fix(image transforms): cleaning up image_transforms implementation in LeRobotDataset

* fix(features copy): adding deepcopy on LeRobot dataset features to avoid shallow copy leaks

* tests(test): adding new test

src/lerobot/datasets/dataset_metadata.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 import contextlib
 from collections.abc import Callable
+from copy import deepcopy
 from pathlib import Path
 
 import numpy as np
@@ -709,7 +710,7 @@ class LeRobotDatasetMetadata:
 
         obj.root.mkdir(parents=True, exist_ok=False)
 
-        features = {**features, **DEFAULT_FEATURES}
+        features = {**deepcopy(features), **DEFAULT_FEATURES}
         _validate_feature_names(features)
 
         obj.tasks = None

src/lerobot/datasets/dataset_reader.py

@@ -74,6 +74,8 @@ class DatasetReader:
         self.episodes = episodes
         self._tolerance_s = tolerance_s
         self._video_backend = video_backend
+        if image_transforms is not None and not callable(image_transforms):
+            raise TypeError("image_transforms must be callable or None.")
         self._image_transforms = image_transforms
         self._return_uint8 = return_uint8
 
@@ -86,6 +88,16 @@ class DatasetReader:
             check_delta_timestamps(delta_timestamps, meta.fps, tolerance_s)
             self.delta_indices = get_delta_indices(delta_timestamps, meta.fps)
 
+    def set_image_transforms(self, image_transforms: Callable | None) -> None:
+        """Replace the transform applied to visual observations."""
+        if image_transforms is not None and not callable(image_transforms):
+            raise TypeError("image_transforms must be callable or None.")
+        self._image_transforms = image_transforms
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform applied to visual observations."""
+        self._image_transforms = None
+
     def try_load(self) -> bool:
         """Attempt to load from local cache. Returns True if data is sufficient."""
         try:

src/lerobot/datasets/dataset_tools.py

@@ -27,6 +27,7 @@ import logging
 import shutil
 from collections.abc import Callable
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
+from copy import deepcopy
 from pathlib import Path
 
 import datasets
@@ -1101,7 +1102,9 @@ def _copy_episodes_metadata_and_stats(
     if dst_meta.video_keys and src_dataset.meta.video_keys:
         for key in dst_meta.video_keys:
             if key in src_dataset.meta.features:
-                dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {})
+                dst_meta.info.features[key]["info"] = deepcopy(
+                    src_dataset.meta.info.features[key].get("info", {})
+                )
 
     write_info(dst_meta.info, dst_meta.root)
 

src/lerobot/datasets/lerobot_dataset.py

@@ -201,8 +201,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         super().__init__()
         self.repo_id = repo_id
         self._requested_root = Path(root) if root else None
-        self.reader = None
-        self.set_image_transforms(image_transforms)
         self.delta_timestamps = delta_timestamps
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
@@ -249,6 +247,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             image_transforms=image_transforms,
             return_uint8=self._return_uint8,
         )
+        self.image_transforms = image_transforms
 
         # Load actual data
         if force_cache_sync or not self.reader.try_load():
@@ -505,15 +504,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def set_image_transforms(self, image_transforms: Callable | None) -> None:
         """Replace the transform applied to visual observations."""
-        if image_transforms is not None and not callable(image_transforms):
-            raise TypeError("image_transforms must be callable or None.")
+        self._ensure_reader().set_image_transforms(image_transforms)
         self.image_transforms = image_transforms
-        if self.reader is not None:
-            self.reader._image_transforms = image_transforms
 
     def clear_image_transforms(self) -> None:
         """Remove the transform applied to visual observations."""
-        self.set_image_transforms(None)
+        if self.reader is not None:
+            self.reader.set_image_transforms(None)
+        self.image_transforms = None
 
     # ── Hub methods (stay on facade) ──────────────────────────────────
 

tests/datasets/test_datasets.py

@@ -51,7 +51,7 @@ from lerobot.robots import make_robot_from_config
 from lerobot.transforms import ImageTransforms, ImageTransformsConfig
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, OBS_STR, REWARD
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_REPO_ID
+from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_MOTOR_FEATURES, DUMMY_REPO_ID
 from tests.mocks.mock_robot import MockRobotConfig
 from tests.utils import require_x86_64_kernel
 
@@ -133,6 +133,21 @@ def test_dataset_feature_with_forward_slash_raises_error():
         )
 
 
+def test_create_does_not_mutate_input_features(tmp_path, empty_lerobot_dataset_factory):
+    # ``create`` must deep-copy features so a dataset built from another's features stays independent.
+    dataset = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds1", features=DUMMY_MOTOR_FEATURES, use_videos=False
+    )
+    dataset_copy = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds2", features=dataset.meta.features, use_videos=False
+    )
+
+    original_shape = dataset.meta.info.features["state"]["shape"]
+    dataset_copy.meta.info.features["state"]["shape"] = (999,)
+
+    assert dataset.meta.info.features["state"]["shape"] == original_shape
+
+
 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)