tangger/lerobot
1
0
Fork 1
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

LeRobotDataset v2.1 (#711)

Browse Source 
Co-authored-by: Remi <remi.cadene@huggingface.co>
Co-authored-by: Remi Cadene <re.cadene@gmail.com>
This commit is contained in:
Simon Alibert
2025-02-25 15:27:29 +01:00
committed by GitHub
parent aca464ca72
commit 3354d919fc
43 changed files with 2023 additions and 1322 deletions
Download Patch File Download Diff File Expand all files Collapse all files

359
tests/test_datasets.py
View File

@@ -15,24 +15,21 @@
# limitations under the License.
import json
import logging
import re
from copy import deepcopy
from itertools import chain
from pathlib import Path
import einops
import numpy as np
import pytest
import torch
from datasets import Dataset
from huggingface_hub import HfApi
from PIL import Image
from safetensors.torch import load_file
import lerobot
from lerobot.common.datasets.compute_stats import (
aggregate_stats,
compute_stats,
get_stats_einops_patterns,
)
from lerobot.common.datasets.factory import make_dataset
from lerobot.common.datasets.image_writer import image_array_to_pil_image
from lerobot.common.datasets.lerobot_dataset import (
LeRobotDataset,
MultiLeRobotDataset,
@@ -40,20 +37,34 @@ from lerobot.common.datasets.lerobot_dataset import (
from lerobot.common.datasets.utils import (
create_branch,
flatten_dict,
hf_transform_to_torch,
unflatten_dict,
)
from lerobot.common.envs.factory import make_env_config
from lerobot.common.policies.factory import make_policy_config
from lerobot.common.robot_devices.robots.utils import make_robot
from lerobot.common.utils.random_utils import seeded_context
from lerobot.configs.default import DatasetConfig
from lerobot.configs.train import TrainPipelineConfig
from tests.fixtures.constants import DUMMY_REPO_ID
from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_REPO_ID
from tests.utils import DEVICE, require_x86_64_kernel
def test_same_attributes_defined(lerobot_dataset_factory, tmp_path):
@pytest.fixture
def image_dataset(tmp_path, empty_lerobot_dataset_factory):
features = {
"image": {
"dtype": "image",
"shape": DUMMY_CHW,
"names": [
"channels",
"height",
"width",
],
}
}
return empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
def test_same_attributes_defined(tmp_path, lerobot_dataset_factory):
"""
Instantiate a LeRobotDataset both ways with '__init__()' and 'create()' and verify that instantiated
objects have the same sets of attributes defined.
@@ -66,24 +77,20 @@ def test_same_attributes_defined(lerobot_dataset_factory, tmp_path):
root_init = tmp_path / "init"
dataset_init = lerobot_dataset_factory(root=root_init)
# Access the '_hub_version' cached_property in both instances to force its creation
_ = dataset_init.meta._hub_version
_ = dataset_create.meta._hub_version
init_attr = set(vars(dataset_init).keys())
create_attr = set(vars(dataset_create).keys())
assert init_attr == create_attr
def test_dataset_initialization(lerobot_dataset_factory, tmp_path):
def test_dataset_initialization(tmp_path, lerobot_dataset_factory):
kwargs = {
"repo_id": DUMMY_REPO_ID,
"total_episodes": 10,
"total_frames": 400,
"episodes": [2, 5, 6],
}
dataset = lerobot_dataset_factory(root=tmp_path, **kwargs)
dataset = lerobot_dataset_factory(root=tmp_path / "test", **kwargs)
assert dataset.repo_id == kwargs["repo_id"]
assert dataset.meta.total_episodes == kwargs["total_episodes"]
@@ -93,12 +100,232 @@ def test_dataset_initialization(lerobot_dataset_factory, tmp_path):
assert dataset.num_frames == len(dataset)
def test_add_frame_missing_task(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError, match="Feature mismatch in `frame` dictionary:\nMissing features: {'task'}\n"
):
dataset.add_frame({"state": torch.randn(1)})
def test_add_frame_missing_feature(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError, match="Feature mismatch in `frame` dictionary:\nMissing features: {'state'}\n"
):
dataset.add_frame({"task": "Dummy task"})
def test_add_frame_extra_feature(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError, match="Feature mismatch in `frame` dictionary:\nExtra features: {'extra'}\n"
):
dataset.add_frame({"state": torch.randn(1), "task": "Dummy task", "extra": "dummy_extra"})
def test_add_frame_wrong_type(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError, match="The feature 'state' of dtype 'float16' is not of the expected dtype 'float32'.\n"
):
dataset.add_frame({"state": torch.randn(1, dtype=torch.float16), "task": "Dummy task"})
def test_add_frame_wrong_shape(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError,
match=re.escape("The feature 'state' of shape '(1,)' does not have the expected shape '(2,)'.\n"),
):
dataset.add_frame({"state": torch.randn(1), "task": "Dummy task"})
def test_add_frame_wrong_shape_python_float(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError,
match=re.escape(
"The feature 'state' is not a 'np.ndarray'. Expected type is 'float32', but type '<class 'float'>' provided instead.\n"
),
):
dataset.add_frame({"state": 1.0, "task": "Dummy task"})
def test_add_frame_wrong_shape_torch_ndim_0(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError,
match=re.escape("The feature 'state' of shape '()' does not have the expected shape '(1,)'.\n"),
