""" check_all_metrics_force.py Force-specific reconstruction metrics over ALL clips (analogous to check_all_metrics.py for contact, but adapted for 6 SIGNED channels). Force differs from contact: - 6 channels (L_fx,L_fy,L_fz, R_fx,R_fy,R_fz), each can be POSITIVE or NEGATIVE - so "peak" uses ABS-max (a strong force may be negative) - we additionally check DIRECTION: sign agreement + cosine similarity, because reconstructing +0.5 vs -0.5 is a fundamental error that magnitude alone hides Per channel it reports: - absmax_ratio : |recon|max / |input|max (magnitude recovery, ~1.0 = good) - active_L1 : L1 error on active pixels - peak_err : pixel distance between input & recon abs-peak location - sign_agree : fraction of active pixels where sign(recon)==sign(input) - cosine : cosine similarity over active pixels (direction+shape match) Run: python examples/wanvideo/model_training/check_all_metrics_force.py \ --clips ... --stats ... --source_root ... \ --ckpt ./vae_ckpt/force_overfit_clip10/force_vae_ep199.pt --force_clip 10.0 """ 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 ForceVAE from physical_dataset import PhysicalClipDataset CH_NAMES = ["L_fx", "L_fy", "L_fz", "R_fx", "R_fy", "R_fz"] def abspeak_xy(img): """(y,x) of the max ABSOLUTE value pixel (force can be negative).""" return np.unravel_index(int(np.argmax(np.abs(img))), img.shape) 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("--force_clip", type=float, default=10.0) ap.add_argument("--dtype", choices=["fp32", "bf16"], default="bf16") ap.add_argument("--save_csv", default=None) 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, "force", force_clip=args.force_clip) vae = ForceVAE(); 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() # per-channel accumulators absmax_ratio = {c: [] for c in range(6)} act_l1 = {c: [] for c in range(6)} peak_err = {c: [] for c in range(6)} sign_agree = {c: [] for c in range(6)} cosine = {c: [] for c in range(6)} bg_l1_all = [] rows = [] for i in tqdm(range(len(ds)), desc="clips", dynamic_ncols=True): item = ds[i] if float(item["active_mask"].mean()) <= 0: continue x = item["data"].unsqueeze(0).to(device, dtype=dtype) with torch.no_grad(): x_rec, _, _ = vae(x, sample=False) xn = x[0].float().cpu().numpy() # (6,17,H,W) rn = x_rec[0].float().cpu().numpy() # background L1: pixels where ALL channels are zero active_any = (np.abs(xn).sum(0) > 0) # (17,H,W) bg = ~active_any err_map = np.abs(rn - xn).sum(0) if bg.any(): bg_l1_all.append(err_map[bg].mean()) row = {"clip": i} for ch in range(6): # frame with the strongest (abs) input for this channel energy = np.abs(xn[ch]).reshape(xn.shape[1], -1).sum(1) if energy.max() < 1e-4: continue fr = int(np.argmax(energy)) in_img = xn[ch, fr]; rec_img = rn[ch, fr] # magnitude recovery via abs-max (handles negative forces) in_am = np.abs(in_img).max() rec_am = np.abs(rec_img).max() absmax_ratio[ch].append(rec_am / (in_am + 1e-9)) # active region (this channel nonzero) am = np.abs(in_img) > 0 if am.any(): act_l1[ch].append(np.abs(rec_img - in_img)[am].mean()) # sign agreement on active pixels (ignore near-zero recon) si = np.sign(in_img[am]); sr = np.sign(rec_img[am]) valid = np.abs(rec_img[am]) > 0.05 * (rec_am + 1e-9) if valid.any(): sign_agree[ch].append((si[valid] == sr[valid]).mean()) # cosine over active pixels (direction + shape) a = in_img[am]; b = rec_img[am] cosine[ch].append((a @ b) / (np.linalg.norm(a)*np.linalg.norm(b)+1e-9)) # abs-peak location error iy, ix = abspeak_xy(in_img); ry, rx = abspeak_xy(rec_img) peak_err[ch].append(float(np.hypot(iy-ry, ix-rx))) row[f"{CH_NAMES[ch]}_absmax_ratio"] = rec_am/(in_am+1e-9) rows.append(row) def summ(name, d): if not d: return f"{name}: (none)" a = np.array(d) return (f"{name}: mean={a.mean():.4f} median={np.median(a):.4f} " f"min={a.min():.4f} max={a.max():.4f}") print("\n" + "="*70) for ch in range(6): n = len(absmax_ratio[ch]) if n == 0: print(f"--- {CH_NAMES[ch]}: no contact ---"); continue print(f"--- {CH_NAMES[ch]} ({n} clips) ---") print(" " + summ("absmax_ratio", absmax_ratio[ch])) print(" " + summ("active_L1 ", act_l1[ch])) print(" " + summ("peak_err(px)", peak_err[ch])) print(" " + summ("sign_agree ", sign_agree[ch])) print(" " + summ("cosine ", cosine[ch])) print("-"*70) if bg_l1_all: print(" " + summ("background_L1", bg_l1_all)) # overall verdict across all channels all_ratio = np.array([v for c in range(6) for v in absmax_ratio[c]]) all_cos = np.array([v for c in range(6) for v in cosine[c]]) all_sign = np.array([v for c in range(6) for v in sign_agree[c]]) print("="*70) print(f"OVERALL absmax_ratio: mean={all_ratio.mean():.3f} " f"| in[0.7,1.3]: {100*((all_ratio>0.7)&(all_ratio<1.3)).mean():.0f}%") print(f"OVERALL cosine: mean={all_cos.mean():.3f} " f"| >0.8: {100*(all_cos>0.8).mean():.0f}%") print(f"OVERALL sign_agree: mean={all_sign.mean():.3f}") if all_cos.mean() > 0.8 and all_sign.mean() > 0.9: print("VERDICT: GOOD - force magnitude AND direction reconstructed.") elif all_cos.mean() > 0.6: print("VERDICT: PARTIAL - shape ok, check weak/negative channels.") else: print("VERDICT: POOR - direction or magnitude failing.") if args.save_csv and rows: import csv keys = sorted({k for r in rows for k in r}) with open(args.save_csv, "w", newline="") as f: w = csv.DictWriter(f, fieldnames=keys); w.writeheader(); w.writerows(rows) print(f"\nper-clip CSV -> {args.save_csv}") if __name__ == "__main__": main()