSON config
|
||||
if self.features and isinstance(list(self.features.values())[0], dict):
|
||||
# Features came from JSON - need to reconstruct PolicyFeature objects
|
||||
reconstructed_features = {}
|
||||
for key, ft_dict in self.features.items():
|
||||
reconstructed_features[key] = PolicyFeature(
|
||||
type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
|
||||
)
|
||||
self.features = reconstructed_features
|
||||
|
||||
if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
|
||||
# norm_map came from JSON - need to reconstruct enum keys and values
|
||||
reconstructed_norm_map = {}
|
||||
for ft_type_str, norm_mode_str in self.norm_map.items():
|
||||
reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
|
||||
self.norm_map = reconstructed_norm_map
|
||||
|
||||
self.stats = self.stats or {}
|
||||
self._tensor_stats = _convert_stats_to_tensors(self.stats)
|
||||
|
||||
def _unnormalize_obs(self, observation):
|
||||
if observation is None:
|
||||
return None
|
||||
keys = [k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION]
|
||||
processed = dict(observation)
|
||||
for key in keys:
|
||||
if key not in processed or key not in self._tensor_stats:
|
||||
continue
|
||||
orig_val = processed[key]
|
||||
tensor = (
|
||||
orig_val.to(dtype=torch.float32)
|
||||
if isinstance(orig_val, torch.Tensor)
|
||||
else torch.as_tensor(orig_val, dtype=torch.float32)
|
||||
)
|
||||
stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
|
||||
if "mean" in stats and "std" in stats:
|
||||
mean, std = stats["mean"], stats["std"]
|
||||
processed[key] = tensor * std + mean
|
||||
elif "min" in stats and "max" in stats:
|
||||
min_val, max_val = stats["min"], stats["max"]
|
||||
processed[key] = (tensor + 1) / 2 * (max_val - min_val) + min_val
|
||||
return processed
|
||||
|
||||
def _unnormalize_action(self, action):
|
||||
if action is None or "action" not in self._tensor_stats:
|
||||
return action
|
||||
tensor = (
|
||||
action.to(dtype=torch.float32)
|
||||
if isinstance(action, torch.Tensor)
|
||||
else torch.as_tensor(action, dtype=torch.float32)
|
||||
)
|
||||
stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
|
||||
if "mean" in stats and "std" in stats:
|
||||
mean, std = stats["mean"], stats["std"]
|
||||
return tensor * std + mean
|
||||
if "min" in stats and "max" in stats:
|
||||
min_val, max_val = stats["min"], stats["max"]
|
||||
return (tensor + 1) / 2 * (max_val - min_val) + min_val
|
||||
raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
|
||||
|
||||
def __call__(self, transition: EnvTransition) -> EnvTransition:
|
||||
observation = self._unnormalize_obs(transition.get(TransitionKey.OBSERVATION))
|
||||
action = self._unnormalize_action(transition.get(TransitionKey.ACTION))
|
||||
|
||||
# Create a new transition with unnormalized values
|
||||
new_transition = transition.copy()
|
||||
new_transition[TransitionKey.OBSERVATION] = observation
|
||||
new_transition[TransitionKey.ACTION] = action
|
||||
return new_transition
|
||||
|
||||
def get_config(self) -> dict[str, Any]:
|
||||
return {
|
||||
"features": {
|
||||
key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
|
||||
},
|
||||
"norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
|
||||
}
|
||||
|
||||
def state_dict(self) -> dict[str, Tensor]:
|
||||
flat = {}
|
||||
for key, sub in self._tensor_stats.items():
|
||||
for stat_name, tensor in sub.items():
|
||||
flat[f"{key}.{stat_name}"] = tensor
|
||||
return flat
|
||||
|
||||
def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
|
||||
self._tensor_stats.clear()
|
||||
for flat_key, tensor in state.items():
|
||||
key, stat_name = flat_key.rsplit(".", 1)
|
||||
self._tensor_stats.setdefault(key, {})[stat_name] = tensor
|
||||
|
||||
def reset(self):
|
||||
pass
|
||||
|
||||
def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
|
||||
return features
|
||||
@@ -1,157 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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 dataclasses import dataclass
|
||||
|
||||
import einops
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch import Tensor
|
||||
|
||||
from lerobot.configs.types import PolicyFeature
|
||||
from lerobot.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
|
||||
from lerobot.processor.pipeline import ObservationProcessor, ProcessorStepRegistry
|
||||
|
||||
|
||||
@dataclass
|
||||
@ProcessorStepRegistry.register(name="observation_processor")
|
||||
class VanillaObservationProcessor(ObservationProcessor):
|
||||
"""
|
||||
Processes environment observations into the LeRobot format by handling both images and states.
|
||||
|
||||
Image processing:
|
||||
- Converts channel-last (H, W, C) images to channel-first (C, H, W)
|
||||
- Normalizes uint8 images ([0, 255]) to float32 ([0, 1])
|
||||
- Adds a batch dimension if missing
|
||||
- Supports single images and image dictionaries
|
||||
|
||||
State processing:
|
||||
- Maps 'environment_state' to observation.environment_state
|
||||
- Maps 'agent_pos' to observation.state
|
||||
- Converts numpy arrays to tensors
|
||||
- Adds a batch dimension if missing
|
||||
"""
|
||||
|
||||
def _process_single_image(self, img: np.ndarray) -> Tensor:
|
||||
"""Process a single image array."""
|
||||
# Convert to tensor
|
||||
img_tensor = torch.from_numpy(img)
|
||||
|
||||
# Add batch dimension if needed
|
||||
if img_tensor.ndim == 3:
|
||||
img_tensor = img_tensor.unsqueeze(0)
|
||||
|
||||
# Validate image format
|
||||
_, h, w, c = img_tensor.shape
|
||||
if not (c < h and c < w):
|
||||
raise ValueError(f"Expected channel-last images, but got shape {img_tensor.shape}")
|
||||
|
||||
if img_tensor.dtype != torch.uint8:
|
||||
raise ValueError(f"Expected torch.uint8 images, but got {img_tensor.dtype}")
|
||||
|
||||
# Convert to channel-first format
|
||||
img_tensor = einops.rearrange(img_tensor, "b h w c -> b c h w").contiguous()
|
||||
|
||||
# Convert to float32 and normalize to [0, 1]
|
||||
img_tensor = img_tensor.type(torch.float32) / 255.0
|
||||
|
||||
return img_tensor
|
||||
|
||||
def _process_observation(self, observation):
|
||||
"""
|
||||
Processes both image and state observations.
|
||||
"""
|
||||
|
||||
processed_obs = observation.copy()
|
||||
|
||||
if "pixels" in processed_obs:
|
||||
pixels = processed_obs.pop("pixels")
|
||||
|
||||
if isinstance(pixels, dict):
|
||||
imgs = {f"{OBS_IMAGES}.{key}": img for key, img in pixels.items()}
|
||||
else:
|
||||
imgs = {OBS_IMAGE: pixels}
|
||||
|
||||
for imgkey, img in imgs.items():
|
||||
processed_obs[imgkey] = self._process_single_image(img)
|
||||
|
||||
if "environment_state" in processed_obs:
|
||||
env_state_np = processed_obs.pop("environment_state")
|
||||
env_state = torch.from_numpy(env_state_np).float()
|
||||
if env_state.dim() == 1:
|
||||
env_state = env_state.unsqueeze(0)
|
||||
processed_obs[OBS_ENV_STATE] = env_state
|
||||
|
||||
if "agent_pos" in processed_obs:
|
||||
agent_pos_np = processed_obs.pop("agent_pos")
|
||||
agent_pos = torch.from_numpy(agent_pos_np).float()
|
||||
if agent_pos.dim() == 1:
|
||||
agent_pos = agent_pos.unsqueeze(0)
|
||||
processed_obs[OBS_STATE] = agent_pos
|
||||
|
||||
return processed_obs
|
||||
|
||||
def observation(self, observation):
|
||||
return self._process_observation(observation)
|
||||
|
||||
def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
|
||||
"""Transforms feature keys to a standardized contract.
|
||||
|
||||
This method handles several renaming patterns:
|
||||
- Exact matches (e.g., 'pixels' -> 'OBS_IMAGE').
|
||||
- Prefixed exact matches (e.g., 'observation.pixels' -> 'OBS_IMAGE').
|
||||
- Prefix matches (e.g., 'pixels.cam1' -> 'OBS_IMAGES.cam1').
|
||||
- Prefixed prefix matches (e.g., 'observation.pixels.cam1' -> 'OBS_IMAGES.cam1').
|
||||
- environment_state -> OBS_ENV_STATE,
|
||||
- agent_pos -> OBS_STATE,
|
||||
- observation.environment_state -> OBS_ENV_STATE,
|
||||
- observation.agent_pos -> OBS_STATE
|
||||
"""
|
||||
exact_pairs = {
|
||||
"pixels": OBS_IMAGE,
|
||||
"environment_state": OBS_ENV_STATE,
|
||||
"agent_pos": OBS_STATE,
|
||||
}
|
||||
|
||||
prefix_pairs = {
|
||||
"pixels.": f"{OBS_IMAGES}.",
|
||||
}
|
||||
|
||||
for key in list(features.keys()):
|
||||
matched_prefix = False
|
||||
for old_prefix, new_prefix in prefix_pairs.items():
|
||||
prefixed_old = f"observation.{old_prefix}"
|
||||
if key.startswith(prefixed_old):
|
||||
suffix = key[len(prefixed_old) :]
|
||||
features[f"{new_prefix}{suffix}"] = features.pop(key)
|
||||
matched_prefix = True
|
||||
break
|
||||
|
||||
if key.startswith(old_prefix):
|
||||
suffix = key[len(old_prefix) :]
|
||||
features[f"{new_prefix}{suffix}"] = features.pop(key)
|
||||
matched_prefix = True
|
||||
break
|
||||
|
||||
if matched_prefix:
|
||||
continue
|
||||
|
||||
for old, new in exact_pairs.items():
|
||||
if key == old or key == f"observation.{old}":
|
||||
if key in features:
|
||||
features[new] = features.pop(key)
|
||||
break
|
||||
|
||||
return features
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,51 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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 dataclasses import dataclass, field
|
||||
from typing import Any
|
||||
|
||||
from lerobot.configs.types import PolicyFeature
|
||||
from lerobot.processor.pipeline import (
|
||||
ObservationProcessor,
|
||||
ProcessorStepRegistry,
|
||||
)
|
||||
|
||||
|
||||
@dataclass
|
||||
@ProcessorStepRegistry.register(name="rename_processor")
|
||||
class RenameProcessor(ObservationProcessor):
|
||||
"""Rename processor that renames keys in the observation."""
|
||||
|
||||
rename_map: dict[str, str] = field(default_factory=dict)
|
||||
|
||||
def observation(self, observation):
|
||||
processed_obs = {}
|
||||
for key, value in observation.items():
|
||||
if key in self.rename_map:
|
||||
processed_obs[self.rename_map[key]] = value
|
||||
else:
|
||||
processed_obs[key] = value
|
||||
|
||||
return processed_obs
|
||||
|
||||
def get_config(self) -> dict[str, Any]:
|
||||
return {"rename_map": self.rename_map}
|
||||
|
||||
def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
|
||||
"""Transforms:
|
||||
- Each key in the observation that appears in `rename_map` is renamed to its value.
|
||||
- Keys not in `rename_map` remain unchanged.
|
||||
"""
|
||||
return {self.rename_map.get(k, k): v for k, v in features.items()}
|
||||
@@ -18,7 +18,7 @@ Records a dataset. Actions for the robot can be either generated by teleoperatio
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.cameras="{laptop: {type: opencv, camera_index: 0, width: 640, height: 480}}" \
|
||||
@@ -36,7 +36,7 @@ lerobot-record \
|
||||
|
||||
Example recording with bimanual so100:
|
||||
```shell
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=bi_so100_follower \
|
||||
--robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
|
||||
--robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
|
||||
|
||||
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
|
||||
Examples:
|
||||
|
||||
```shell
|
||||
lerobot-replay \
|
||||
python -m lerobot.replay \
|
||||
--robot.type=so100_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.id=black \
|
||||
@@ -28,7 +28,7 @@ lerobot-replay \
|
||||
|
||||
Example replay with bimanual so100:
|
||||
```shell
|
||||
lerobot-replay \
|
||||
python -m lerobot.replay \
|
||||
--robot.type=bi_so100_follower \
|
||||
--robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
|
||||
--robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
|
||||
|
||||
@@ -161,11 +161,6 @@ class SO100Follower(Robot):
|
||||
self.bus.write("I_Coefficient", motor, 0)
|
||||
self.bus.write("D_Coefficient", motor, 32)
|
||||
|
||||
if motor == "gripper":
|
||||
self.bus.write("Max_Torque_Limit", motor, 500) # 50% of max torque to avoid burnout
|
||||
self.bus.write("Protection_Current", motor, 250) # 50% of max current to avoid burnout
|
||||
self.bus.write("Overload_Torque", motor, 25) # 25% torque when overloaded
|
||||
|
||||
def setup_motors(self) -> None:
|
||||
for motor in reversed(self.bus.motors):
|
||||
input(f"Connect the controller board to the '{motor}' motor only and press enter.")
|
||||
|
||||
@@ -157,13 +157,6 @@ class SO101Follower(Robot):
|
||||
self.bus.write("I_Coefficient", motor, 0)
|
||||
self.bus.write("D_Coefficient", motor, 32)
|
||||
|
||||
if motor == "gripper":
|
||||
self.bus.write(
|
||||
"Max_Torque_Limit", motor, 500
|
||||
) # 50% of the max torque limit to avoid burnout
|
||||
self.bus.write("Protection_Current", motor, 250) # 50% of max current to avoid burnout
|
||||
self.bus.write("Overload_Torque", motor, 25) # 25% torque when overloaded
|
||||
|
||||
def setup_motors(self) -> None:
|
||||
for motor in reversed(self.bus.motors):
|
||||
input(f"Connect the controller board to the '{motor}' motor only and press enter.")
|
||||
|
||||
@@ -141,10 +141,10 @@ python lerobot/scripts/control_robot.py \
|
||||
|
||||
## Train a policy
|
||||
|
||||
To train a policy to control your robot, use the [`lerobot-train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=${HF_USER}/aloha_test \
|
||||
--policy.type=act \
|
||||
--output_dir=outputs/train/act_aloha_test \
|
||||
|
||||
@@ -21,7 +21,7 @@ You want to evaluate a model from the hub (eg: https://huggingface.co/lerobot/di
|
||||
for 10 episodes.
