admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-16 09:09:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Organize test folders (#856)

Browse Source 
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>
This commit is contained in:
Simon Alibert
2025-03-13 14:05:55 +01:00
committed by GitHub
parent a36ed39487
commit 974028bd28
79 changed files with 63 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/scripts/save_dataset_to_safetensors.py
-91
View File
@@ -1,91 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script provides a utility for saving a dataset as safetensors files for the purpose of testing backward compatibility
when updating the data format. It uses the `PushtDataset` to create a DataLoader and saves selected frame from the
dataset into a corresponding safetensors file in a specified output directory.
If you know that your change will break backward compatibility, you should write a shortlived test by modifying
`tests/test_datasets.py::test_backward_compatibility` accordingly, and make sure this custom test pass. Your custom test
doesnt need to be merged into the `main` branch. Then you need to run this script and update the tests artifacts.
Example usage:
`python tests/scripts/save_dataset_to_safetensors.py`
"""
import shutil
from pathlib import Path
from safetensors.torch import save_file
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
def save_dataset_to_safetensors(output_dir, repo_id="lerobot/pusht"):
repo_dir = Path(output_dir) / repo_id
if repo_dir.exists():
shutil.rmtree(repo_dir)
repo_dir.mkdir(parents=True, exist_ok=True)
dataset = LeRobotDataset(
repo_id=repo_id,
episodes=[0],
)
# save 2 first frames of first episode
i = dataset.episode_data_index["from"][0].item()
save_file(dataset[i], repo_dir / f"frame_{i}.safetensors")
save_file(dataset[i + 1], repo_dir / f"frame_{i + 1}.safetensors")
# save 2 frames at the middle of first episode
i = int((dataset.episode_data_index["to"][0].item() - dataset.episode_data_index["from"][0].item()) / 2)
save_file(dataset[i], repo_dir / f"frame_{i}.safetensors")
save_file(dataset[i + 1], repo_dir / f"frame_{i + 1}.safetensors")
# save 2 last frames of first episode
i = dataset.episode_data_index["to"][0].item()
save_file(dataset[i - 2], repo_dir / f"frame_{i - 2}.safetensors")
save_file(dataset[i - 1], repo_dir / f"frame_{i - 1}.safetensors")
# TODO(rcadene): Enable testing on second and last episode
# We currently cant because our test dataset only contains the first episode
# # save 2 first frames of second episode
# i = dataset.episode_data_index["from"][1].item()
# save_file(dataset[i], repo_dir / f"frame_{i}.safetensors")
# save_file(dataset[i + 1], repo_dir / f"frame_{i+1}.safetensors")
# # save 2 last frames of second episode
# i = dataset.episode_data_index["to"][1].item()
# save_file(dataset[i - 2], repo_dir / f"frame_{i-2}.safetensors")
# save_file(dataset[i - 1], repo_dir / f"frame_{i-1}.safetensors")
# # save 2 last frames of last episode
# i = dataset.episode_data_index["to"][-1].item()
# save_file(dataset[i - 2], repo_dir / f"frame_{i-2}.safetensors")
# save_file(dataset[i - 1], repo_dir / f"frame_{i-1}.safetensors")
if __name__ == "__main__":
for dataset in [
"lerobot/pusht",
"lerobot/aloha_sim_insertion_human",
"lerobot/xarm_lift_medium",
"lerobot/nyu_franka_play_dataset",
"lerobot/cmu_stretch",
]:
save_dataset_to_safetensors("tests/data/save_dataset_to_safetensors", repo_id=dataset)
tests/scripts/save_image_transforms_to_safetensors.py
-75
View File
@@ -1,75 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from pathlib import Path
import torch
from safetensors.torch import save_file
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
from lerobot.common.datasets.transforms import (
ImageTransformConfig,
ImageTransforms,
ImageTransformsConfig,
make_transform_from_config,
)
from lerobot.common.utils.random_utils import seeded_context
ARTIFACT_DIR = Path("tests/data/save_image_transforms_to_safetensors")
DATASET_REPO_ID = "lerobot/aloha_mobile_shrimp"
def save_default_config_transform(original_frame: torch.Tensor, output_dir: Path):
cfg = ImageTransformsConfig(enable=True)
default_tf = ImageTransforms(cfg)
with seeded_context(1337):
img_tf = default_tf(original_frame)
save_file({"default": img_tf}, output_dir / "default_transforms.safetensors")
def save_single_transforms(original_frame: torch.Tensor, output_dir: Path):
transforms = {
("ColorJitter", "brightness", [(0.5, 0.5), (2.0, 2.0)]),
("ColorJitter", "contrast", [(0.5, 0.5), (2.0, 2.0)]),
("ColorJitter", "saturation", [(0.5, 0.5), (2.0, 2.0)]),
("ColorJitter", "hue", [(-0.25, -0.25), (0.25, 0.25)]),
("SharpnessJitter", "sharpness", [(0.5, 0.5), (2.0, 2.0)]),
}
frames = {"original_frame": original_frame}
for tf_type, tf_name, min_max_values in transforms.items():
for min_max in min_max_values:
tf_cfg = ImageTransformConfig(type=tf_type, kwargs={tf_name: min_max})
