add scripts/add_xyzrgb.py

scripts/add_xyzrgb.py

@@ -0,0 +1,148 @@
+"""Add ``obs/pointcloud/xyzrgb (T, 1024, 6) float32`` to an MSHAB h5 produced by
+``add_pointcloud.py``. This is the final training-ready field that
+``data/hdf5_mshab_dataset.py`` reads.
+
+Matches the spec in ``third_party/mshab/DATA.md``:
+
+  * Foreground only (xyzw[:,:,3] > 0.5)
+  * Crop to a box around the robot base_pose:
+      x: [base_x - 1.5, base_x + 1.5]  m
+      y: [base_y - 1.5, base_y + 1.5]  m
+      z: [base_z + 0.0, base_z + 3.0]  m
+  * Exactly 1024 points per timestep (random sample w/ replacement if needed)
+  * Output format per-timestep: ``[x, y, z, r, g, b]`` where rgb is float in [0,1]
+
+Idempotent: if xyzrgb already exists with the right shape & dtype, skip.
+
+Usage:
+    python add_xyzrgb.py --data-dir path/to/gen_data_save_trajectories [--max-workers 16]
+"""
+
+from __future__ import annotations
+import argparse
+from pathlib import Path
+
+import h5py
+import numpy as np
+from tqdm import tqdm
+
+N_SAMPLE = 1024
+BOX_XY_HALF = 1.5
+BOX_Z_MIN = 0.0
+BOX_Z_MAX = 3.0
+
+
+def compute_xyzrgb(xyzw: np.ndarray, rgb_uint8: np.ndarray, base_pose: np.ndarray,
+                   rng: np.random.Generator) -> np.ndarray:
+    """xyzw: (N, 4) float32, rgb_uint8: (N, 3) uint8, base_pose: (>=3,) float.
+    Returns (1024, 6) float32 — xyz in meters, rgb in [0,1]."""
+    # 1. foreground mask
+    fg = xyzw[:, 3] > 0.5
+    pts = xyzw[fg, :3]
+    rgb = rgb_uint8[fg]
+
+    # 2. crop box around base
+    bx, by, bz = float(base_pose[0]), float(base_pose[1]), float(base_pose[2])
+    crop = (
+        (pts[:, 0] >= bx - BOX_XY_HALF) & (pts[:, 0] <= bx + BOX_XY_HALF) &
+        (pts[:, 1] >= by - BOX_XY_HALF) & (pts[:, 1] <= by + BOX_XY_HALF) &
+        (pts[:, 2] >= bz + BOX_Z_MIN)   & (pts[:, 2] <= bz + BOX_Z_MAX)
+    )
+    pts = pts[crop]
+    rgb = rgb[crop]
+
+    # 3. fixed-size sample
+    if pts.shape[0] == 0:
+        # degenerate: no points in box → return zeros
+        return np.zeros((N_SAMPLE, 6), dtype=np.float32)
+    if pts.shape[0] >= N_SAMPLE:
+        idx = rng.choice(pts.shape[0], size=N_SAMPLE, replace=False)
+    else:
+        idx = rng.choice(pts.shape[0], size=N_SAMPLE, replace=True)
+    pts = pts[idx]
+    rgb = rgb[idx].astype(np.float32) / 255.0
+    return np.concatenate([pts.astype(np.float32), rgb], axis=1)
+
+
+def process_trajectory(traj_group: h5py.Group) -> tuple[bool, str]:
+    if "obs" not in traj_group:
+        return False, "no_obs"
+    obs = traj_group["obs"]
+    if "pointcloud" not in obs:
+        return False, "no_pointcloud"
+    pc = obs["pointcloud"]
+    if "xyzw" not in pc or "rgb" not in pc:
+        return False, "missing xyzw/rgb"
+    if "extra" not in obs or "base_pose" not in obs["extra"]:
+        return False, "missing base_pose"
+
+    xyzw_all = pc["xyzw"]        # (T, N, 4)
+    rgb_all = pc["rgb"]          # (T, N, 3)
+    base_pose = obs["extra"]["base_pose"][:]   # (T, 3) or (T, 7)
+    T = xyzw_all.shape[0]
+
+    expected = (T, N_SAMPLE, 6)
+    if "xyzrgb" in pc:
+        ds = pc["xyzrgb"]
+        if tuple(ds.shape) == expected and np.dtype(ds.dtype) == np.float32:
+            return True, "already_done"
+        del pc["xyzrgb"]
+
+    rng = np.random.default_rng(2024)
+    out = np.empty(expected, dtype=np.float32)
+    for t in range(T):
+        out[t] = compute_xyzrgb(xyzw_all[t][:], rgb_all[t][:], base_pose[t], rng)
+
+    pc.create_dataset("xyzrgb", data=out, dtype=np.float32)
+    return True, "written"
+
+
+def process_h5_file(h5_path: Path):
+    with h5py.File(h5_path, "r+") as f:
+        traj_keys = [k for k in f.keys() if k.startswith("traj_")]
+        stats = dict(written=0, already_done=0, skipped=0)
+        for tk in tqdm(traj_keys, desc=h5_path.name):
+            try:
+                ok, reason = process_trajectory(f[tk])
+                if reason == "written": stats["written"] += 1
+                elif reason == "already_done": stats["already_done"] += 1
+                else: stats["skipped"] += 1
+            except Exception as e:
+                stats["skipped"] += 1
+                print(f"  [SKIP] {tk}: {type(e).__name__}: {e}")
+        print(f"[{h5_path.name}] {stats}")
+
+
+def iter_h5(data_dir: Path):
+    for p in sorted(data_dir.rglob("*.h5")):
+        if p.stat().st_size < 1024 * 1024:
+            continue
+        try:
+            with h5py.File(p, "r") as _f:
+                _ = list(_f.keys())[:1]
+        except Exception as e:
+            print(f"[SKIP corrupted] {p}: {e}")
+            continue
+        yield p
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--data-dir", required=True, type=Path)
+    args = ap.parse_args()
+
+    h5s = list(iter_h5(args.data_dir))
+    print(f"Found {len(h5s)} h5 file(s):")
+    for p in h5s:
+        print(f"  {p}  ({p.stat().st_size/1e9:.1f} GB)")
+
+    for p in h5s:
+        try:
+            process_h5_file(p)
+        except Exception as e:
+            import traceback
+            print(f"[ERROR] {p}: {e}"); traceback.print_exc()
+
+
+if __name__ == "__main__":
+    main()