|
||||
|
||||
```
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=lerobot/diffusion_pusht \
|
||||
--env.type=pusht \
|
||||
--eval.batch_size=10 \
|
||||
@@ -32,7 +32,7 @@ lerobot-eval \
|
||||
|
||||
OR, you want to evaluate a model checkpoint from the LeRobot training script for 10 episodes.
|
||||
```
|
||||
lerobot-eval \
|
||||
python -m lerobot.scripts.eval \
|
||||
--policy.path=outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
|
||||
--env.type=pusht \
|
||||
--eval.batch_size=10 \
|
||||
@@ -243,11 +243,7 @@ def eval_policy(
|
||||
if max_episodes_rendered > 0 and not videos_dir:
|
||||
raise ValueError("If max_episodes_rendered > 0, videos_dir must be provided.")
|
||||
|
||||
# Handle accelerate-wrapped models by unwrapping them
|
||||
if hasattr(policy, 'module') and isinstance(policy.module, PreTrainedPolicy):
|
||||
# This is likely an accelerate-wrapped model (DistributedDataParallel)
|
||||
policy = policy.module
|
||||
elif not isinstance(policy, PreTrainedPolicy):
|
||||
if not isinstance(policy, PreTrainedPolicy):
|
||||
raise ValueError(
|
||||
f"Policy of type 'PreTrainedPolicy' is expected, but type '{type(policy)}' was provided."
|
||||
)
|
||||
|
||||
@@ -302,6 +302,11 @@ class RobotClient:
|
||||
|
||||
self.logger.debug(f"Current latest action: {latest_action}")
|
||||
|
||||
# Get queue state before changes
|
||||
old_size, old_timesteps = self._inspect_action_queue()
|
||||
if not old_timesteps:
|
||||
old_timesteps = [latest_action] # queue was empty
|
||||
|
||||
# Get queue state before changes
|
||||
old_size, old_timesteps = self._inspect_action_queue()
|
||||
if not old_timesteps:
|
||||
|
||||
@@ -16,7 +16,6 @@
|
||||
import logging
|
||||
import time
|
||||
from contextlib import nullcontext
|
||||
from functools import partial
|
||||
from pprint import pformat
|
||||
from typing import Any
|
||||
|
||||
@@ -24,8 +23,6 @@ import torch
|
||||
from termcolor import colored
|
||||
from torch.amp import GradScaler
|
||||
from torch.optim import Optimizer
|
||||
import os
|
||||
from datetime import timedelta
|
||||
|
||||
from lerobot.configs import parser
|
||||
from lerobot.configs.train import TrainPipelineConfig
|
||||
@@ -55,8 +52,6 @@ from lerobot.utils.utils import (
|
||||
)
|
||||
from lerobot.utils.wandb_utils import WandBLogger
|
||||
|
||||
def is_launched_with_accelerate() -> bool:
|
||||
return "ACCELERATE_MIXED_PRECISION" in os.environ
|
||||
|
||||
def update_policy(
|
||||
train_metrics: MetricsTracker,
|
||||
@@ -64,65 +59,36 @@ def update_policy(
|
||||
batch: Any,
|
||||
optimizer: Optimizer,
|
||||
grad_clip_norm: float,
|
||||
grad_scaler: GradScaler | None,
|
||||
grad_scaler: GradScaler,
|
||||
lr_scheduler=None,
|
||||
use_amp: bool = False,
|
||||
lock=None,
|
||||
accelerator=None,
|
||||
) -> tuple[MetricsTracker, dict]:
|
||||
start_time = time.perf_counter()
|
||||
device = get_device_from_parameters(policy)
|
||||
policy.train()
|
||||
|
||||
grad_norm = 0.0 # Initialize grad_norm to avoid undefined variable
|
||||
|
||||
if accelerator:
|
||||
with accelerator.accumulate(policy):
|
||||
with torch.autocast(device_type=device.type) if use_amp else nullcontext():
|
||||
loss, output_dict = policy.forward(batch)
|
||||
# TODO(rcadene): policy.unnormalize_outputs(out_dict)
|
||||
accelerator.backward(loss)
|
||||
if accelerator.sync_gradients:
|
||||
grad_norm = torch.nn.utils.clip_grad_norm_(
|
||||
policy.parameters(),
|
||||
grad_clip_norm,
|
||||
error_if_nonfinite=False,
|
||||
)
|
||||
optimizer.step()
|
||||
optimizer.zero_grad()
|
||||
else:
|
||||
# Standard training loop without accelerate
|
||||
with torch.autocast(device_type=device.type) if use_amp else nullcontext():
|
||||
loss, output_dict = policy.forward(batch)
|
||||
with torch.autocast(device_type=device.type) if use_amp else nullcontext():
|
||||
loss, output_dict = policy.forward(batch)
|
||||
# TODO(rcadene): policy.unnormalize_outputs(out_dict)
|
||||
|
||||
if grad_scaler is not None:
|
||||
grad_scaler.scale(loss).backward()
|
||||
# Unscale the gradient of the optimizer's assigned params in-place **prior to gradient clipping**.
|
||||
grad_scaler.unscale_(optimizer)
|
||||
grad_norm = torch.nn.utils.clip_grad_norm_(
|
||||
policy.parameters(),
|
||||
grad_clip_norm,
|
||||
error_if_nonfinite=False,
|
||||
)
|
||||
# Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
|
||||
# although it still skips optimizer.step() if the gradients contain infs or NaNs.
|
||||
with lock if lock is not None else nullcontext():
|
||||
grad_scaler.step(optimizer)
|
||||
# Updates the scale for next iteration.
|
||||
grad_scaler.update()
|
||||
else:
|
||||
# Without GradScaler (fallback)
|
||||
loss.backward()
|
||||
grad_norm = torch.nn.utils.clip_grad_norm_(
|
||||
policy.parameters(),
|
||||
grad_clip_norm,
|
||||
error_if_nonfinite=False,
|
||||
)
|
||||
with lock if lock is not None else nullcontext():
|
||||
optimizer.step()
|
||||
grad_scaler.scale(loss).backward()
|
||||
|
||||
optimizer.zero_grad()
|
||||
# Unscale the gradient of the optimizer's assigned params in-place **prior to gradient clipping**.
|
||||
grad_scaler.unscale_(optimizer)
|
||||
|
||||
grad_norm = torch.nn.utils.clip_grad_norm_(
|
||||
policy.parameters(),
|
||||
grad_clip_norm,
|
||||
error_if_nonfinite=False,
|
||||
)
|
||||
|
||||
# Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
|
||||
# although it still skips optimizer.step() if the gradients contain infs or NaNs.
|
||||
with lock if lock is not None else nullcontext():
|
||||
grad_scaler.step(optimizer)
|
||||
# Updates the scale for next iteration.
|
||||
grad_scaler.update()
|
||||
|
||||
optimizer.zero_grad()
|
||||
|
||||
# Step through pytorch scheduler at every batch instead of epoch
|
||||
if lr_scheduler is not None:
|
||||
@@ -133,7 +99,7 @@ def update_policy(
|
||||
policy.update()
|
||||
|
||||
train_metrics.loss = loss.item()
|
||||
train_metrics.grad_norm = grad_norm.item() if isinstance(grad_norm, torch.Tensor) else grad_norm
|
||||
train_metrics.grad_norm = grad_norm.item()
|
||||
train_metrics.lr = optimizer.param_groups[0]["lr"]
|
||||
train_metrics.update_s = time.perf_counter() - start_time
|
||||
return train_metrics, output_dict
|
||||
@@ -142,33 +108,8 @@ def update_policy(
|
||||
@parser.wrap()
|
||||
def train(cfg: TrainPipelineConfig):
|
||||
cfg.validate()
|
||||
|
||||
accelerator = None
|
||||
if is_launched_with_accelerate():
|
||||
import accelerate
|
||||
|
||||
# For example pi0 has unused params (last llm block)
|
||||
from accelerate import DistributedDataParallelKwargs
|
||||
ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
||||
# accelerator = accelerate.Accelerator(step_scheduler_with_optimizer=False, kwargs_handlers=[ddp_kwargs])
|
||||
from accelerate import InitProcessGroupKwargs
|
||||
# Set NCCL timeout (default 30 minutes = 1800 seconds)
|
||||
nccl_timeout = getattr(cfg, 'nccl_timeout', 1800)
|
||||
ddp_init_kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=nccl_timeout)) # FIXME(mshukor): allow user to set timeout. This should be longer than the evaluation time
|
||||
# Set gradient accumulation steps (default 1)
|
||||
gradient_accumulation_steps = getattr(cfg, 'gradient_accumulation_steps', 1)
|
||||
accelerator = accelerate.Accelerator(step_scheduler_with_optimizer=False, gradient_accumulation_steps=gradient_accumulation_steps, kwargs_handlers=[ddp_init_kwargs, ddp_kwargs])
|
||||
if accelerator is not None and not accelerator.is_main_process:
|
||||
# Disable duplicate logging on non-main processes
|
||||
logging.info(f"Setting logging level on non-main process {accelerator.process_index} to WARNING.")
|
||||
logging.getLogger().setLevel(logging.WARNING)
|
||||
|
||||
logging.info(pformat(cfg.to_dict()))
|
||||
|
||||
if accelerator and not accelerator.is_main_process:
|
||||
# Disable logging on non-main processes.
|
||||
cfg.wandb.enable = False
|
||||
|
||||
if cfg.wandb.enable and cfg.wandb.project:
|
||||
wandb_logger = WandBLogger(cfg)
|
||||
else:
|
||||
@@ -202,8 +143,7 @@ def train(cfg: TrainPipelineConfig):
|
||||
|
||||
logging.info("Creating optimizer and scheduler")
|
||||
optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
|
||||
# Only use GradScaler when not using accelerate (accelerate handles mixed precision internally)
|
||||
grad_scaler = None if accelerator else GradScaler(device.type, enabled=cfg.policy.use_amp)
|
||||
grad_scaler = GradScaler(device.type, enabled=cfg.policy.use_amp)
|
||||
|
||||
step = 0 # number of policy updates (forward + backward + optim)
|
||||
|
||||
@@ -245,11 +185,6 @@ def train(cfg: TrainPipelineConfig):
|
||||
)
|
||||
dl_iter = cycle(dataloader)
|
||||
|
||||
# Prepare models for accelerate if using multi-GPU
|
||||
if accelerator:
|
||||
policy, optimizer, dataloader = accelerator.prepare(policy, optimizer, dataloader)
|
||||
dl_iter = cycle(dataloader)
|
||||
|
||||
policy.train()
|
||||
|
||||
train_metrics = {
|
||||
@@ -270,10 +205,9 @@ def train(cfg: TrainPipelineConfig):
|
||||
batch = next(dl_iter)
|
||||
train_tracker.dataloading_s = time.perf_counter() - start_time
|
||||
|
||||
if not accelerator:
|
||||
for key in batch:
|
||||
if isinstance(batch[key], torch.Tensor):
|
||||
batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")
|
||||
for key in batch:
|
||||
if isinstance(batch[key], torch.Tensor):
|
||||
batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")
|
||||
|
||||
train_tracker, output_dict = update_policy(
|
||||
train_tracker,
|
||||
@@ -284,7 +218,6 @@ def train(cfg: TrainPipelineConfig):
|
||||
grad_scaler=grad_scaler,
|
||||
lr_scheduler=lr_scheduler,
|
||||
use_amp=cfg.policy.use_amp,
|
||||
accelerator=accelerator,
|
||||
)
|
||||
|
||||
# Note: eval and checkpoint happens *after* the `step`th training update has completed, so we
|
||||
@@ -304,17 +237,15 @@ def train(cfg: TrainPipelineConfig):
|
||||
wandb_logger.log_dict(wandb_log_dict, step)
|
||||
train_tracker.reset_averages()
|
||||
|
||||
if cfg.save_checkpoint and is_saving_step and (not accelerator or accelerator.is_main_process):
|
||||
if cfg.save_checkpoint and is_saving_step:
|
||||
logging.info(f"Checkpoint policy after step {step}")
|
||||
checkpoint_dir = get_step_checkpoint_dir(cfg.output_dir, cfg.steps, step)
|
||||
# Unwrap model for accelerate
|
||||
policy_to_save = accelerator.unwrap_model(policy) if accelerator else policy
|
||||
save_checkpoint(checkpoint_dir, step, cfg, policy_to_save, optimizer, lr_scheduler)
|
||||
save_checkpoint(checkpoint_dir, step, cfg, policy, optimizer, lr_scheduler)
|
||||
update_last_checkpoint(checkpoint_dir)
|
||||
if wandb_logger:
|
||||
wandb_logger.log_policy(checkpoint_dir)
|
||||
|
||||
if cfg.env and is_eval_step and (not accelerator or accelerator.is_main_process):
|
||||
if cfg.env and is_eval_step:
|
||||
step_id = get_step_identifier(step, cfg.steps)
|
||||
logging.info(f"Eval policy at step {step}")
|
||||
with (
|
||||
@@ -323,7 +254,7 @@ def train(cfg: TrainPipelineConfig):
|
||||
):
|
||||
eval_info = eval_policy(
|
||||
eval_env,
|
||||
accelerator.unwrap_model(policy) if accelerator else policy,
|
||||
policy,
|
||||
cfg.eval.n_episodes,
|
||||
videos_dir=cfg.output_dir / "eval" / f"videos_step_{step_id}",
|
||||
max_episodes_rendered=4,
|
||||
|
||||
@@ -18,7 +18,7 @@ Helper to set motor ids and baudrate.