):
dataset.add_frame({"state": torch.tensor(1.0), "task": "Dummy task"})
def test_add_frame_wrong_shape_numpy_ndim_0(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
with pytest.raises(
ValueError,
match=re.escape(
"The feature 'state' is not a 'np.ndarray'. Expected type is 'float32', but type '<class 'numpy.float32'>' provided instead.\n"
),
):
dataset.add_frame({"state": np.float32(1.0), "task": "Dummy task"})
def test_add_frame(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(1), "task": "Dummy task"})
dataset.save_episode()
assert len(dataset) == 1
assert dataset[0]["task"] == "Dummy task"
assert dataset[0]["task_index"] == 0
assert dataset[0]["state"].ndim == 0
def test_add_frame_state_1d(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(2), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].shape == torch.Size([2])
def test_add_frame_state_2d(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2, 4), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(2, 4), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].shape == torch.Size([2, 4])
def test_add_frame_state_3d(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2, 4, 3), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(2, 4, 3), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].shape == torch.Size([2, 4, 3])
def test_add_frame_state_4d(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2, 4, 3, 5), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(2, 4, 3, 5), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].shape == torch.Size([2, 4, 3, 5])
def test_add_frame_state_5d(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (2, 4, 3, 5, 1), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": torch.randn(2, 4, 3, 5, 1), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].shape == torch.Size([2, 4, 3, 5, 1])
def test_add_frame_state_numpy(tmp_path, empty_lerobot_dataset_factory):
features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": np.array([1], dtype=np.float32), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["state"].ndim == 0
def test_add_frame_string(tmp_path, empty_lerobot_dataset_factory):
features = {"caption": {"dtype": "string", "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"caption": "Dummy caption", "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["caption"] == "Dummy caption"
def test_add_frame_image_wrong_shape(image_dataset):
dataset = image_dataset
with pytest.raises(
ValueError,
match=re.escape(
"The feature 'image' of shape '(3, 128, 96)' does not have the expected shape '(3, 96, 128)' or '(96, 128, 3)'.\n"
),
):
c, h, w = DUMMY_CHW
dataset.add_frame({"image": torch.randn(c, w, h), "task": "Dummy task"})
def test_add_frame_image_wrong_range(image_dataset):
"""This test will display the following error message from a thread:
```
Error writing image ...test_add_frame_image_wrong_ran0/test/images/image/episode_000000/frame_000000.png:
The image data type is float, which requires values in the range [0.0, 1.0]. However, the provided range is [0.009678772038470007, 254.9776492089887].
Please adjust the range or provide a uint8 image with values in the range [0, 255]
```
Hence the image won't be saved on disk and save_episode will raise `FileNotFoundError`.
"""
dataset = image_dataset
dataset.add_frame({"image": np.random.rand(*DUMMY_CHW) * 255, "task": "Dummy task"})
with pytest.raises(FileNotFoundError):
dataset.save_episode()
def test_add_frame_image(image_dataset):
dataset = image_dataset
dataset.add_frame({"image": np.random.rand(*DUMMY_CHW), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
def test_add_frame_image_h_w_c(image_dataset):
dataset = image_dataset
dataset.add_frame({"image": np.random.rand(*DUMMY_HWC), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
def test_add_frame_image_uint8(image_dataset):
dataset = image_dataset
image = np.random.randint(0, 256, DUMMY_HWC, dtype=np.uint8)
dataset.add_frame({"image": image, "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
def test_add_frame_image_pil(image_dataset):
dataset = image_dataset
image = np.random.randint(0, 256, DUMMY_HWC, dtype=np.uint8)
dataset.add_frame({"image": Image.fromarray(image), "task": "Dummy task"})
dataset.save_episode()
assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
def test_image_array_to_pil_image_wrong_range_float_0_255():
image = np.random.rand(*DUMMY_HWC) * 255
with pytest.raises(ValueError):
image_array_to_pil_image(image)
# TODO(aliberts):
# - [ ] test various attributes & state from init and create
# - [ ] test init with episodes and check num_frames
# - [ ] test add_frame
# - [ ] test add_episode
# - [ ] test consolidate
# - [ ] test push_to_hub
# - [ ] test smaller methods
@@ -210,67 +437,6 @@ def test_multidataset_frames():
assert torch.equal(sub_dataset_item[k], dataset_item[k])
# TODO(aliberts, rcadene): Refactor and move this to a tests/test_compute_stats.py
def test_compute_stats_on_xarm():
"""Check that the statistics are computed correctly according to the stats_patterns property.
We compare with taking a straight min, mean, max, std of all the data in one pass (which we can do
because we are working with a small dataset).
