Create diagnose_hard_clips.py

diagnose_hard_clips.py

@@ -0,0 +1,96 @@
+"""
+diagnose_hard_clips.py
+
+Run the force VAE over all clips, rank them by reconstruction loss, and report
+which clips are the 'stubborn' high-loss ones + their force magnitude stats.
+This tells you whether a few extreme clips are dragging the average, vs broad
+failure.
+
+Run:
+  python examples/wanvideo/model_training/diagnose_hard_clips.py \
+      --clips ... --stats ... --source_root ... \
+      --ckpt ./vae_ckpt/force_overfit/force_vae_ep39.pt \
+      --modality force --force_clip 5.0
+"""
+import argparse, os, sys
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import numpy as np
+import torch
+from tqdm import tqdm
+from diffsynth.models.physical_vae import ContactVAE, ForceVAE
+from physical_dataset import PhysicalClipDataset
+
+
+def main():
+    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("--ckpt", required=True)
+    ap.add_argument("--modality", choices=["contact","force"], default="force")
+    ap.add_argument("--force_clip", type=float, default=5.0)
+    ap.add_argument("--contact_gain", type=float, default=5.0)
+    ap.add_argument("--w_active", type=float, default=50.0)
+    ap.add_argument("--dtype", choices=["fp32","bf16"], default="bf16")
+    args = ap.parse_args()
+
+    device = "cuda"
+    dtype = torch.bfloat16 if args.dtype=="bf16" else torch.float32
+    ds = PhysicalClipDataset(args.clips, args.stats, args.source_root, args.modality,
+                             force_clip=args.force_clip, contact_gain=args.contact_gain)
+    Cls = ContactVAE if args.modality=="contact" else ForceVAE
+    vae = Cls(); vae._adapt_channels()
+    vae = vae.to(device=device, dtype=dtype)
+    sd = torch.load(args.ckpt, map_location="cpu")
+    vae.model.load_state_dict(sd["model"]); vae.model.eval()
+
+    rows = []
+    for i in tqdm(range(len(ds)), desc="clips"):
+        item = ds[i]
+        x = item["data"].unsqueeze(0).to(device, dtype=dtype)
+        active = item["active_mask"].unsqueeze(0).to(device)
+        with torch.no_grad():
+            x_rec, _, _ = vae(x, sample=False)
+        xn = x[0].float().cpu().numpy()
+        rn = x_rec[0].float().cpu().numpy()
+        # weighted L1 (same as training, w_active)
+        a = active[0,0].cpu().numpy()       # (T,H,W)
+        w = 1.0 + (args.w_active - 1.0) * a
+        err = np.abs(rn - xn).sum(0)        # sum over channels -> (T,H,W)
+        wl1 = (w * err).sum() / w.sum()
+        # active-region L1
+        act = a > 0
+        act_l1 = err[act].mean() if act.any() else 0.0
+        # input magnitude stats
+        in_absmax = np.abs(xn).max()
+        in_active_frac = float(a.mean())
+        rows.append((i, wl1, act_l1, in_absmax, in_active_frac))
+
+    rows.sort(key=lambda r: -r[1])   # sort by weighted L1 desc
+    print("\n" + "="*72)
+    print(f"{'clip':>4s} {'w_L1':>8s} {'act_L1':>8s} {'in_absmax':>10s} {'act_frac':>9s}")
+    print("-"*72)
+    print("HARDEST 10:")
+    for r in rows[:10]:
+        print(f"{r[0]:4d} {r[1]:8.4f} {r[2]:8.4f} {r[3]:10.4f} {r[4]:9.4f}")
+    print("EASIEST 5:")
+    for r in rows[-5:]:
+        print(f"{r[0]:4d} {r[1]:8.4f} {r[2]:8.4f} {r[3]:10.4f} {r[4]:9.4f}")
+
+    wl1s = np.array([r[1] for r in rows])
+    print("="*72)
+    print(f"weighted L1: mean={wl1s.mean():.4f} median={np.median(wl1s):.4f} "
+          f"max={wl1s.max():.4f}")
+    print(f"clips with w_L1 > 1.0 (stubborn): {(wl1s>1.0).sum()}/{len(wl1s)}")
+    print(f"clips with w_L1 < 0.1 (learned):  {(wl1s<0.1).sum()}/{len(wl1s)}")
+    # correlation between hardness and input magnitude
+    absmaxes = np.array([r[3] for r in rows])
+    if len(rows) > 5:
+        hard = wl1s > np.median(wl1s)
+        print(f"\nhard clips avg in_absmax:  {absmaxes[hard].mean():.3f}")
+        print(f"easy clips avg in_absmax:  {absmaxes[~hard].mean():.3f}")
+        print("(if hard >> easy, the stubborn clips are the extreme-force ones)")
+
+
+if __name__ == "__main__":
+    main()