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-setup-motors \
|
||||
python -m lerobot.setup_motors \
|
||||
--teleop.type=so100_leader \
|
||||
--teleop.port=/dev/tty.usbmodem575E0031751
|
||||
```
|
||||
|
||||
@@ -18,7 +18,7 @@ Simple script to control a robot from teleoperation.
|
||||
Example:
|
||||
|
||||
```shell
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=so101_follower \
|
||||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||||
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
|
||||
@@ -32,7 +32,7 @@ lerobot-teleoperate \
|
||||
Example teleoperation with bimanual so100:
|
||||
|
||||
```shell
|
||||
lerobot-teleoperate \
|
||||
python -m lerobot.teleoperate \
|
||||
--robot.type=bi_so100_follower \
|
||||
--robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
|
||||
--robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
|
||||
|
||||
@@ -44,7 +44,7 @@ Below is the short version on how to train and run inference/eval:
|
||||
### Train from scratch
|
||||
|
||||
```bash
|
||||
lerobot-train \
|
||||
python -m lerobot.scripts.train \
|
||||
--dataset.repo_id=${HF_USER}/<dataset> \
|
||||
--policy.type=act \
|
||||
--output_dir=outputs/train/<desired_policy_repo_id> \
|
||||
@@ -59,7 +59,7 @@ _Writes checkpoints to `outputs/train/<desired_policy_repo_id>/checkpoints/`._
|
||||
### Evaluate the policy/run inference
|
||||
|
||||
```bash
|
||||
lerobot-record \
|
||||
python -m lerobot.record \
|
||||
--robot.type=so100_follower \
|
||||
--dataset.repo_id=<hf_user>/eval_<dataset> \
|
||||
--policy.path=<hf_user>/<desired_policy_repo_id> \
|
||||
|
||||
@@ -17,9 +17,10 @@ import time
|
||||
|
||||
|
||||
def busy_wait(seconds):
|
||||
if platform.system() == "Darwin" or platform.system() == "Windows":
|
||||
# On Mac and Windows, `time.sleep` is not accurate and we need to use this while loop trick,
|
||||
if platform.system() == "Darwin":
|
||||
# On Mac, `time.sleep` is not accurate and we need to use this while loop trick,
|
||||
# but it consumes CPU cycles.
|
||||
# TODO(rcadene): find an alternative: from python 11, time.sleep is precise
|
||||
end_time = time.perf_counter() + seconds
|
||||
while time.perf_counter() < end_time:
|
||||
pass
|
||||
|
||||
@@ -60,39 +60,11 @@ def load_training_step(save_dir: Path) -> int:
|
||||
|
||||
|
||||
def update_last_checkpoint(checkpoint_dir: Path) -> Path:
|
||||
import fcntl
|
||||
import tempfile
|
||||
import os
|
||||
|
||||
last_checkpoint_dir = checkpoint_dir.parent / LAST_CHECKPOINT_LINK
|
||||
if last_checkpoint_dir.is_symlink():
|
||||
last_checkpoint_dir.unlink()
|
||||
relative_target = checkpoint_dir.relative_to(checkpoint_dir.parent)
|
||||
|
||||
# Use file locking to prevent race conditions in multi-GPU training
|
||||
lock_file = checkpoint_dir.parent / ".symlink_lock"
|
||||
|
||||
try:
|
||||
with open(lock_file, 'w') as f:
|
||||
# Get exclusive lock
|
||||
fcntl.flock(f.fileno(), fcntl.LOCK_EX)
|
||||
|
||||
# Update symlink atomically
|
||||
if last_checkpoint_dir.exists() or last_checkpoint_dir.is_symlink():
|
||||
last_checkpoint_dir.unlink()
|
||||
last_checkpoint_dir.symlink_to(relative_target)
|
||||
|
||||
except (OSError, FileExistsError) as e:
|
||||
# Handle race conditions gracefully - another process may have already updated
|
||||
if not last_checkpoint_dir.exists():
|
||||
try:
|
||||
last_checkpoint_dir.symlink_to(relative_target)
|
||||
except FileExistsError:
|
||||
pass # Another process created it, that's fine
|
||||
finally:
|
||||
# Clean up lock file
|
||||
try:
|
||||
lock_file.unlink()
|
||||
except FileNotFoundError:
|
||||
pass
|
||||
last_checkpoint_dir.symlink_to(relative_target)
|
||||
|
||||
|
||||
def save_checkpoint(
|
||||
|
||||
@@ -1,45 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
echo "=== Local 1-GPU Accelerate Training Test with SmolVLA ==="
|
||||
echo "Environment: multi"
|
||||
echo "GPU: 1"
|
||||
echo "Steps: 50 (quick local test)"
|
||||
echo ""
|
||||
|
||||
# Activate conda environment
|
||||
source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
|
||||
conda activate multi
|
||||
|
||||
# Set CUDA environment for 1 GPU
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
|
||||
export TORCH_DISTRIBUTED_DEBUG=OFF
|
||||
export CUDA_LAUNCH_BLOCKING=0
|
||||
export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
|
||||
|
||||
# Change to working directory
|
||||
cd /fsx/dana_aubakirova/vla/pr/lerobot
|
||||
|
||||
# Set output directory with timestamp
|
||||
export OUTPUT_DIR="outputs/test_accelerate_1gpu_local_$(date +%Y%m%d_%H%M%S)"
|
||||
|
||||
echo "Output directory: $OUTPUT_DIR"
|
||||
echo ""
|
||||
|
||||
# Test accelerate training with 1 GPU
|
||||
accelerate launch --config_file accelerate_configs/1gpu_config.yaml -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/smolvla_base \
|
||||
--policy.push_to_hub=false \
|
||||
--dataset.repo_id=lerobot/svla_so100_sorting \
|
||||
--dataset.video_backend=pyav \
|
||||
--steps=50 \
|
||||
--save_freq=25 \
|
||||
--log_freq=5 \
|
||||
--batch_size=1 \
|
||||
--num_workers=0 \
|
||||
--output_dir=$OUTPUT_DIR \
|
||||
--wandb.enable=false
|
||||
|
||||
echo ""
|
||||
echo "=== Training completed! ==="
|
||||
echo "Check outputs in: $OUTPUT_DIR"
|
||||
@@ -1,67 +0,0 @@
|
||||
#!/bin/bash
|
||||
#SBATCH --job-name=test_accelerate
|
||||
#SBATCH --nodes=1
|
||||
#SBATCH --ntasks-per-node=1
|
||||
#SBATCH --cpus-per-task=16
|
||||
#SBATCH --gres=gpu:2
|
||||
#SBATCH --time=1:00:00
|
||||
#SBATCH --partition=hopper-prod
|
||||
#SBATCH --output=/fsx/dana_aubakirova/vla/logs/test_accelerate_%j.out
|
||||
#SBATCH --error=/fsx/dana_aubakirova/vla/logs/test_accelerate_%j.err
|
||||
|
||||
# Create logs directory if it doesn't exist
|
||||
mkdir -p /fsx/dana_aubakirova/vla/pr/lerobot/logs
|
||||
|
||||
# Activate conda environment
|
||||
source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
|
||||
conda activate multi
|
||||
|
||||
# 2-GPU Test CUDA environment
|
||||
export CUDA_VISIBLE_DEVICES=0,1
|
||||
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
|
||||
export TORCH_DISTRIBUTED_DEBUG=OFF
|
||||
export NCCL_DEBUG=INFO
|
||||
export CUDA_LAUNCH_BLOCKING=0
|
||||
export ACCELERATE_USE_FSDP=false
|
||||
export ACCELERATE_USE_DEEPSPEED=false
|
||||
export HF_ACCELERATE_DEVICE_MAP=false
|
||||
export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
|
||||
export SAFETENSORS_FAST_GPU=1
|
||||
export HF_HUB_ENABLE_HF_TRANSFER=1
|
||||
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
|
||||
export ACCELERATE_TORCH_DEVICE_MAP_AUTO=false
|
||||
|
||||
# Change to working directory
|
||||
cd /fsx/dana_aubakirova/vla/pr/lerobot
|
||||
|
||||
echo "=== Testing Accelerate Multi-GPU Training with SmolVLA ==="
|
||||
echo "Dataset: lerobot/svla_so100_sorting"
|
||||
echo "GPUs: 2"
|
||||
echo "Steps: 100 (for quick test)"
|
||||
echo "Job ID: $SLURM_JOB_ID"
|
||||
echo ""
|
||||
|
||||
# Set output directory with job ID
|
||||
export OUTPUT_DIR="outputs/test_accelerate_2gpu_job_${SLURM_JOB_ID}"
|
||||
|
||||
echo "Output directory: $OUTPUT_DIR"
|
||||
echo ""
|
||||
|
||||
# Test accelerate training
|
||||
accelerate launch --config_file accelerate_configs/2gpu_config_safe.yaml -m lerobot.scripts.train \
|
||||
--policy.type=smolvla \
|
||||
--policy.push_to_hub=false \
|
||||
--dataset.repo_id=lerobot/svla_so100_sorting \
|
||||
--dataset.video_backend=pyav \
|
||||
--steps=100 \
|
||||
--save_freq=50 \
|
||||
--log_freq=5 \
|
||||
--batch_size=2 \
|
||||
--num_workers=0 \
|
||||
--output_dir=$OUTPUT_DIR \
|
||||
--wandb.enable=false
|
||||
|
||||
echo ""
|
||||
echo "=== Training completed! ==="
|
||||
echo "Check logs and outputs in: $OUTPUT_DIR"
|
||||
echo "Job ID: $SLURM_JOB_ID"
|
||||
@@ -1,45 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
echo "=== Direct 1-GPU Training Test with SmolVLA (no accelerate) ==="
|
||||
echo "Environment: multi"
|
||||
echo "GPU: 1"
|
||||
echo "Steps: 50 (quick local test)"
|
||||
echo ""
|
||||
|
||||
# Activate conda environment
|
||||
source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
|
||||
conda activate multi
|
||||
|
||||
# Set CUDA environment for 1 GPU
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
|
||||
export TORCH_DISTRIBUTED_DEBUG=OFF
|
||||
export CUDA_LAUNCH_BLOCKING=0
|
||||
export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
|
||||
|
||||
# Change to working directory
|
||||
cd /fsx/dana_aubakirova/vla/pr/lerobot
|
||||
|
||||
# Set output directory with timestamp
|
||||
export OUTPUT_DIR="outputs/test_direct_1gpu_local_$(date +%Y%m%d_%H%M%S)"
|
||||
|
||||
echo "Output directory: $OUTPUT_DIR"
|
||||
echo ""
|
||||
|
||||
# Test direct training with 1 GPU (no accelerate)
|
||||
python -m lerobot.scripts.train \
|
||||
--policy.path=lerobot/smolvla_base \
|
||||
--policy.push_to_hub=false \
|
||||
--dataset.repo_id=lerobot/svla_so100_sorting \
|
||||
--dataset.video_backend=pyav \
|
||||
--steps=50 \
|
||||
--save_freq=25 \
|
||||
--log_freq=5 \
|
||||
--batch_size=1 \
|
||||
--num_workers=0 \
|
||||
--output_dir=$OUTPUT_DIR \
|
||||
--wandb.enable=false
|
||||
|
||||
echo ""
|
||||
echo "=== Training completed! ==="
|
||||
echo "Check outputs in: $OUTPUT_DIR"
|
||||
@@ -19,7 +19,6 @@ import traceback
|
||||
import pytest
|
||||
from serial import SerialException
|
||||
|
||||
from lerobot.configs.types import FeatureType, PolicyFeature
|
||||
from tests.utils import DEVICE
|
||||
|
||||
# Import fixture modules as plugins
|
||||
@@ -70,19 +69,3 @@ def patch_builtins_input(monkeypatch):
|
||||
print(text)
|
||||
|
||||
monkeypatch.setattr("builtins.input", print_text)
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def policy_feature_factory():
|
||||
"""PolicyFeature factory"""
|
||||
|
||||
def _pf(ft: FeatureType, shape: tuple[int, ...]) -> PolicyFeature:
|
||||
return PolicyFeature(type=ft, shape=shape)
|
||||
|
||||
return _pf
|
||||
|
||||
|
||||
def assert_contract_is_typed(features: dict[str, PolicyFeature]) -> None:
|
||||
assert isinstance(features, dict)
|
||||
assert all(isinstance(k, str) for k in features.keys())
|
||||
assert all(isinstance(v, PolicyFeature) for v in features.values())
|
||||
|
||||
@@ -27,13 +27,11 @@ from lerobot import available_policies
|
||||
from lerobot.configs.default import DatasetConfig
|
||||
from lerobot.configs.train import TrainPipelineConfig
|
||||
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
|
||||
from lerobot.constants import ACTION, OBS_STATE
|
||||
from lerobot.datasets.factory import make_dataset
|
||||
from lerobot.datasets.utils import cycle, dataset_to_policy_features
|
||||
from lerobot.envs.factory import make_env, make_env_config
|
||||
from lerobot.envs.utils import preprocess_observation
|
||||
from lerobot.optim.factory import make_optimizer_and_scheduler
|
||||
from lerobot.policies.act.configuration_act import ACTConfig
|
||||
from lerobot.policies.act.modeling_act import ACTTemporalEnsembler
|
||||
from lerobot.policies.factory import (
|
||||
get_policy_class,
|
||||
@@ -365,54 +363,6 @@ def test_normalize(insert_temporal_dim):
|
||||
unnormalize(output_batch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("multikey", [True, False])
|
||||
def test_multikey_construction(multikey: bool):
|
||||
"""
|
||||
Asserts that multiple keys with type State/Action are correctly processed by the policy constructor,
|
||||
preventing erroneous creation of the policy object.
|
||||
"""
|
||||
input_features = {
|
||||
"observation.state": PolicyFeature(
|
||||
type=FeatureType.STATE,
|
||||
shape=(10,),
|
||||
),
|
||||
}
|
||||
output_features = {
|
||||
"action": PolicyFeature(
|
||||
type=FeatureType.ACTION,
|
||||
shape=(5,),
|
||||
),
|
||||
}
|
||||
|
||||
if multikey:
|
||||
"""Simulates the complete state/action is constructed from more granular multiple
|
||||
keys, of the same type as the overall state/action"""
|
||||
input_features = {}
|
||||
input_features["observation.state.subset1"] = PolicyFeature(type=FeatureType.STATE, shape=(5,))
|
||||
input_features["observation.state.subset2"] = PolicyFeature(type=FeatureType.STATE, shape=(5,))
|
||||
input_features["observation.state"] = PolicyFeature(type=FeatureType.STATE, shape=(10,))
|
||||
|
||||
output_features = {}
|
||||
output_features["action.first_three_motors"] = PolicyFeature(type=FeatureType.ACTION, shape=(3,))
|
||||
output_features["action.last_two_motors"] = PolicyFeature(type=FeatureType.ACTION, shape=(2,))
|
||||
output_features["action"] = PolicyFeature(
|
||||
type=FeatureType.ACTION,
|
||||
shape=(5,),
|
||||
)
|
||||
|
||||
config = ACTConfig(input_features=input_features, output_features=output_features)
|
||||
|
||||
state_condition = config.robot_state_feature == input_features[OBS_STATE]
|
||||
action_condition = config.action_feature == output_features[ACTION]
|
||||
|
||||
assert state_condition, (
|
||||
f"Discrepancy detected. Robot state feature is {config.robot_state_feature} but policy expects {input_features[OBS_STATE]}"
|
||||
)
|
||||
assert action_condition, (
|
||||
f"Discrepancy detected. Action feature is {config.action_feature} but policy expects {output_features[ACTION]}"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"ds_repo_id, policy_name, policy_kwargs, file_name_extra",
|
||||
[
|
||||
|
||||
@@ -1,282 +0,0 @@
|
||||
import torch
|
||||
|
||||
from lerobot.processor.pipeline import (
|
||||
RobotProcessor,
|
||||
TransitionKey,
|
||||
_default_batch_to_transition,
|
||||
_default_transition_to_batch,
|
||||
)
|
||||
|
||||
|
||||
def _dummy_batch():
|
||||
"""Create a dummy batch using the new format with observation.* and next.* keys."""