"""
# TODO(rcadene, aliberts): remove dataset download
dataset = LeRobotDataset("lerobot/xarm_lift_medium", episodes=[0])
# reduce size of dataset sample on which stats compute is tested to 10 frames
dataset.hf_dataset = dataset.hf_dataset.select(range(10))
# Note: we set the batch size to be smaller than the whole dataset to make sure we are testing batched
# computation of the statistics. While doing this, we also make sure it works when we don't divide the
# dataset into even batches.
computed_stats = compute_stats(dataset, batch_size=int(len(dataset) * 0.25), num_workers=0)
# get einops patterns to aggregate batches and compute statistics
stats_patterns = get_stats_einops_patterns(dataset)
# get all frames from the dataset in the same dtype and range as during compute_stats
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=0,
batch_size=len(dataset),
shuffle=False,
)
full_batch = next(iter(dataloader))
# compute stats based on all frames from the dataset without any batching
expected_stats = {}
for k, pattern in stats_patterns.items():
full_batch[k] = full_batch[k].float()
expected_stats[k] = {}
expected_stats[k]["mean"] = einops.reduce(full_batch[k], pattern, "mean")
expected_stats[k]["std"] = torch.sqrt(
einops.reduce((full_batch[k] - expected_stats[k]["mean"]) ** 2, pattern, "mean")
)
expected_stats[k]["min"] = einops.reduce(full_batch[k], pattern, "min")
expected_stats[k]["max"] = einops.reduce(full_batch[k], pattern, "max")
# test computed stats match expected stats
for k in stats_patterns:
assert torch.allclose(computed_stats[k]["mean"], expected_stats[k]["mean"])
assert torch.allclose(computed_stats[k]["std"], expected_stats[k]["std"])
assert torch.allclose(computed_stats[k]["min"], expected_stats[k]["min"])
assert torch.allclose(computed_stats[k]["max"], expected_stats[k]["max"])
# load stats used during training which are expected to match the ones returned by computed_stats
loaded_stats = dataset.meta.stats # noqa: F841
# TODO(rcadene): we can't test this because expected_stats is computed on a subset
# # test loaded stats match expected stats
# for k in stats_patterns:
# assert torch.allclose(loaded_stats[k]["mean"], expected_stats[k]["mean"])
# assert torch.allclose(loaded_stats[k]["std"], expected_stats[k]["std"])
# assert torch.allclose(loaded_stats[k]["min"], expected_stats[k]["min"])
# assert torch.allclose(loaded_stats[k]["max"], expected_stats[k]["max"])
# TODO(aliberts): Move to more appropriate location
def test_flatten_unflatten_dict():
d = {
@@ -374,35 +540,6 @@ def test_backward_compatibility(repo_id):
# load_and_compare(i - 1)
@pytest.mark.skip("TODO after fix multidataset")
def test_multidataset_aggregate_stats():
"""Makes 3 basic datasets and checks that aggregate stats are computed correctly."""
with seeded_context(0):
data_a = torch.rand(30, dtype=torch.float32)
data_b = torch.rand(20, dtype=torch.float32)
data_c = torch.rand(20, dtype=torch.float32)
hf_dataset_1 = Dataset.from_dict(
{"a": data_a[:10], "b": data_b[:10], "c": data_c[:10], "index": torch.arange(10)}
)
hf_dataset_1.set_transform(hf_transform_to_torch)
hf_dataset_2 = Dataset.from_dict({"a": data_a[10:20], "b": data_b[10:], "index": torch.arange(10)})
hf_dataset_2.set_transform(hf_transform_to_torch)
hf_dataset_3 = Dataset.from_dict({"a": data_a[20:], "c": data_c[10:], "index": torch.arange(10)})
hf_dataset_3.set_transform(hf_transform_to_torch)
dataset_1 = LeRobotDataset.from_preloaded("d1", hf_dataset=hf_dataset_1)
dataset_1.stats = compute_stats(dataset_1, batch_size=len(hf_dataset_1), num_workers=0)
dataset_2 = LeRobotDataset.from_preloaded("d2", hf_dataset=hf_dataset_2)
dataset_2.stats = compute_stats(dataset_2, batch_size=len(hf_dataset_2), num_workers=0)
dataset_3 = LeRobotDataset.from_preloaded("d3", hf_dataset=hf_dataset_3)
dataset_3.stats = compute_stats(dataset_3, batch_size=len(hf_dataset_3), num_workers=0)
stats = aggregate_stats([dataset_1, dataset_2, dataset_3])
for data_key, data in zip(["a", "b", "c"], [data_a, data_b, data_c], strict=True):
for agg_fn in ["mean", "min", "max"]:
assert torch.allclose(stats[data_key][agg_fn], einops.reduce(data, "n -> 1", agg_fn))
assert torch.allclose(stats[data_key]["std"], torch.std(data, correction=0))
@pytest.mark.skip("Requires internet access")
def test_create_branch():
api = HfApi()
@@ -431,9 +568,9 @@ def test_create_branch():
def test_dataset_feature_with_forward_slash_raises_error():
# make sure dir does not exist
from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME
from lerobot.common.constants import HF_LEROBOT_HOME
dataset_dir = LEROBOT_HOME / "lerobot/test/with/slash"
dataset_dir = HF_LEROBOT_HOME / "lerobot/test/with/slash"
# make sure does not exist
if dataset_dir.exists():
dataset_dir.rmdir()