check_active.py

"""
Quick check: how many clips actually contain contact (nonzero) frames?
If most clips are all-zero, the VAE will collapse to predicting 0.

Usage:
  python check_clip_activity.py --clips ... --stats ... --source_root ... --modality contact --n 200
"""
import argparse, sys, os
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from physical_dataset import PhysicalClipDataset
import numpy as np

ap = argparse.ArgumentParser()
ap.add_argument("--clips", required=True)
ap.add_argument("--stats", required=True)
ap.add_argument("--source_root", required=True)
ap.add_argument("--modality", default="contact")
ap.add_argument("--n", type=int, default=200, help="how many clips to scan")
args = ap.parse_args()

ds = PhysicalClipDataset(args.clips, args.stats, args.source_root, args.modality)
n = min(args.n, len(ds))
print(f"scanning {n}/{len(ds)} clips...")

all_zero = 0
active_fracs = []
contact_frame_counts = []   # how many of the 17 frames have any contact
for i in range(n):
    item = ds[i]
    m = item["active_mask"]            # (1,T,H,W)
    af = float(m.mean())
    active_fracs.append(af)
    if af == 0.0:
        all_zero += 1
    # frames with any contact
    per_frame = m[0].reshape(m.shape[1], -1).sum(1)   # (T,)
    contact_frame_counts.append(int((per_frame > 0).sum()))

active_fracs = np.array(active_fracs)
cfc = np.array(contact_frame_counts)
print(f"\nall-zero clips: {all_zero}/{n} ({100*all_zero/n:.1f}%)")
print(f"active frac: mean={active_fracs.mean():.4f} max={active_fracs.max():.4f}")
print(f"contact frames per clip (of 17): mean={cfc.mean():.1f} "
      f"min={cfc.min()} max={cfc.max()}")
print(f"clips with >=1 contact frame: {(cfc>0).sum()}/{n} ({100*(cfc>0).sum()/n:.1f}%)")
print(f"clips with >=8 contact frames: {(cfc>=8).sum()}/{n}")