|
||||
return {
|
||||
"observation.image.left": torch.randn(1, 3, 128, 128),
|
||||
"observation.image.right": torch.randn(1, 3, 128, 128),
|
||||
"observation.state": torch.tensor([[0.1, 0.2, 0.3, 0.4]]),
|
||||
"action": torch.tensor([[0.5]]),
|
||||
"next.reward": 1.0,
|
||||
"next.done": False,
|
||||
"next.truncated": False,
|
||||
"info": {"key": "value"},
|
||||
}
|
||||
|
||||
|
||||
def test_observation_grouping_roundtrip():
|
||||
"""Test that observation.* keys are properly grouped and ungrouped."""
|
||||
proc = RobotProcessor([])
|
||||
batch_in = _dummy_batch()
|
||||
batch_out = proc(batch_in)
|
||||
|
||||
# Check that all observation.* keys are preserved
|
||||
original_obs_keys = {k: v for k, v in batch_in.items() if k.startswith("observation.")}
|
||||
reconstructed_obs_keys = {k: v for k, v in batch_out.items() if k.startswith("observation.")}
|
||||
|
||||
assert set(original_obs_keys.keys()) == set(reconstructed_obs_keys.keys())
|
||||
|
||||
# Check tensor values
|
||||
assert torch.allclose(batch_out["observation.image.left"], batch_in["observation.image.left"])
|
||||
assert torch.allclose(batch_out["observation.image.right"], batch_in["observation.image.right"])
|
||||
assert torch.allclose(batch_out["observation.state"], batch_in["observation.state"])
|
||||
|
||||
# Check other fields
|
||||
assert torch.allclose(batch_out["action"], batch_in["action"])
|
||||
assert batch_out["next.reward"] == batch_in["next.reward"]
|
||||
assert batch_out["next.done"] == batch_in["next.done"]
|
||||
assert batch_out["next.truncated"] == batch_in["next.truncated"]
|
||||
assert batch_out["info"] == batch_in["info"]
|
||||
|
||||
|
||||
def test_batch_to_transition_observation_grouping():
|
||||
"""Test that _default_batch_to_transition correctly groups observation.* keys."""
|
||||
batch = {
|
||||
"observation.image.top": torch.randn(1, 3, 128, 128),
|
||||
"observation.image.left": torch.randn(1, 3, 128, 128),
|
||||
"observation.state": [1, 2, 3, 4],
|
||||
"action": "action_data",
|
||||
"next.reward": 1.5,
|
||||
"next.done": True,
|
||||
"next.truncated": False,
|
||||
"info": {"episode": 42},
|
||||
}
|
||||
|
||||
transition = _default_batch_to_transition(batch)
|
||||
|
||||
# Check observation is a dict with all observation.* keys
|
||||
assert isinstance(transition[TransitionKey.OBSERVATION], dict)
|
||||
assert "observation.image.top" in transition[TransitionKey.OBSERVATION]
|
||||
assert "observation.image.left" in transition[TransitionKey.OBSERVATION]
|
||||
assert "observation.state" in transition[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check values are preserved
|
||||
assert torch.allclose(
|
||||
transition[TransitionKey.OBSERVATION]["observation.image.top"], batch["observation.image.top"]
|
||||
)
|
||||
assert torch.allclose(
|
||||
transition[TransitionKey.OBSERVATION]["observation.image.left"], batch["observation.image.left"]
|
||||
)
|
||||
assert transition[TransitionKey.OBSERVATION]["observation.state"] == [1, 2, 3, 4]
|
||||
|
||||
# Check other fields
|
||||
assert transition[TransitionKey.ACTION] == "action_data"
|
||||
assert transition[TransitionKey.REWARD] == 1.5
|
||||
assert transition[TransitionKey.DONE]
|
||||
assert not transition[TransitionKey.TRUNCATED]
|
||||
assert transition[TransitionKey.INFO] == {"episode": 42}
|
||||
assert transition[TransitionKey.COMPLEMENTARY_DATA] == {}
|
||||
|
||||
|
||||
def test_transition_to_batch_observation_flattening():
|
||||
"""Test that _default_transition_to_batch correctly flattens observation dict."""
|
||||
observation_dict = {
|
||||
"observation.image.top": torch.randn(1, 3, 128, 128),
|
||||
"observation.image.left": torch.randn(1, 3, 128, 128),
|
||||
"observation.state": [1, 2, 3, 4],
|
||||
}
|
||||
|
||||
transition = {
|
||||
TransitionKey.OBSERVATION: observation_dict,
|
||||
TransitionKey.ACTION: "action_data",
|
||||
TransitionKey.REWARD: 1.5,
|
||||
TransitionKey.DONE: True,
|
||||
TransitionKey.TRUNCATED: False,
|
||||
TransitionKey.INFO: {"episode": 42},
|
||||
TransitionKey.COMPLEMENTARY_DATA: {},
|
||||
}
|
||||
|
||||
batch = _default_transition_to_batch(transition)
|
||||
|
||||
# Check that observation.* keys are flattened back to batch
|
||||
assert "observation.image.top" in batch
|
||||
assert "observation.image.left" in batch
|
||||
assert "observation.state" in batch
|
||||
|
||||
# Check values are preserved
|
||||
assert torch.allclose(batch["observation.image.top"], observation_dict["observation.image.top"])
|
||||
assert torch.allclose(batch["observation.image.left"], observation_dict["observation.image.left"])
|
||||
assert batch["observation.state"] == [1, 2, 3, 4]
|
||||
|
||||
# Check other fields are mapped to next.* format
|
||||
assert batch["action"] == "action_data"
|
||||
assert batch["next.reward"] == 1.5
|
||||
assert batch["next.done"]
|
||||
assert not batch["next.truncated"]
|
||||
assert batch["info"] == {"episode": 42}
|
||||
|
||||
|
||||
def test_no_observation_keys():
|
||||
"""Test behavior when there are no observation.* keys."""
|
||||
batch = {
|
||||
"action": "action_data",
|
||||
"next.reward": 2.0,
|
||||
"next.done": False,
|
||||
"next.truncated": True,
|
||||
"info": {"test": "no_obs"},
|
||||
}
|
||||
|
||||
transition = _default_batch_to_transition(batch)
|
||||
|
||||
# Observation should be None when no observation.* keys
|
||||
assert transition[TransitionKey.OBSERVATION] is None
|
||||
|
||||
# Check other fields
|
||||
assert transition[TransitionKey.ACTION] == "action_data"
|
||||
assert transition[TransitionKey.REWARD] == 2.0
|
||||
assert not transition[TransitionKey.DONE]
|
||||
assert transition[TransitionKey.TRUNCATED]
|
||||
assert transition[TransitionKey.INFO] == {"test": "no_obs"}
|
||||
|
||||
# Round trip should work
|
||||
reconstructed_batch = _default_transition_to_batch(transition)
|
||||
assert reconstructed_batch["action"] == "action_data"
|
||||
assert reconstructed_batch["next.reward"] == 2.0
|
||||
assert not reconstructed_batch["next.done"]
|
||||
assert reconstructed_batch["next.truncated"]
|
||||
assert reconstructed_batch["info"] == {"test": "no_obs"}
|
||||
|
||||
|
||||
def test_minimal_batch():
|
||||
"""Test with minimal batch containing only observation.* and action."""
|
||||
batch = {"observation.state": "minimal_state", "action": "minimal_action"}
|
||||
|
||||
transition = _default_batch_to_transition(batch)
|
||||
|
||||
# Check observation
|
||||
assert transition[TransitionKey.OBSERVATION] == {"observation.state": "minimal_state"}
|
||||
assert transition[TransitionKey.ACTION] == "minimal_action"
|
||||
|
||||
# Check defaults
|
||||
assert transition[TransitionKey.REWARD] == 0.0
|
||||
assert not transition[TransitionKey.DONE]
|
||||
assert not transition[TransitionKey.TRUNCATED]
|
||||
assert transition[TransitionKey.INFO] == {}
|
||||
assert transition[TransitionKey.COMPLEMENTARY_DATA] == {}
|
||||
|
||||
# Round trip
|
||||
reconstructed_batch = _default_transition_to_batch(transition)
|
||||
assert reconstructed_batch["observation.state"] == "minimal_state"
|
||||
assert reconstructed_batch["action"] == "minimal_action"
|
||||
assert reconstructed_batch["next.reward"] == 0.0
|
||||
assert not reconstructed_batch["next.done"]
|
||||
assert not reconstructed_batch["next.truncated"]
|
||||
assert reconstructed_batch["info"] == {}
|
||||
|
||||
|
||||
def test_empty_batch():
|
||||
"""Test behavior with empty batch."""
|
||||
batch = {}
|
||||
|
||||
transition = _default_batch_to_transition(batch)
|
||||
|
||||
# All fields should have defaults
|
||||
assert transition[TransitionKey.OBSERVATION] is None
|
||||
assert transition[TransitionKey.ACTION] is None
|
||||
assert transition[TransitionKey.REWARD] == 0.0
|
||||
assert not transition[TransitionKey.DONE]
|
||||
assert not transition[TransitionKey.TRUNCATED]
|
||||
assert transition[TransitionKey.INFO] == {}
|
||||
assert transition[TransitionKey.COMPLEMENTARY_DATA] == {}
|
||||
|
||||
# Round trip
|
||||
reconstructed_batch = _default_transition_to_batch(transition)
|
||||
assert reconstructed_batch["action"] is None
|
||||
assert reconstructed_batch["next.reward"] == 0.0
|
||||
assert not reconstructed_batch["next.done"]
|
||||
assert not reconstructed_batch["next.truncated"]
|
||||
assert reconstructed_batch["info"] == {}
|
||||
|
||||
|
||||
def test_complex_nested_observation():
|
||||
"""Test with complex nested observation data."""
|
||||
batch = {
|
||||
"observation.image.top": {"image": torch.randn(1, 3, 128, 128), "timestamp": 1234567890},
|
||||
"observation.image.left": {"image": torch.randn(1, 3, 128, 128), "timestamp": 1234567891},
|
||||
"observation.state": torch.randn(7),
|
||||
"action": torch.randn(8),
|
||||
"next.reward": 3.14,
|
||||
"next.done": False,
|
||||
"next.truncated": True,
|
||||
"info": {"episode_length": 200, "success": True},
|
||||
}
|
||||
|
||||
transition = _default_batch_to_transition(batch)
|
||||
reconstructed_batch = _default_transition_to_batch(transition)
|
||||
|
||||
# Check that all observation keys are preserved
|
||||
original_obs_keys = {k for k in batch if k.startswith("observation.")}
|
||||
reconstructed_obs_keys = {k for k in reconstructed_batch if k.startswith("observation.")}
|
||||
|
||||
assert original_obs_keys == reconstructed_obs_keys
|
||||
|
||||
# Check tensor values
|
||||
assert torch.allclose(batch["observation.state"], reconstructed_batch["observation.state"])
|
||||
|
||||
# Check nested dict with tensors
|
||||
assert torch.allclose(
|
||||
batch["observation.image.top"]["image"], reconstructed_batch["observation.image.top"]["image"]
|
||||
)
|
||||
assert torch.allclose(
|
||||
batch["observation.image.left"]["image"], reconstructed_batch["observation.image.left"]["image"]
|
||||
)
|
||||
|
||||
# Check action tensor
|
||||
assert torch.allclose(batch["action"], reconstructed_batch["action"])
|
||||
|
||||
# Check other fields
|
||||
assert batch["next.reward"] == reconstructed_batch["next.reward"]
|
||||
assert batch["next.done"] == reconstructed_batch["next.done"]
|
||||
assert batch["next.truncated"] == reconstructed_batch["next.truncated"]
|
||||
assert batch["info"] == reconstructed_batch["info"]
|
||||
|
||||
|
||||
def test_custom_converter():
|
||||
"""Test that custom converters can still be used."""
|
||||
|
||||
def to_tr(batch):
|
||||
# Custom converter that modifies the reward
|
||||
tr = _default_batch_to_transition(batch)
|
||||
# Double the reward
|
||||
reward = tr.get(TransitionKey.REWARD, 0.0)
|
||||
new_tr = tr.copy()
|
||||
new_tr[TransitionKey.REWARD] = reward * 2 if reward is not None else 0.0
|
||||
return new_tr
|
||||
|
||||
def to_batch(tr):
|
||||
batch = _default_transition_to_batch(tr)
|
||||
return batch
|
||||
|
||||
processor = RobotProcessor(steps=[], to_transition=to_tr, to_output=to_batch)
|
||||
|
||||
batch = {
|
||||
"observation.state": torch.randn(1, 4),
|
||||
"action": torch.randn(1, 2),
|
||||
"next.reward": 1.0,
|
||||
"next.done": False,
|
||||
}
|
||||
|
||||
result = processor(batch)
|
||||
|
||||
# Check the reward was doubled by our custom converter
|
||||
assert result["next.reward"] == 2.0
|
||||
assert torch.allclose(result["observation.state"], batch["observation.state"])
|
||||
assert torch.allclose(result["action"], batch["action"])
|
||||
@@ -1,628 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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 unittest.mock import Mock
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
|
||||
from lerobot.processor.normalize_processor import (
|
||||
NormalizerProcessor,
|
||||
UnnormalizerProcessor,
|
||||
_convert_stats_to_tensors,
|
||||
)
|
||||
from lerobot.processor.pipeline import RobotProcessor, TransitionKey
|
||||
|
||||
|
||||
def create_transition(
|
||||
observation=None, action=None, reward=None, done=None, truncated=None, info=None, complementary_data=None
|
||||
):
|
||||
"""Helper to create an EnvTransition dictionary."""