tf = make_transform_from_config(tf_cfg)
key = f"{tf_name}_{min_max[0]}_{min_max[1]}"
frames[key] = tf(original_frame)
save_file(frames, output_dir / "single_transforms.safetensors")
def main():
dataset = LeRobotDataset(DATASET_REPO_ID, episodes=[0], image_transforms=None)
output_dir = Path(ARTIFACT_DIR)
output_dir.mkdir(parents=True, exist_ok=True)
original_frame = dataset[0][dataset.meta.camera_keys[0]]
save_single_transforms(original_frame, output_dir)
save_default_config_transform(original_frame, output_dir)
if __name__ == "__main__":
main()
tests/scripts/save_policy_to_safetensors.py
-145
View File
@@ -1,145 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import shutil
from pathlib import Path
import torch
from safetensors.torch import save_file
from lerobot.common.datasets.factory import make_dataset
from lerobot.common.optim.factory import make_optimizer_and_scheduler
from lerobot.common.policies.factory import make_policy, make_policy_config
from lerobot.common.utils.random_utils import set_seed
from lerobot.configs.default import DatasetConfig
from lerobot.configs.train import TrainPipelineConfig
def get_policy_stats(ds_repo_id: str, policy_name: str, policy_kwargs: dict):
set_seed(1337)
train_cfg = TrainPipelineConfig(
# TODO(rcadene, aliberts): remove dataset download
dataset=DatasetConfig(repo_id=ds_repo_id, episodes=[0]),
policy=make_policy_config(policy_name, **policy_kwargs),
)
train_cfg.validate() # Needed for auto-setting some parameters
dataset = make_dataset(train_cfg)
policy = make_policy(train_cfg.policy, ds_meta=dataset.meta)
policy.train()
optimizer, _ = make_optimizer_and_scheduler(train_cfg, policy)
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=0,
batch_size=train_cfg.batch_size,
shuffle=False,
)
batch = next(iter(dataloader))
loss, output_dict = policy.forward(batch)
if output_dict is not None:
output_dict = {k: v for k, v in output_dict.items() if isinstance(v, torch.Tensor)}
output_dict["loss"] = loss
else:
output_dict = {"loss": loss}
loss.backward()
grad_stats = {}
for key, param in policy.named_parameters():
if param.requires_grad:
grad_stats[f"{key}_mean"] = param.grad.mean()
grad_stats[f"{key}_std"] = (
param.grad.std() if param.grad.numel() > 1 else torch.tensor(float(0.0))
)
optimizer.step()
param_stats = {}
for key, param in policy.named_parameters():
param_stats[f"{key}_mean"] = param.mean()
param_stats[f"{key}_std"] = param.std() if param.numel() > 1 else torch.tensor(float(0.0))
optimizer.zero_grad()
policy.reset()
# HACK: We reload a batch with no delta_indices as `select_action` won't expect a timestamps dimension
# We simulate having an environment using a dataset by setting delta_indices to None and dropping tensors
# indicating padding (those ending with "_is_pad")
dataset.delta_indices = None
batch = next(iter(dataloader))
obs = {}
for k in batch:
# TODO: regenerate the safetensors
# for backward compatibility
if k.endswith("_is_pad"):
continue
# for backward compatibility
if k == "task":
continue
if k.startswith("observation"):
obs[k] = batch[k]
if hasattr(train_cfg.policy, "n_action_steps"):
actions_queue = train_cfg.policy.n_action_steps
else:
actions_queue = train_cfg.policy.n_action_repeats
actions = {str(i): policy.select_action(obs).contiguous() for i in range(actions_queue)}
return output_dict, grad_stats, param_stats, actions
def save_policy_to_safetensors(output_dir: Path, ds_repo_id: str, policy_name: str, policy_kwargs: dict):
if output_dir.exists():
print(f"Overwrite existing safetensors in '{output_dir}':")
print(f" - Validate with: `git add {output_dir}`")
print(f" - Revert with: `git checkout -- {output_dir}`")
shutil.rmtree(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
output_dict, grad_stats, param_stats, actions = get_policy_stats(ds_repo_id, policy_name, policy_kwargs)
save_file(output_dict, output_dir / "output_dict.safetensors")
save_file(grad_stats, output_dir / "grad_stats.safetensors")
save_file(param_stats, output_dir / "param_stats.safetensors")
save_file(actions, output_dir / "actions.safetensors")
if __name__ == "__main__":
artifacts_cfg = [
("lerobot/xarm_lift_medium", "tdmpc", {"use_mpc": False}, "use_policy"),
("lerobot/xarm_lift_medium", "tdmpc", {"use_mpc": True}, "use_mpc"),
(
"lerobot/pusht",
"diffusion",
{
"n_action_steps": 8,
"num_inference_steps": 10,
"down_dims": [128, 256, 512],
},
"",
),
("lerobot/aloha_sim_insertion_human", "act", {"n_action_steps": 10}, ""),
(
"lerobot/aloha_sim_insertion_human",
"act",
{"n_action_steps": 1000, "chunk_size": 1000},
"1000_steps",
),
]
if len(artifacts_cfg) == 0:
raise RuntimeError("No policies were provided!")
for ds_repo_id, policy, policy_kwargs, file_name_extra in artifacts_cfg:
ds_name = ds_repo_id.split("/")[-1]
output_dir = Path("tests/data/save_policy_to_safetensors") / f"{ds_name}_{policy}_{file_name_extra}"
save_policy_to_safetensors(output_dir, ds_repo_id, policy, policy_kwargs)