action-worldmodel-bench / compute_force_stats.py
syCen's picture
Create compute_force_stats.py
3fc1615 verified
Raw
History Blame Contribute Delete
4.33 kB
"""
compute_force_stats.py
Compute per-channel ACTIVE (nonzero) statistics for FORCE only.
Why: the original force_ch_std was computed over ALL pixels including the ~98%
zero background, which dilutes std to a tiny value. Dividing by that tiny std
inflates real contact forces and makes the VAE loss spike. This computes std
over NONZERO pixels only -> the correct normalization scale.
Contact is NOT recomputed (it uses contact_ch_max, which is unaffected by zeros).
Output: writes/updates force_ch_active_mean / force_ch_active_std into a json.
You can either merge these into your existing dataset_norm_params.json, or pass
this new json to the dataloader (it reads force_ch_active_std if present).
Usage:
python compute_force_stats.py --root ./grasping
python compute_force_stats.py --root ./grasping --out ./grasping/force_stats.json
# merge into existing stats file:
python compute_force_stats.py --root ./grasping --merge_into ./grasping/dataset_norm_params.json
"""
import argparse
import json
from pathlib import Path
import numpy as np
from tqdm import tqdm
CH_NAMES = ["L_fx", "L_fy", "L_fz", "R_fx", "R_fy", "R_fz"]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--root", required=True)
ap.add_argument("--out", default=None,
help="output json (default: <root>/force_stats.json)")
ap.add_argument("--merge_into", default=None,
help="if set, merge force_ch_active_* into this existing json")
args = ap.parse_args()
root = Path(args.root)
episodes = [p for p in root.rglob("*")
if p.is_dir() and (p / "masks.json").exists()]
print(f"Found {len(episodes)} valid episodes")
# per-channel accumulators over NONZERO pixels only
f_sum = np.zeros(6); f_sq = np.zeros(6); f_cnt = np.zeros(6)
f_absmax = np.zeros(6)
f_min = np.full(6, np.inf); f_max = np.full(6, -np.inf)
total_frames = 0
skipped = 0
for ep in tqdm(episodes, desc="episodes"):
fdir = ep / "modalities" / "force"
if not fdir.exists():
skipped += 1
continue
for ff in sorted(fdir.glob("*.npy")):
f = np.load(ff).astype(np.float64) # (6,H,W)
if f.ndim != 3 or f.shape[0] != 6:
continue
total_frames += 1
for ch in range(6):
v = f[ch][f[ch] != 0] # NONZERO only
if v.size:
f_sum[ch] += v.sum()
f_sq[ch] += (v ** 2).sum()
f_cnt[ch] += v.size
f_absmax[ch] = max(f_absmax[ch], np.abs(v).max())
f_min[ch] = min(f_min[ch], v.min())
f_max[ch] = max(f_max[ch], v.max())
if total_frames == 0:
raise RuntimeError("no force frames found")
f_cnt_safe = np.maximum(f_cnt, 1)
f_active_mean = f_sum / f_cnt_safe
f_active_std = np.sqrt(np.maximum(f_sq / f_cnt_safe - f_active_mean ** 2, 1e-12))
f_min[~np.isfinite(f_min)] = 0.0
f_max[~np.isfinite(f_max)] = 0.0
force_stats = {
"force_ch_active_mean": f_active_mean.tolist(),
"force_ch_active_std": f_active_std.tolist(),
"force_ch_abs_max": f_absmax.tolist(),
"force_ch_min": f_min.tolist(),
"force_ch_max": f_max.tolist(),
"num_frames": int(total_frames),
}
# print summary
print(f"\nframes used: {total_frames}")
print(f"{'ch':6s} {'active_std':>11s} {'abs_max':>9s} {'max/std':>8s}")
print("-" * 40)
for i in range(6):
ratio = f_absmax[i] / (f_active_std[i] + 1e-12)
print(f"{CH_NAMES[i]:6s} {f_active_std[i]:11.5f} {f_absmax[i]:9.4f} {ratio:8.1f}")
print("\n(if max/std > ~20, a clip after normalization is still recommended)")
# write output
if args.merge_into:
with open(args.merge_into) as fp:
existing = json.load(fp)
existing.update(force_stats)
with open(args.merge_into, "w") as fp:
json.dump(existing, fp, indent=2)
print(f"\nmerged force_ch_active_* into {args.merge_into}")
else:
out = args.out or str(root / "force_stats.json")
with open(out, "w") as fp:
json.dump(force_stats, fp, indent=2)
print(f"\nwrote {out}")
if __name__ == "__main__":
main()