|
||||
return {
|
||||
TransitionKey.OBSERVATION: observation,
|
||||
TransitionKey.ACTION: action,
|
||||
TransitionKey.REWARD: reward,
|
||||
TransitionKey.DONE: done,
|
||||
TransitionKey.TRUNCATED: truncated,
|
||||
TransitionKey.INFO: info,
|
||||
TransitionKey.COMPLEMENTARY_DATA: complementary_data,
|
||||
}
|
||||
|
||||
|
||||
def test_numpy_conversion():
|
||||
stats = {
|
||||
"observation.image": {
|
||||
"mean": np.array([0.5, 0.5, 0.5]),
|
||||
"std": np.array([0.2, 0.2, 0.2]),
|
||||
}
|
||||
}
|
||||
tensor_stats = _convert_stats_to_tensors(stats)
|
||||
|
||||
assert isinstance(tensor_stats["observation.image"]["mean"], torch.Tensor)
|
||||
assert isinstance(tensor_stats["observation.image"]["std"], torch.Tensor)
|
||||
assert torch.allclose(tensor_stats["observation.image"]["mean"], torch.tensor([0.5, 0.5, 0.5]))
|
||||
assert torch.allclose(tensor_stats["observation.image"]["std"], torch.tensor([0.2, 0.2, 0.2]))
|
||||
|
||||
|
||||
def test_tensor_conversion():
|
||||
stats = {
|
||||
"action": {
|
||||
"mean": torch.tensor([0.0, 0.0]),
|
||||
"std": torch.tensor([1.0, 1.0]),
|
||||
}
|
||||
}
|
||||
tensor_stats = _convert_stats_to_tensors(stats)
|
||||
|
||||
assert tensor_stats["action"]["mean"].dtype == torch.float32
|
||||
assert tensor_stats["action"]["std"].dtype == torch.float32
|
||||
|
||||
|
||||
def test_scalar_conversion():
|
||||
stats = {
|
||||
"reward": {
|
||||
"mean": 0.5,
|
||||
"std": 0.1,
|
||||
}
|
||||
}
|
||||
tensor_stats = _convert_stats_to_tensors(stats)
|
||||
|
||||
assert torch.allclose(tensor_stats["reward"]["mean"], torch.tensor(0.5))
|
||||
assert torch.allclose(tensor_stats["reward"]["std"], torch.tensor(0.1))
|
||||
|
||||
|
||||
def test_list_conversion():
|
||||
stats = {
|
||||
"observation.state": {
|
||||
"min": [0.0, -1.0, -2.0],
|
||||
"max": [1.0, 1.0, 2.0],
|
||||
}
|
||||
}
|
||||
tensor_stats = _convert_stats_to_tensors(stats)
|
||||
|
||||
assert torch.allclose(tensor_stats["observation.state"]["min"], torch.tensor([0.0, -1.0, -2.0]))
|
||||
assert torch.allclose(tensor_stats["observation.state"]["max"], torch.tensor([1.0, 1.0, 2.0]))
|
||||
|
||||
|
||||
def test_unsupported_type():
|
||||
stats = {
|
||||
"bad_key": {
|
||||
"mean": "string_value",
|
||||
}
|
||||
}
|
||||
with pytest.raises(TypeError, match="Unsupported type"):
|
||||
_convert_stats_to_tensors(stats)
|
||||
|
||||
|
||||
# Helper functions to create feature maps and norm maps
|
||||
def _create_observation_features():
|
||||
return {
|
||||
"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96)),
|
||||
"observation.state": PolicyFeature(FeatureType.STATE, (2,)),
|
||||
}
|
||||
|
||||
|
||||
def _create_observation_norm_map():
|
||||
return {
|
||||
FeatureType.VISUAL: NormalizationMode.MEAN_STD,
|
||||
FeatureType.STATE: NormalizationMode.MIN_MAX,
|
||||
}
|
||||
|
||||
|
||||
# Fixtures for observation normalisation tests using NormalizerProcessor
|
||||
@pytest.fixture
|
||||
def observation_stats():
|
||||
return {
|
||||
"observation.image": {
|
||||
"mean": np.array([0.5, 0.5, 0.5]),
|
||||
"std": np.array([0.2, 0.2, 0.2]),
|
||||
},
|
||||
"observation.state": {
|
||||
"min": np.array([0.0, -1.0]),
|
||||
"max": np.array([1.0, 1.0]),
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def observation_normalizer(observation_stats):
|
||||
"""Return a NormalizerProcessor that only has observation stats (no action)."""
|
||||
features = _create_observation_features()
|
||||
norm_map = _create_observation_norm_map()
|
||||
return NormalizerProcessor(features=features, norm_map=norm_map, stats=observation_stats)
|
||||
|
||||
|
||||
def test_mean_std_normalization(observation_normalizer):
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]),
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
normalized_transition = observation_normalizer(transition)
|
||||
normalized_obs = normalized_transition[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check mean/std normalization
|
||||
expected_image = (torch.tensor([0.7, 0.5, 0.3]) - 0.5) / 0.2
|
||||
assert torch.allclose(normalized_obs["observation.image"], expected_image)
|
||||
|
||||
|
||||
def test_min_max_normalization(observation_normalizer):
|
||||
observation = {
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
normalized_transition = observation_normalizer(transition)
|
||||
normalized_obs = normalized_transition[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check min/max normalization to [-1, 1]
|
||||
# For state[0]: 2 * (0.5 - 0.0) / (1.0 - 0.0) - 1 = 0.0
|
||||
# For state[1]: 2 * (0.0 - (-1.0)) / (1.0 - (-1.0)) - 1 = 0.0
|
||||
expected_state = torch.tensor([0.0, 0.0])
|
||||
assert torch.allclose(normalized_obs["observation.state"], expected_state, atol=1e-6)
|
||||
|
||||
|
||||
def test_selective_normalization(observation_stats):
|
||||
features = _create_observation_features()
|
||||
norm_map = _create_observation_norm_map()
|
||||
normalizer = NormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats=observation_stats, normalize_keys={"observation.image"}
|
||||
)
|
||||
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]),
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
normalized_transition = normalizer(transition)
|
||||
normalized_obs = normalized_transition[TransitionKey.OBSERVATION]
|
||||
|
||||
# Only image should be normalized
|
||||
assert torch.allclose(normalized_obs["observation.image"], (torch.tensor([0.7, 0.5, 0.3]) - 0.5) / 0.2)
|
||||
# State should remain unchanged
|
||||
assert torch.allclose(normalized_obs["observation.state"], observation["observation.state"])
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
|
||||
def test_device_compatibility(observation_stats):
|
||||
features = _create_observation_features()
|
||||
norm_map = _create_observation_norm_map()
|
||||
normalizer = NormalizerProcessor(features=features, norm_map=norm_map, stats=observation_stats)
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]).cuda(),
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
normalized_transition = normalizer(transition)
|
||||
normalized_obs = normalized_transition[TransitionKey.OBSERVATION]
|
||||
|
||||
assert normalized_obs["observation.image"].device.type == "cuda"
|
||||
|
||||
|
||||
def test_from_lerobot_dataset():
|
||||
# Mock dataset
|
||||
mock_dataset = Mock()
|
||||
mock_dataset.meta.stats = {
|
||||
"observation.image": {"mean": [0.5], "std": [0.2]},
|
||||
"action": {"mean": [0.0], "std": [1.0]},
|
||||
}
|
||||
|
||||
features = {
|
||||
"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96)),
|
||||
"action": PolicyFeature(FeatureType.ACTION, (1,)),
|
||||
}
|
||||
norm_map = {
|
||||
FeatureType.VISUAL: NormalizationMode.MEAN_STD,
|
||||
FeatureType.ACTION: NormalizationMode.MEAN_STD,
|
||||
}
|
||||
|
||||
normalizer = NormalizerProcessor.from_lerobot_dataset(mock_dataset, features, norm_map)
|
||||
|
||||
# Both observation and action statistics should be present in tensor stats
|
||||
assert "observation.image" in normalizer._tensor_stats
|
||||
assert "action" in normalizer._tensor_stats
|
||||
|
||||
|
||||
def test_state_dict_save_load(observation_normalizer):
|
||||
# Save state
|
||||
state_dict = observation_normalizer.state_dict()
|
||||
|
||||
# Create new normalizer and load state
|
||||
features = _create_observation_features()
|
||||
norm_map = _create_observation_norm_map()
|
||||
new_normalizer = NormalizerProcessor(features=features, norm_map=norm_map, stats={})
|
||||
new_normalizer.load_state_dict(state_dict)
|
||||
|
||||
# Test that it works the same
|
||||
observation = {"observation.image": torch.tensor([0.7, 0.5, 0.3])}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result1 = observation_normalizer(transition)[TransitionKey.OBSERVATION]
|
||||
result2 = new_normalizer(transition)[TransitionKey.OBSERVATION]
|
||||
|
||||
assert torch.allclose(result1["observation.image"], result2["observation.image"])
|
||||
|
||||
|
||||
# Fixtures for ActionUnnormalizer tests
|
||||
@pytest.fixture
|
||||
def action_stats_mean_std():
|
||||
return {
|
||||
"mean": np.array([0.0, 0.0, 0.0]),
|
||||
"std": np.array([1.0, 2.0, 0.5]),
|
||||
}
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def action_stats_min_max():
|
||||
return {
|
||||
"min": np.array([-1.0, -2.0, 0.0]),
|
||||
"max": np.array([1.0, 2.0, 1.0]),
|
||||
}
|
||||
|
||||
|
||||
def _create_action_features():
|
||||
return {
|
||||
"action": PolicyFeature(FeatureType.ACTION, (3,)),
|
||||
}
|
||||
|
||||
|
||||
def _create_action_norm_map_mean_std():
|
||||
return {
|
||||
FeatureType.ACTION: NormalizationMode.MEAN_STD,
|
||||
}
|
||||
|
||||
|
||||
def _create_action_norm_map_min_max():
|
||||
return {
|
||||
FeatureType.ACTION: NormalizationMode.MIN_MAX,
|
||||
}
|
||||
|
||||
|
||||
def test_mean_std_unnormalization(action_stats_mean_std):
|
||||
features = _create_action_features()
|
||||
norm_map = _create_action_norm_map_mean_std()
|
||||
unnormalizer = UnnormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats={"action": action_stats_mean_std}
|
||||
)
|
||||
|
||||
normalized_action = torch.tensor([1.0, -0.5, 2.0])
|
||||
transition = create_transition(action=normalized_action)
|
||||
|
||||
unnormalized_transition = unnormalizer(transition)
|
||||
unnormalized_action = unnormalized_transition[TransitionKey.ACTION]
|
||||
|
||||
# action * std + mean
|
||||
expected = torch.tensor([1.0 * 1.0 + 0.0, -0.5 * 2.0 + 0.0, 2.0 * 0.5 + 0.0])
|
||||
assert torch.allclose(unnormalized_action, expected)
|
||||
|
||||
|
||||
def test_min_max_unnormalization(action_stats_min_max):
|
||||
features = _create_action_features()
|
||||
norm_map = _create_action_norm_map_min_max()
|
||||
unnormalizer = UnnormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats={"action": action_stats_min_max}
|
||||
)
|
||||
|
||||
# Actions in [-1, 1]
|
||||
normalized_action = torch.tensor([0.0, -1.0, 1.0])
|
||||
transition = create_transition(action=normalized_action)
|
||||
|
||||
unnormalized_transition = unnormalizer(transition)
|
||||
unnormalized_action = unnormalized_transition[TransitionKey.ACTION]
|
||||
|
||||
# Map from [-1, 1] to [min, max]
|
||||
# (action + 1) / 2 * (max - min) + min
|
||||
expected = torch.tensor(
|
||||
[
|
||||
(0.0 + 1) / 2 * (1.0 - (-1.0)) + (-1.0), # 0.0
|
||||
(-1.0 + 1) / 2 * (2.0 - (-2.0)) + (-2.0), # -2.0
|
||||
(1.0 + 1) / 2 * (1.0 - 0.0) + 0.0, # 1.0
|
||||
]
|
||||
)
|
||||
assert torch.allclose(unnormalized_action, expected)
|
||||
|
||||
|
||||
def test_numpy_action_input(action_stats_mean_std):
|
||||
features = _create_action_features()
|
||||
norm_map = _create_action_norm_map_mean_std()
|
||||
unnormalizer = UnnormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats={"action": action_stats_mean_std}
|
||||
)
|
||||
|
||||
normalized_action = np.array([1.0, -0.5, 2.0], dtype=np.float32)
|
||||
transition = create_transition(action=normalized_action)
|
||||
|
||||
unnormalized_transition = unnormalizer(transition)
|
||||
