chore(installation): remove libero installation patch (#2416)

* chore(installation): remove libero installation patch

* fix(ci): exclude groot for unbound deps test

.github/workflows/release.yml

@@ -83,11 +83,11 @@ jobs:
           fi
 
       - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove libero and pi from PyPi 0.4.0 release due to its reliance on git dependencies.
+        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
         run: |
           echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]|lerobot\[libero\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]|lerobot\[libero\]/d' pyproject.toml
+          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
+          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
           echo "::info:: Git dependencies removed. Proceeding with build."
 
       - name: Install build dependencies

.github/workflows/unbound_deps_tests.yml

@@ -70,7 +70,7 @@ jobs:
           echo "Dependencies unbound:" && cat pyproject.toml
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv

README.md

@@ -186,7 +186,7 @@ For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
 > [!NOTE]
-> For lerobot 0.4.0, if you want to install libero or pi tags, you will have to do: `pip install "lerobot[pi,libero]@git+https://github.com/huggingface/lerobot.git"`.
+> For lerobot 0.4.0, if you want to install pi tags, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
 >
 > This will be solved in the next patch release
 

docs/source/installation.mdx

@@ -82,7 +82,7 @@ For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
 > [!NOTE]
-> For lerobot 0.4.0, if you want to install libero or pi, you will have to do: `pip install "lerobot[pi,libero]@git+https://github.com/huggingface/lerobot.git"`
+> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
 
 ### Troubleshooting
 

docs/source/libero.mdx

@@ -28,11 +28,6 @@ LIBERO is now part of our **multi-eval supported simulation**, meaning you can b
 To Install LIBERO, after following LeRobot official instructions, just do:
 `pip install -e ".[libero]"`
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install libero tag, you will have to do: `pip install "lerobot[libero]@git+https://github.com/huggingface/lerobot.git"`.
->
-> This will be solved in the next patch release
-
 ### Single-suite evaluation
 
 Evaluate a policy on one LIBERO suite:

src/lerobot/datasets/lerobot_dataset.py

@@ -940,11 +940,26 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return query_timestamps
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
-        return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
-            for key, q_idx in query_indices.items()
-            if key not in self.meta.video_keys
-        }
+        """
+        Query dataset for indices across keys, skipping video keys.
+
+        Tries column-first [key][indices] for speed, falls back to row-first.
+
+        Args:
+            query_indices: Dict mapping keys to index lists to retrieve
+
+        Returns:
+            Dict with stacked tensors of queried data (video keys excluded)
+        """
+        result: dict = {}
+        for key, q_idx in query_indices.items():
+            if key in self.meta.video_keys:
+                continue
+            try:
+                result[key] = torch.stack(self.hf_dataset[key][q_idx])
+            except (KeyError, TypeError, IndexError):
+                result[key] = torch.stack(self.hf_dataset[q_idx][key])
+        return result
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
         """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -50,9 +50,9 @@ from typing import Any
 
 import jsonlines
 import pandas as pd
-import pyarrow.parquet as pq
+import pyarrow as pa
 import tqdm
-from datasets import Dataset, concatenate_datasets
+from datasets import Dataset, Features, Image
 from huggingface_hub import HfApi, snapshot_download
 from requests import HTTPError
 
@@ -68,7 +68,6 @@ from lerobot.datasets.utils import (
     LEGACY_EPISODES_STATS_PATH,
     LEGACY_TASKS_PATH,
     cast_stats_to_numpy,
-    embed_images,
     flatten_dict,
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
@@ -175,33 +174,25 @@ def convert_tasks(root, new_root):
     write_tasks(df_tasks, new_root)
 
 
-def concat_data_files(
-    paths_to_cat: list[Path], new_root: Path, chunk_idx: int, file_idx: int, image_keys: list[str]
-):
-    """Concatenate multiple parquet data files into a single file.
-
-    Args:
-        paths_to_cat: List of parquet file paths to concatenate
-        new_root: Root directory for the new dataset
-        chunk_idx: Chunk index for the output file
-        file_idx: File index within the chunk
-        image_keys: List of feature keys that contain images
-    """
-
-    datasets_list: list[Dataset] = [Dataset.from_parquet(str(file)) for file in paths_to_cat]
-    concatenated_ds: Dataset = concatenate_datasets(datasets_list)
-
-    if len(image_keys) > 0:
-        logging.debug(f"Embedding {len(image_keys)} image features for optimal training performance")
-        concatenated_ds = embed_images(concatenated_ds)
+def concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys):
+    # TODO(rcadene): to save RAM use Dataset.from_parquet(file) and concatenate_datasets
+    dataframes = [pd.read_parquet(file) for file in paths_to_cat]
+    # Concatenate all DataFrames along rows
+    concatenated_df = pd.concat(dataframes, ignore_index=True)
 
     path = new_root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
     path.parent.mkdir(parents=True, exist_ok=True)
 
-    table = concatenated_ds.with_format("arrow")[:]
-    writer = pq.ParquetWriter(path, schema=table.schema, compression="snappy", use_dictionary=True)
-    writer.write_table(table)
-    writer.close()
+    if len(image_keys) > 0:
+        schema = pa.Schema.from_pandas(concatenated_df)
+        features = Features.from_arrow_schema(schema)
+        for key in image_keys:
+            features[key] = Image()
+        schema = features.arrow_schema
+    else:
+        schema = None
+
+    concatenated_df.to_parquet(path, index=False, schema=schema)
 
 
 def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):