Upload convert_hand_actions.py with huggingface_hub

convert_hand_actions.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+"""Convert binary hand actions to next-state continuous actions.
+
+For each timestep t, replaces:
+  action[t][7:13]  (left_hand)  <- observation.state[t+1][7:13]
+  action[t][29:35] (right_hand) <- observation.state[t+1][29:35]
+
+For the last timestep of each episode, repeats the current state.
+
+Creates new datasets with suffix '_nextstate' and symlinks videos to save space.
+"""
+
+import argparse
+import json
+import os
+import shutil
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+
+LEFT_HAND_SLICE = slice(7, 13)
+RIGHT_HAND_SLICE = slice(29, 35)
+
+
+def convert_episode_parquet(src_path: str, dst_path: str) -> int:
+    """Convert a single episode parquet file. Returns number of frames."""
+    df = pd.read_parquet(src_path)
+
+    actions = np.array(df["action"].tolist())
+    states = np.array(df["observation.state"].tolist())
+
+    n = len(df)
+
+    # Replace hand actions with next state (shift by 1)
+    # For t=0..n-2: action[t] = state[t+1]
+    actions[: n - 1, LEFT_HAND_SLICE] = states[1:n, LEFT_HAND_SLICE]
+    actions[: n - 1, RIGHT_HAND_SLICE] = states[1:n, RIGHT_HAND_SLICE]
+
+    # For last timestep: repeat current state
+    actions[n - 1, LEFT_HAND_SLICE] = states[n - 1, LEFT_HAND_SLICE]
+    actions[n - 1, RIGHT_HAND_SLICE] = states[n - 1, RIGHT_HAND_SLICE]
+
+    # Write back
+    df["action"] = actions.tolist()
+    os.makedirs(os.path.dirname(dst_path), exist_ok=True)
+    df.to_parquet(dst_path, index=False)
+    return n
+
+
+def convert_dataset(src_dir: str, dst_dir: str):
+    src = Path(src_dir)
+    dst = Path(dst_dir)
+
+    if dst.exists():
+        print(f"Destination {dst} already exists, skipping. Delete it first to reconvert.")
+        return
+
+    dst.mkdir(parents=True)
+
+    # 1. Symlink videos directory (saves huge disk space)
+    src_videos = src / "videos"
+    dst_videos = dst / "videos"
+    if src_videos.exists():
+        os.symlink(src_videos.resolve(), dst_videos)
+        print(f"  Symlinked videos: {dst_videos} -> {src_videos.resolve()}")
+
+    # 2. Copy meta directory (small)
+    src_meta = src / "meta"
+    dst_meta = dst / "meta"
+    shutil.copytree(src_meta, dst_meta)
+    print(f"  Copied meta/")
+
+    # 3. Remove stats.json so training auto-regenerates it
+    stats_path = dst_meta / "stats.json"
+    if stats_path.exists():
+        stats_path.unlink()
+        print(f"  Removed stats.json (will be auto-regenerated)")
+
+    relative_stats = dst_meta / "relative_stats.json"
+    if relative_stats.exists():
+        relative_stats.unlink()
+        print(f"  Removed relative_stats.json")
+
+    # 4. Convert parquet files
+    src_data = src / "data"
+    dst_data = dst / "data"
+    parquet_files = sorted(src_data.rglob("*.parquet"))
+    print(f"  Converting {len(parquet_files)} parquet files...")
+
+    total_frames = 0
+    for pf in tqdm(parquet_files, desc=f"  {src.name}"):
+        rel = pf.relative_to(src_data)
+        dst_pf = dst_data / rel
+        n = convert_episode_parquet(str(pf), str(dst_pf))
+        total_frames += n
+
+    print(f"  Done: {len(parquet_files)} episodes, {total_frames} frames converted")
+
+    # 5. Copy README with note about modification
+    readme_src = src / "README.md"
+    if readme_src.exists():
+        readme_text = readme_src.read_text()
+        note = (
+            "\n\n## Modification: Next-State Hand Actions\n\n"
+            "Hand action channels (left_hand indices 7-12, right_hand indices 29-34) "
+            "have been replaced with next-state values:\n"
+            "- `action[t][7:13] = observation.state[t+1][7:13]`\n"
+            "- `action[t][29:35] = observation.state[t+1][29:35]`\n"
+            "- Last timestep repeats current state.\n"
+            "- `stats.json` removed so training auto-regenerates normalization stats.\n"
+        )
+        (dst / "README.md").write_text(readme_text + note)
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert binary hand actions to next-state")
+    parser.add_argument(
+        "--src-dirs",
+        nargs="+",
+        default=[
+            "datasets/gr1_100x24",
+            "datasets/gr1_300x24",
+            "datasets/gr1_1000x24",
+        ],
+        help="Source dataset directories",
+    )
+    parser.add_argument(
+        "--suffix",
+        default="_nextstate",
+        help="Suffix for output directories (default: _nextstate)",
+    )
+    args = parser.parse_args()
+
+    for src_dir in args.src_dirs:
+        src = Path(src_dir)
+        dst = src.parent / (src.name + args.suffix)
+        print(f"\nConverting {src} -> {dst}")
+        convert_dataset(str(src), str(dst))
+
+    print("\nAll done!")
+
+
+if __name__ == "__main__":
+    main()