unnormalized_action = unnormalized_transition[TransitionKey.ACTION]
|
||||
|
||||
assert isinstance(unnormalized_action, torch.Tensor)
|
||||
expected = torch.tensor([1.0, -1.0, 1.0])
|
||||
assert torch.allclose(unnormalized_action, expected)
|
||||
|
||||
|
||||
def test_none_action(action_stats_mean_std):
|
||||
features = _create_action_features()
|
||||
norm_map = _create_action_norm_map_mean_std()
|
||||
unnormalizer = UnnormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats={"action": action_stats_mean_std}
|
||||
)
|
||||
|
||||
transition = create_transition()
|
||||
result = unnormalizer(transition)
|
||||
|
||||
# Should return transition unchanged
|
||||
assert result == transition
|
||||
|
||||
|
||||
def test_action_from_lerobot_dataset():
|
||||
mock_dataset = Mock()
|
||||
mock_dataset.meta.stats = {"action": {"mean": [0.0], "std": [1.0]}}
|
||||
features = {"action": PolicyFeature(FeatureType.ACTION, (1,))}
|
||||
norm_map = {FeatureType.ACTION: NormalizationMode.MEAN_STD}
|
||||
unnormalizer = UnnormalizerProcessor.from_lerobot_dataset(mock_dataset, features, norm_map)
|
||||
assert "mean" in unnormalizer._tensor_stats["action"]
|
||||
|
||||
|
||||
# Fixtures for NormalizerProcessor tests
|
||||
@pytest.fixture
|
||||
def full_stats():
|
||||
return {
|
||||
"observation.image": {
|
||||
"mean": np.array([0.5, 0.5, 0.5]),
|
||||
"std": np.array([0.2, 0.2, 0.2]),
|
||||
},
|
||||
"observation.state": {
|
||||
"min": np.array([0.0, -1.0]),
|
||||
"max": np.array([1.0, 1.0]),
|
||||
},
|
||||
"action": {
|
||||
"mean": np.array([0.0, 0.0]),
|
||||
"std": np.array([1.0, 2.0]),
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def _create_full_features():
|
||||
return {
|
||||
"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96)),
|
||||
"observation.state": PolicyFeature(FeatureType.STATE, (2,)),
|
||||
"action": PolicyFeature(FeatureType.ACTION, (2,)),
|
||||
}
|
||||
|
||||
|
||||
def _create_full_norm_map():
|
||||
return {
|
||||
FeatureType.VISUAL: NormalizationMode.MEAN_STD,
|
||||
FeatureType.STATE: NormalizationMode.MIN_MAX,
|
||||
FeatureType.ACTION: NormalizationMode.MEAN_STD,
|
||||
}
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def normalizer_processor(full_stats):
|
||||
features = _create_full_features()
|
||||
norm_map = _create_full_norm_map()
|
||||
return NormalizerProcessor(features=features, norm_map=norm_map, stats=full_stats)
|
||||
|
||||
|
||||
def test_combined_normalization(normalizer_processor):
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]),
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
action = torch.tensor([1.0, -0.5])
|
||||
transition = create_transition(
|
||||
observation=observation,
|
||||
action=action,
|
||||
reward=1.0,
|
||||
done=False,
|
||||
truncated=False,
|
||||
info={},
|
||||
complementary_data={},
|
||||
)
|
||||
|
||||
processed_transition = normalizer_processor(transition)
|
||||
|
||||
# Check normalized observations
|
||||
processed_obs = processed_transition[TransitionKey.OBSERVATION]
|
||||
expected_image = (torch.tensor([0.7, 0.5, 0.3]) - 0.5) / 0.2
|
||||
assert torch.allclose(processed_obs["observation.image"], expected_image)
|
||||
|
||||
# Check normalized action
|
||||
processed_action = processed_transition[TransitionKey.ACTION]
|
||||
expected_action = torch.tensor([(1.0 - 0.0) / 1.0, (-0.5 - 0.0) / 2.0])
|
||||
assert torch.allclose(processed_action, expected_action)
|
||||
|
||||
# Check other fields remain unchanged
|
||||
assert processed_transition[TransitionKey.REWARD] == 1.0
|
||||
assert not processed_transition[TransitionKey.DONE]
|
||||
|
||||
|
||||
def test_processor_from_lerobot_dataset(full_stats):
|
||||
# Mock dataset
|
||||
mock_dataset = Mock()
|
||||
mock_dataset.meta.stats = full_stats
|
||||
|
||||
features = _create_full_features()
|
||||
norm_map = _create_full_norm_map()
|
||||
|
||||
processor = NormalizerProcessor.from_lerobot_dataset(
|
||||
mock_dataset, features, norm_map, normalize_keys={"observation.image"}
|
||||
)
|
||||
|
||||
assert processor.normalize_keys == {"observation.image"}
|
||||
assert "observation.image" in processor._tensor_stats
|
||||
assert "action" in processor._tensor_stats
|
||||
|
||||
|
||||
def test_get_config(full_stats):
|
||||
features = _create_full_features()
|
||||
norm_map = _create_full_norm_map()
|
||||
processor = NormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats=full_stats, normalize_keys={"observation.image"}, eps=1e-6
|
||||
)
|
||||
|
||||
config = processor.get_config()
|
||||
expected_config = {
|
||||
"normalize_keys": ["observation.image"],
|
||||
"eps": 1e-6,
|
||||
"features": {
|
||||
"observation.image": {"type": "VISUAL", "shape": (3, 96, 96)},
|
||||
"observation.state": {"type": "STATE", "shape": (2,)},
|
||||
"action": {"type": "ACTION", "shape": (2,)},
|
||||
},
|
||||
"norm_map": {
|
||||
"VISUAL": "MEAN_STD",
|
||||
"STATE": "MIN_MAX",
|
||||
"ACTION": "MEAN_STD",
|
||||
},
|
||||
}
|
||||
assert config == expected_config
|
||||
|
||||
|
||||
def test_integration_with_robot_processor(normalizer_processor):
|
||||
"""Test integration with RobotProcessor pipeline"""
|
||||
robot_processor = RobotProcessor([normalizer_processor])
|
||||
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]),
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
action = torch.tensor([1.0, -0.5])
|
||||
transition = create_transition(
|
||||
observation=observation,
|
||||
action=action,
|
||||
reward=1.0,
|
||||
done=False,
|
||||
truncated=False,
|
||||
info={},
|
||||
complementary_data={},
|
||||
)
|
||||
|
||||
processed_transition = robot_processor(transition)
|
||||
|
||||
# Verify the processing worked
|
||||
assert isinstance(processed_transition[TransitionKey.OBSERVATION], dict)
|
||||
assert isinstance(processed_transition[TransitionKey.ACTION], torch.Tensor)
|
||||
|
||||
|
||||
# Edge case tests
|
||||
def test_empty_observation():
|
||||
stats = {"observation.image": {"mean": [0.5], "std": [0.2]}}
|
||||
features = {"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96))}
|
||||
norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}
|
||||
normalizer = NormalizerProcessor(features=features, norm_map=norm_map, stats=stats)
|
||||
|
||||
transition = create_transition()
|
||||
result = normalizer(transition)
|
||||
|
||||
assert result == transition
|
||||
|
||||
|
||||
def test_empty_stats():
|
||||
features = {"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96))}
|
||||
norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}
|
||||
normalizer = NormalizerProcessor(features=features, norm_map=norm_map, stats={})
|
||||
observation = {"observation.image": torch.tensor([0.5])}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = normalizer(transition)
|
||||
# Should return observation unchanged since no stats are available
|
||||
assert torch.allclose(
|
||||
result[TransitionKey.OBSERVATION]["observation.image"], observation["observation.image"]
|
||||
)
|
||||
|
||||
|
||||
def test_partial_stats():
|
||||
"""If statistics are incomplete, the value should pass through unchanged."""
|
||||
stats = {"observation.image": {"mean": [0.5]}} # Missing std / (min,max)
|
||||
features = {"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96))}
|
||||
norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}
|
||||
normalizer = NormalizerProcessor(features=features, norm_map=norm_map, stats=stats)
|
||||
observation = {"observation.image": torch.tensor([0.7])}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
processed = normalizer(transition)[TransitionKey.OBSERVATION]
|
||||
assert torch.allclose(processed["observation.image"], observation["observation.image"])
|
||||
|
||||
|
||||
def test_missing_action_stats_no_error():
|
||||
mock_dataset = Mock()
|
||||
mock_dataset.meta.stats = {"observation.image": {"mean": [0.5], "std": [0.2]}}
|
||||
|
||||
features = {"observation.image": PolicyFeature(FeatureType.VISUAL, (3, 96, 96))}
|
||||
norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}
|
||||
|
||||
processor = UnnormalizerProcessor.from_lerobot_dataset(mock_dataset, features, norm_map)
|
||||
# The tensor stats should not contain the 'action' key
|
||||
assert "action" not in processor._tensor_stats
|
||||
|
||||
|
||||
def test_serialization_roundtrip(full_stats):
|
||||
"""Test that features and norm_map can be serialized and deserialized correctly."""
|
||||
features = _create_full_features()
|
||||
norm_map = _create_full_norm_map()
|
||||
original_processor = NormalizerProcessor(
|
||||
features=features, norm_map=norm_map, stats=full_stats, normalize_keys={"observation.image"}, eps=1e-6
|
||||
)
|
||||
|
||||
# Get config (serialization)
|
||||
config = original_processor.get_config()
|
||||
|
||||
# Create a new processor from the config (deserialization)
|
||||
new_processor = NormalizerProcessor(
|
||||
features=config["features"],
|
||||
norm_map=config["norm_map"],
|
||||
stats=full_stats,
|
||||
normalize_keys=set(config["normalize_keys"]),
|
||||
eps=config["eps"],
|
||||
)
|
||||
|
||||
# Test that both processors work the same way
|
||||
observation = {
|
||||
"observation.image": torch.tensor([0.7, 0.5, 0.3]),
|
||||
"observation.state": torch.tensor([0.5, 0.0]),
|
||||
}
|
||||
action = torch.tensor([1.0, -0.5])
|
||||
transition = create_transition(
|
||||
observation=observation,
|
||||
action=action,
|
||||
reward=1.0,
|
||||
done=False,
|
||||
truncated=False,
|
||||
info={},
|
||||
complementary_data={},
|
||||
)
|
||||
|
||||
result1 = original_processor(transition)
|
||||
result2 = new_processor(transition)
|
||||
|
||||
# Compare results
|
||||
assert torch.allclose(
|
||||
result1[TransitionKey.OBSERVATION]["observation.image"],
|
||||
result2[TransitionKey.OBSERVATION]["observation.image"],
|
||||
)
|
||||
assert torch.allclose(result1[TransitionKey.ACTION], result2[TransitionKey.ACTION])
|
||||
|
||||
# Verify features and norm_map are correctly reconstructed
|
||||
assert new_processor.features.keys() == original_processor.features.keys()
|
||||
for key in new_processor.features:
|
||||
assert new_processor.features[key].type == original_processor.features[key].type
|
||||
assert new_processor.features[key].shape == original_processor.features[key].shape
|
||||
|
||||
assert new_processor.norm_map == original_processor.norm_map
|
||||
@@ -1,486 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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.
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import FeatureType
|
||||
from lerobot.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
|
||||
from lerobot.processor import VanillaObservationProcessor
|
||||
from lerobot.processor.pipeline import TransitionKey
|
||||
from tests.conftest import assert_contract_is_typed
|
||||
|
||||
|
||||
def create_transition(
|
||||
observation=None, action=None, reward=None, done=None, truncated=None, info=None, complementary_data=None
|
||||
):
|
||||
"""Helper to create an EnvTransition dictionary."""
|
||||
return {
|
||||
TransitionKey.OBSERVATION: observation,
|
||||
TransitionKey.ACTION: action,
|
||||
TransitionKey.REWARD: reward,
|
||||
TransitionKey.DONE: done,
|
||||
TransitionKey.TRUNCATED: truncated,
|
||||
TransitionKey.INFO: info,
|
||||
TransitionKey.COMPLEMENTARY_DATA: complementary_data,
|
||||
}
|
||||
|
||||
|
||||
def test_process_single_image():
|
||||
"""Test processing a single image."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create a mock image (H, W, C) format, uint8
|
||||
image = np.random.randint(0, 256, size=(64, 64, 3), dtype=np.uint8)
|
||||
|
||||
observation = {"pixels": image}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that the image was processed correctly
|
||||
assert "observation.image" in processed_obs
|
||||
processed_img = processed_obs["observation.image"]
|
||||
|
||||
# Check shape: should be (1, 3, 64, 64) - batch, channels, height, width
|
||||
assert processed_img.shape == (1, 3, 64, 64)
|
||||
|
||||
# Check dtype and range
|
||||
assert processed_img.dtype == torch.float32
|
||||
assert processed_img.min() >= 0.0
|
||||
assert processed_img.max() <= 1.0
|
||||
|
||||
|
||||
def test_process_image_dict():
|
||||
"""Test processing multiple images in a dictionary."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create mock images
|
||||
image1 = np.random.randint(0, 256, size=(32, 32, 3), dtype=np.uint8)
|
||||
image2 = np.random.randint(0, 256, size=(48, 48, 3), dtype=np.uint8)
|
||||
|
||||
observation = {"pixels": {"camera1": image1, "camera2": image2}}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that both images were processed
|
||||
assert "observation.images.camera1" in processed_obs
|
||||
assert "observation.images.camera2" in processed_obs
|
||||
|
||||
# Check shapes
|
||||
assert processed_obs["observation.images.camera1"].shape == (1, 3, 32, 32)
|
||||
assert processed_obs["observation.images.camera2"].shape == (1, 3, 48, 48)
|
||||
|
||||
|
||||
def test_process_batched_image():
|
||||
"""Test processing already batched images."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create a batched image (B, H, W, C)
|
||||
image = np.random.randint(0, 256, size=(2, 64, 64, 3), dtype=np.uint8)
|
||||
|
||||
observation = {"pixels": image}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that batch dimension is preserved
|
||||
assert processed_obs["observation.image"].shape == (2, 3, 64, 64)
|
||||
|
||||
|
||||
def test_invalid_image_format():
|
||||
"""Test error handling for invalid image formats."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Test wrong channel order (channels first)
|
||||
image = np.random.randint(0, 256, size=(3, 64, 64), dtype=np.uint8)
|
||||
observation = {"pixels": image}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
with pytest.raises(ValueError, match="Expected channel-last images"):
|
||||
processor(transition)
|
||||
|
||||
|
||||
def test_invalid_image_dtype():
|
||||
"""Test error handling for invalid image dtype."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Test wrong dtype
|
||||
image = np.random.rand(64, 64, 3).astype(np.float32)
|
||||
observation = {"pixels": image}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
with pytest.raises(ValueError, match="Expected torch.uint8 images"):
|
||||
processor(transition)
|
||||
|
||||
|
||||
def test_no_pixels_in_observation():
|
||||
"""Test processor when no pixels are in observation."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
observation = {"other_data": np.array([1, 2, 3])}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Should preserve other data unchanged
|
||||
assert "other_data" in processed_obs
|
||||
np.testing.assert_array_equal(processed_obs["other_data"], np.array([1, 2, 3]))
|
||||
|
||||
|
||||
def test_none_observation():
|
||||
"""Test processor with None observation."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
transition = create_transition()
|
||||
result = processor(transition)
|
||||
|
||||
assert result == transition
|
||||
|
||||
|
||||
def test_serialization_methods():
|
||||
"""Test serialization methods."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Test get_config
|
||||
config = processor.get_config()
|
||||
assert isinstance(config, dict)
|
||||
|
||||
# Test state_dict
|
||||
state = processor.state_dict()
|
||||
assert isinstance(state, dict)
|
||||
|
||||
# Test load_state_dict (should not raise)
|
||||
processor.load_state_dict(state)
|
||||
|
||||
# Test reset (should not raise)
|
||||
processor.reset()
|
||||
|
||||
|
||||
def test_process_environment_state():
|
||||
"""Test processing environment_state."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
env_state = np.array([1.0, 2.0, 3.0], dtype=np.float32)
|
||||
observation = {"environment_state": env_state}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that environment_state was renamed and processed
|
||||
assert "observation.environment_state" in processed_obs
|
||||
assert "environment_state" not in processed_obs
|
||||
|
||||
processed_state = processed_obs["observation.environment_state"]
|
||||
assert processed_state.shape == (1, 3) # Batch dimension added
|
||||
assert processed_state.dtype == torch.float32
|
||||
torch.testing.assert_close(processed_state, torch.tensor([[1.0, 2.0, 3.0]]))
|
||||
|
||||
|
||||
def test_process_agent_pos():
|
||||
"""Test processing agent_pos."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
agent_pos = np.array([0.5, -0.5, 1.0], dtype=np.float32)
|
||||
observation = {"agent_pos": agent_pos}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that agent_pos was renamed and processed
|
||||
assert "observation.state" in processed_obs
|
||||
assert "agent_pos" not in processed_obs
|
||||
|
||||
processed_state = processed_obs["observation.state"]
|
||||
assert processed_state.shape == (1, 3) # Batch dimension added
|
||||
assert processed_state.dtype == torch.float32
|
||||
torch.testing.assert_close(processed_state, torch.tensor([[0.5, -0.5, 1.0]]))
|
||||
|
||||
|
||||
def test_process_batched_states():
|
||||
"""Test processing already batched states."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
env_state = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
|
||||
agent_pos = np.array([[0.5, -0.5], [1.0, -1.0]], dtype=np.float32)
|
||||
|
||||
observation = {"environment_state": env_state, "agent_pos": agent_pos}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that batch dimensions are preserved
|
||||
assert processed_obs["observation.environment_state"].shape == (2, 2)
|
||||
assert processed_obs["observation.state"].shape == (2, 2)
|
||||
|
||||
|
||||
def test_process_both_states():
|
||||
"""Test processing both environment_state and agent_pos."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
env_state = np.array([1.0, 2.0], dtype=np.float32)
|
||||
agent_pos = np.array([0.5, -0.5], dtype=np.float32)
|
||||
|
||||
observation = {"environment_state": env_state, "agent_pos": agent_pos, "other_data": "keep_me"}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that both states were processed
|
||||
assert "observation.environment_state" in processed_obs
|
||||
assert "observation.state" in processed_obs
|
||||
|
||||
# Check that original keys were removed
|
||||
assert "environment_state" not in processed_obs
|
||||
assert "agent_pos" not in processed_obs
|
||||
|
||||
# Check that other data was preserved
|
||||
assert processed_obs["other_data"] == "keep_me"
|
||||
|
||||
|
||||
def test_no_states_in_observation():
|
||||
"""Test processor when no states are in observation."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
observation = {"other_data": np.array([1, 2, 3])}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Should preserve data unchanged
|
||||
np.testing.assert_array_equal(processed_obs, observation)
|
||||
|
||||
|
||||
def test_complete_observation_processing():
|
||||
"""Test processing a complete observation with both images and states."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create mock data
|
||||
image = np.random.randint(0, 256, size=(32, 32, 3), dtype=np.uint8)
|
||||
env_state = np.array([1.0, 2.0, 3.0], dtype=np.float32)
|
||||
agent_pos = np.array([0.5, -0.5, 1.0], dtype=np.float32)
|
||||
|
||||
observation = {
|
||||
"pixels": image,
|
||||
"environment_state": env_state,
|
||||
"agent_pos": agent_pos,
|
||||
"other_data": "preserve_me",
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that image was processed
|
||||
assert "observation.image" in processed_obs
|
||||
assert processed_obs["observation.image"].shape == (1, 3, 32, 32)
|
||||
|
||||
# Check that states were processed
|
||||
assert "observation.environment_state" in processed_obs
|
||||
assert "observation.state" in processed_obs
|
||||
|
||||
# Check that original keys were removed
|
||||
assert "pixels" not in processed_obs
|
||||
assert "environment_state" not in processed_obs
|
||||
assert "agent_pos" not in processed_obs
|
||||
|
||||
# Check that other data was preserved
|
||||
assert processed_obs["other_data"] == "preserve_me"
|
||||
|
||||
|
||||
def test_image_only_processing():
|
||||
"""Test processing observation with only images."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
image = np.random.randint(0, 256, size=(64, 64, 3), dtype=np.uint8)
|
||||
observation = {"pixels": image}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
assert "observation.image" in processed_obs
|
||||
assert len(processed_obs) == 1
|
||||
|
||||
|
||||
def test_state_only_processing():
|
||||
"""Test processing observation with only states."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
agent_pos = np.array([1.0, 2.0], dtype=np.float32)
|
||||
observation = {"agent_pos": agent_pos}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
assert "observation.state" in processed_obs
|
||||
assert "agent_pos" not in processed_obs
|
||||
|
||||
|
||||
def test_empty_observation():
|
||||
"""Test processing empty observation."""
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
observation = {}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
assert processed_obs == {}
|
||||
|
||||
|
||||
def test_equivalent_to_original_function():
|
||||
"""Test that ObservationProcessor produces equivalent results to preprocess_observation."""
|
||||
# Import the original function for comparison
|
||||
from lerobot.envs.utils import preprocess_observation
|
||||
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create test data similar to what the original function expects
|
||||
image = np.random.randint(0, 256, size=(64, 64, 3), dtype=np.uint8)
|
||||
env_state = np.array([1.0, 2.0, 3.0], dtype=np.float32)
|
||||
agent_pos = np.array([0.5, -0.5, 1.0], dtype=np.float32)
|
||||
|
||||
observation = {"pixels": image, "environment_state": env_state, "agent_pos": agent_pos}
|
||||
|
||||
# Process with original function
|
||||
original_result = preprocess_observation(observation)
|
||||
|
||||
# Process with new processor
|
||||
transition = create_transition(observation=observation)
|
||||
processor_result = processor(transition)[TransitionKey.OBSERVATION]
|
||||
|
||||
# Compare results
|
||||
assert set(original_result.keys()) == set(processor_result.keys())
|
||||
|
||||
for key in original_result:
|
||||
torch.testing.assert_close(original_result[key], processor_result[key])
|
||||
|
||||
|
||||
def test_equivalent_with_image_dict():
|
||||
"""Test equivalence with dictionary of images."""
|
||||
from lerobot.envs.utils import preprocess_observation
|
||||
|
||||
processor = VanillaObservationProcessor()
|
||||
|
||||
# Create test data with multiple cameras
|
||||
image1 = np.random.randint(0, 256, size=(32, 32, 3), dtype=np.uint8)
|
||||
image2 = np.random.randint(0, 256, size=(48, 48, 3), dtype=np.uint8)
|
||||
agent_pos = np.array([1.0, 2.0], dtype=np.float32)
|
||||
|
||||
observation = {"pixels": {"cam1": image1, "cam2": image2}, "agent_pos": agent_pos}
|
||||
|
||||
# Process with original function
|
||||
original_result = preprocess_observation(observation)
|
||||
|
||||
# Process with new processor
|
||||
transition = create_transition(observation=observation)
|
||||
processor_result = processor(transition)[TransitionKey.OBSERVATION]
|
||||
|
||||
# Compare results
|
||||
assert set(original_result.keys()) == set(processor_result.keys())
|
||||
|
||||
for key in original_result:
|
||||
torch.testing.assert_close(original_result[key], processor_result[key])
|
||||
|
||||
|
||||
def test_image_processor_feature_contract_pixels_to_image(policy_feature_factory):
|
||||
processor = VanillaObservationProcessor()
|
||||
features = {
|
||||
"pixels": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"keep": policy_feature_factory(FeatureType.ENV, (1,)),
|
||||
}
|
||||
out = processor.feature_contract(features.copy())
|
||||
|
||||
assert OBS_IMAGE in out and out[OBS_IMAGE] == features["pixels"]
|
||||
assert "pixels" not in out
|
||||
assert out["keep"] == features["keep"]
|
||||
assert_contract_is_typed(out)
|
||||
|
||||
|
||||
def test_image_processor_feature_contract_observation_pixels_to_image(policy_feature_factory):
|
||||
processor = VanillaObservationProcessor()
|
||||
features = {
|
||||
"observation.pixels": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"keep": policy_feature_factory(FeatureType.ENV, (1,)),
|
||||
}
|
||||
out = processor.feature_contract(features.copy())
|
||||
|
||||
assert OBS_IMAGE in out and out[OBS_IMAGE] == features["observation.pixels"]
|
||||
assert "observation.pixels" not in out
|
||||
assert out["keep"] == features["keep"]
|
||||
assert_contract_is_typed(out)
|
||||
|
||||
|
||||
def test_image_processor_feature_contract_multi_camera_and_prefixed(policy_feature_factory):
|
||||
processor = VanillaObservationProcessor()
|
||||
features = {
|
||||
"pixels.front": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"pixels.wrist": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"observation.pixels.rear": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"keep": policy_feature_factory(FeatureType.ENV, (7,)),
|
||||
}
|
||||
out = processor.feature_contract(features.copy())
|
||||
|
||||
assert f"{OBS_IMAGES}.front" in out and out[f"{OBS_IMAGES}.front"] == features["pixels.front"]
|
||||
assert f"{OBS_IMAGES}.wrist" in out and out[f"{OBS_IMAGES}.wrist"] == features["pixels.wrist"]
|
||||
assert f"{OBS_IMAGES}.rear" in out and out[f"{OBS_IMAGES}.rear"] == features["observation.pixels.rear"]
|
||||
assert "pixels.front" not in out and "pixels.wrist" not in out and "observation.pixels.rear" not in out
|
||||
assert out["keep"] == features["keep"]
|
||||
assert_contract_is_typed(out)
|
||||
|
||||
|
||||
def test_state_processor_feature_contract_environment_and_agent_pos(policy_feature_factory):
|
||||
processor = VanillaObservationProcessor()
|
||||
features = {
|
||||
"environment_state": policy_feature_factory(FeatureType.STATE, (3,)),
|
||||
"agent_pos": policy_feature_factory(FeatureType.STATE, (7,)),
|
||||
"keep": policy_feature_factory(FeatureType.ENV, (1,)),
|
||||
}
|
||||
out = processor.feature_contract(features.copy())
|
||||
|
||||
assert OBS_ENV_STATE in out and out[OBS_ENV_STATE] == features["environment_state"]
|
||||
assert OBS_STATE in out and out[OBS_STATE] == features["agent_pos"]
|
||||
assert "environment_state" not in out and "agent_pos" not in out
|
||||
assert out["keep"] == features["keep"]
|
||||
assert_contract_is_typed(out)
|
||||
|
||||
|
||||
def test_state_processor_feature_contract_prefixed_inputs(policy_feature_factory):
|
||||
proc = VanillaObservationProcessor()
|
||||
features = {
|
||||
"observation.environment_state": policy_feature_factory(FeatureType.STATE, (2,)),
|
||||
"observation.agent_pos": policy_feature_factory(FeatureType.STATE, (4,)),
|
||||
}
|
||||
out = proc.feature_contract(features.copy())
|
||||
|
||||
assert OBS_ENV_STATE in out and out[OBS_ENV_STATE] == features["observation.environment_state"]
|
||||
assert OBS_STATE in out and out[OBS_STATE] == features["observation.agent_pos"]
|
||||
assert "environment_state" not in out and "agent_pos" not in out
|
||||
assert_contract_is_typed(out)
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,467 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# 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.
|
||||
import tempfile
|
||||
from pathlib import Path
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import FeatureType
|
||||
from lerobot.processor import ProcessorStepRegistry, RenameProcessor, RobotProcessor, TransitionKey
|
||||
from tests.conftest import assert_contract_is_typed
|
||||
|
||||
|
||||
def create_transition(
|
||||
observation=None, action=None, reward=None, done=None, truncated=None, info=None, complementary_data=None
|
||||
):
|
||||
"""Helper to create an EnvTransition dictionary."""
|
||||
return {
|
||||
TransitionKey.OBSERVATION: observation,
|
||||
TransitionKey.ACTION: action,
|
||||
TransitionKey.REWARD: reward,
|
||||
TransitionKey.DONE: done,
|
||||
TransitionKey.TRUNCATED: truncated,
|
||||
TransitionKey.INFO: info,
|
||||
TransitionKey.COMPLEMENTARY_DATA: complementary_data,
|
||||
}
|
||||
|
||||
|
||||
def test_basic_renaming():
|
||||
"""Test basic key renaming functionality."""
|
||||
rename_map = {
|
||||
"old_key1": "new_key1",
|
||||
"old_key2": "new_key2",
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
observation = {
|
||||
"old_key1": torch.tensor([1.0, 2.0]),
|
||||
"old_key2": np.array([3.0, 4.0]),
|
||||
"unchanged_key": "keep_me",
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check renamed keys
|
||||
assert "new_key1" in processed_obs
|
||||
assert "new_key2" in processed_obs
|
||||
assert "old_key1" not in processed_obs
|
||||
assert "old_key2" not in processed_obs
|
||||
|
||||
# Check values are preserved
|
||||
torch.testing.assert_close(processed_obs["new_key1"], torch.tensor([1.0, 2.0]))
|
||||
np.testing.assert_array_equal(processed_obs["new_key2"], np.array([3.0, 4.0]))
|
||||
|
||||
# Check unchanged key is preserved
|
||||
assert processed_obs["unchanged_key"] == "keep_me"
|
||||
|
||||
|
||||
def test_empty_rename_map():
|
||||
"""Test processor with empty rename map (should pass through unchanged)."""
|
||||
processor = RenameProcessor(rename_map={})
|
||||
|
||||
observation = {
|
||||
"key1": torch.tensor([1.0]),
|
||||
"key2": "value2",
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# All keys should be unchanged
|
||||
assert processed_obs.keys() == observation.keys()
|
||||
torch.testing.assert_close(processed_obs["key1"], observation["key1"])
|
||||
assert processed_obs["key2"] == observation["key2"]
|
||||
|
||||
|
||||
def test_none_observation():
|
||||
"""Test processor with None observation."""
|
||||
processor = RenameProcessor(rename_map={"old": "new"})
|
||||
|
||||
transition = create_transition()
|
||||
result = processor(transition)
|
||||
|
||||
# Should return transition unchanged
|
||||
assert result == transition
|
||||
|
||||
|
||||
def test_overlapping_rename():
|
||||
"""Test renaming when new names might conflict."""
|
||||
rename_map = {
|
||||
"a": "b",
|
||||
"b": "c", # This creates a potential conflict
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
observation = {
|
||||
"a": 1,
|
||||
"b": 2,
|
||||
"x": 3,
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that renaming happens correctly
|
||||
assert "a" not in processed_obs
|
||||
assert processed_obs["b"] == 1 # 'a' renamed to 'b'
|
||||
assert processed_obs["c"] == 2 # original 'b' renamed to 'c'
|
||||
assert processed_obs["x"] == 3
|
||||
|
||||
|
||||
def test_partial_rename():
|
||||
"""Test renaming only some keys."""
|
||||
rename_map = {
|
||||
"observation.state": "observation.proprio_state",
|
||||
"pixels": "observation.image",
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
observation = {
|
||||
"observation.state": torch.randn(10),
|
||||
"pixels": np.random.randint(0, 256, (64, 64, 3), dtype=np.uint8),
|
||||
"reward": 1.0,
|
||||
"info": {"episode": 1},
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check renamed keys
|
||||
assert "observation.proprio_state" in processed_obs
|
||||
assert "observation.image" in processed_obs
|
||||
assert "observation.state" not in processed_obs
|
||||
assert "pixels" not in processed_obs
|
||||
|
||||
# Check unchanged keys
|
||||
assert processed_obs["reward"] == 1.0
|
||||
assert processed_obs["info"] == {"episode": 1}
|
||||
|
||||
|
||||
def test_get_config():
|
||||
"""Test configuration serialization."""
|
||||
rename_map = {
|
||||
"old1": "new1",
|
||||
"old2": "new2",
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
config = processor.get_config()
|
||||
assert config == {"rename_map": rename_map}
|
||||
|
||||
|
||||
def test_state_dict():
|
||||
"""Test state dict (should be empty for RenameProcessor)."""
|
||||
processor = RenameProcessor(rename_map={"old": "new"})
|
||||
|
||||
state = processor.state_dict()
|
||||
assert state == {}
|
||||
|
||||
# Load state dict should work even with empty dict
|
||||
processor.load_state_dict({})
|
||||
|
||||
|
||||
def test_integration_with_robot_processor():
|
||||
"""Test integration with RobotProcessor pipeline."""
|
||||
rename_map = {
|
||||
"agent_pos": "observation.state",
|
||||
"pixels": "observation.image",
|
||||
}
|
||||
rename_processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
pipeline = RobotProcessor([rename_processor])
|
||||
|
||||
observation = {
|
||||
"agent_pos": np.array([1.0, 2.0, 3.0]),
|
||||
"pixels": np.zeros((32, 32, 3), dtype=np.uint8),
|
||||
"other_data": "preserve_me",
|
||||
}
|
||||
transition = create_transition(
|
||||
observation=observation, reward=0.5, done=False, truncated=False, info={}, complementary_data={}
|
||||
)
|
||||
|
||||
result = pipeline(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check renaming worked through pipeline
|
||||
assert "observation.state" in processed_obs
|
||||
assert "observation.image" in processed_obs
|
||||
assert "agent_pos" not in processed_obs
|
||||
assert "pixels" not in processed_obs
|
||||
assert processed_obs["other_data"] == "preserve_me"
|
||||
|
||||
# Check other transition elements unchanged
|
||||
assert result[TransitionKey.REWARD] == 0.5
|
||||
assert result[TransitionKey.DONE] is False
|
||||
|
||||
|
||||
def test_save_and_load_pretrained():
|
||||
"""Test saving and loading processor with RobotProcessor."""
|
||||
rename_map = {
|
||||
"old_state": "observation.state",
|
||||
"old_image": "observation.image",
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
pipeline = RobotProcessor([processor], name="TestRenameProcessor")
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmp_dir:
|
||||
# Save pipeline
|
||||
pipeline.save_pretrained(tmp_dir)
|
||||
|
||||
# Check files were created
|
||||
config_path = Path(tmp_dir) / "testrenameprocessor.json" # Based on name="TestRenameProcessor"
|
||||
assert config_path.exists()
|
||||
|
||||
# No state files should be created for RenameProcessor
|
||||
state_files = list(Path(tmp_dir).glob("*.safetensors"))
|
||||
assert len(state_files) == 0
|
||||
|
||||
# Load pipeline
|
||||
loaded_pipeline = RobotProcessor.from_pretrained(tmp_dir)
|
||||
|
||||
assert loaded_pipeline.name == "TestRenameProcessor"
|
||||
assert len(loaded_pipeline) == 1
|
||||
|
||||
# Check that loaded processor works correctly
|
||||
loaded_processor = loaded_pipeline.steps[0]
|
||||
assert isinstance(loaded_processor, RenameProcessor)
|
||||
assert loaded_processor.rename_map == rename_map
|
||||
|
||||
# Test functionality after loading
|
||||
observation = {"old_state": [1, 2, 3], "old_image": "image_data"}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = loaded_pipeline(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
assert "observation.state" in processed_obs
|
||||
assert "observation.image" in processed_obs
|
||||
assert processed_obs["observation.state"] == [1, 2, 3]
|
||||
assert processed_obs["observation.image"] == "image_data"
|
||||
|
||||
|
||||
def test_registry_functionality():
|
||||
"""Test that RenameProcessor is properly registered."""
|
||||
# Check that it's registered
|
||||
assert "rename_processor" in ProcessorStepRegistry.list()
|
||||
|
||||
# Get from registry
|
||||
retrieved_class = ProcessorStepRegistry.get("rename_processor")
|
||||
assert retrieved_class is RenameProcessor
|
||||
|
||||
# Create instance from registry
|
||||
instance = retrieved_class(rename_map={"old": "new"})
|
||||
assert isinstance(instance, RenameProcessor)
|
||||
assert instance.rename_map == {"old": "new"}
|
||||
|
||||
|
||||
def test_registry_based_save_load():
|
||||
"""Test save/load using registry name instead of module path."""
|
||||
processor = RenameProcessor(rename_map={"key1": "renamed_key1"})
|
||||
pipeline = RobotProcessor([processor])
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmp_dir:
|
||||
# Save and load
|
||||
pipeline.save_pretrained(tmp_dir)
|
||||
|
||||
# Verify config uses registry name
|
||||
import json
|
||||
|
||||
with open(Path(tmp_dir) / "robotprocessor.json") as f: # Default name is "RobotProcessor"
|
||||
config = json.load(f)
|
||||
|
||||
assert "registry_name" in config["steps"][0]
|
||||
assert config["steps"][0]["registry_name"] == "rename_processor"
|
||||
assert "class" not in config["steps"][0] # Should use registry, not module path
|
||||
|
||||
# Load should work
|
||||
loaded_pipeline = RobotProcessor.from_pretrained(tmp_dir)
|
||||
loaded_processor = loaded_pipeline.steps[0]
|
||||
assert isinstance(loaded_processor, RenameProcessor)
|
||||
assert loaded_processor.rename_map == {"key1": "renamed_key1"}
|
||||
|
||||
|
||||
def test_chained_rename_processors():
|
||||
"""Test multiple RenameProcessors in a pipeline."""
|
||||
# First processor: rename raw keys to intermediate format
|
||||
processor1 = RenameProcessor(
|
||||
rename_map={
|
||||
"pos": "agent_position",
|
||||
"img": "camera_image",
|
||||
}
|
||||
)
|
||||
|
||||
# Second processor: rename to final format
|
||||
processor2 = RenameProcessor(
|
||||
rename_map={
|
||||
"agent_position": "observation.state",
|
||||
"camera_image": "observation.image",
|
||||
}
|
||||
)
|
||||
|
||||
pipeline = RobotProcessor([processor1, processor2])
|
||||
|
||||
observation = {
|
||||
"pos": np.array([1.0, 2.0]),
|
||||
"img": "image_data",
|
||||
"extra": "keep_me",
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
# Step through to see intermediate results
|
||||
results = list(pipeline.step_through(transition))
|
||||
|
||||
# After first processor
|
||||
assert "agent_position" in results[1][TransitionKey.OBSERVATION]
|
||||
assert "camera_image" in results[1][TransitionKey.OBSERVATION]
|
||||
|
||||
# After second processor
|
||||
final_obs = results[2][TransitionKey.OBSERVATION]
|
||||
assert "observation.state" in final_obs
|
||||
assert "observation.image" in final_obs
|
||||
assert final_obs["extra"] == "keep_me"
|
||||
|
||||
# Original keys should be gone
|
||||
assert "pos" not in final_obs
|
||||
assert "img" not in final_obs
|
||||
assert "agent_position" not in final_obs
|
||||
assert "camera_image" not in final_obs
|
||||
|
||||
|
||||
def test_nested_observation_rename():
|
||||
"""Test renaming with nested observation structures."""
|
||||
rename_map = {
|
||||
"observation.images.left": "observation.camera.left_view",
|
||||
"observation.images.right": "observation.camera.right_view",
|
||||
"observation.proprio": "observation.proprioception",
|
||||
}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
observation = {
|
||||
"observation.images.left": torch.randn(3, 64, 64),
|
||||
"observation.images.right": torch.randn(3, 64, 64),
|
||||
"observation.proprio": torch.randn(7),
|
||||
"observation.gripper": torch.tensor([0.0]), # Not renamed
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check renames
|
||||
assert "observation.camera.left_view" in processed_obs
|
||||
assert "observation.camera.right_view" in processed_obs
|
||||
assert "observation.proprioception" in processed_obs
|
||||
|
||||
# Check unchanged key
|
||||
assert "observation.gripper" in processed_obs
|
||||
|
||||
# Check old keys removed
|
||||
assert "observation.images.left" not in processed_obs
|
||||
assert "observation.images.right" not in processed_obs
|
||||
assert "observation.proprio" not in processed_obs
|
||||
|
||||
|
||||
def test_value_types_preserved():
|
||||
"""Test that various value types are preserved during renaming."""
|
||||
rename_map = {"old_tensor": "new_tensor", "old_array": "new_array", "old_scalar": "new_scalar"}
|
||||
processor = RenameProcessor(rename_map=rename_map)
|
||||
|
||||
tensor_value = torch.randn(3, 3)
|
||||
array_value = np.random.rand(2, 2)
|
||||
|
||||
observation = {
|
||||
"old_tensor": tensor_value,
|
||||
"old_array": array_value,
|
||||
"old_scalar": 42,
|
||||
"old_string": "hello",
|
||||
"old_dict": {"nested": "value"},
|
||||
"old_list": [1, 2, 3],
|
||||
}
|
||||
transition = create_transition(observation=observation)
|
||||
|
||||
result = processor(transition)
|
||||
processed_obs = result[TransitionKey.OBSERVATION]
|
||||
|
||||
# Check that values and types are preserved
|
||||
assert torch.equal(processed_obs["new_tensor"], tensor_value)
|
||||
assert np.array_equal(processed_obs["new_array"], array_value)
|
||||
assert processed_obs["new_scalar"] == 42
|
||||
assert processed_obs["old_string"] == "hello"
|
||||
assert processed_obs["old_dict"] == {"nested": "value"}
|
||||
assert processed_obs["old_list"] == [1, 2, 3]
|
||||
|
||||
|
||||
def test_feature_contract_basic_renaming(policy_feature_factory):
|
||||
processor = RenameProcessor(rename_map={"a": "x", "b": "y"})
|
||||
features = {
|
||||
"a": policy_feature_factory(FeatureType.STATE, (2,)),
|
||||
"b": policy_feature_factory(FeatureType.ACTION, (3,)),
|
||||
"c": policy_feature_factory(FeatureType.ENV, (1,)),
|
||||
}
|
||||
|
||||
out = processor.feature_contract(features.copy())
|
||||
|
||||
# Values preserved and typed
|
||||
assert out["x"] == features["a"]
|
||||
assert out["y"] == features["b"]
|
||||
assert out["c"] == features["c"]
|
||||
|
||||
assert_contract_is_typed(out)
|
||||
# Input not mutated
|
||||
assert set(features) == {"a", "b", "c"}
|
||||
|
||||
|
||||
def test_feature_contract_overlapping_keys(policy_feature_factory):
|
||||
# Overlapping renames: both 'a' and 'b' exist. 'a'->'b', 'b'->'c'
|
||||
processor = RenameProcessor(rename_map={"a": "b", "b": "c"})
|
||||
features = {
|
||||
"a": policy_feature_factory(FeatureType.STATE, (1,)),
|
||||
"b": policy_feature_factory(FeatureType.STATE, (2,)),
|
||||
}
|
||||
out = processor.feature_contract(features)
|
||||
|
||||
assert set(out) == {"b", "c"}
|
||||
assert out["b"] == features["a"] # 'a' renamed to'b'
|
||||
assert out["c"] == features["b"] # 'b' renamed to 'c'
|
||||
assert_contract_is_typed(out)
|
||||
|
||||
|
||||
def test_feature_contract_chained_processors(policy_feature_factory):
|
||||
# Chain two rename processors at the contract level
|
||||
processor1 = RenameProcessor(rename_map={"pos": "agent_position", "img": "camera_image"})
|
||||
processor2 = RenameProcessor(
|
||||
rename_map={"agent_position": "observation.state", "camera_image": "observation.image"}
|
||||
)
|
||||
pipeline = RobotProcessor([processor1, processor2])
|
||||
|
||||
spec = {
|
||||
"pos": policy_feature_factory(FeatureType.STATE, (7,)),
|
||||
"img": policy_feature_factory(FeatureType.VISUAL, (3, 64, 64)),
|
||||
"extra": policy_feature_factory(FeatureType.ENV, (1,)),
|
||||
}
|
||||
out = pipeline.feature_contract(initial_features=spec)
|
||||
|
||||
assert set(out) == {"observation.state", "observation.image", "extra"}
|
||||
assert out["observation.state"] == spec["pos"]
|
||||
assert out["observation.image"] == spec["img"]
|
||||
assert out["extra"] == spec["extra"]
|
||||
assert_contract_is_typed(out)
|
||||
Reference in New Issue
Block a user