Upload det04_find_batch_culprit.py with huggingface_hub

det04_find_batch_culprit.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+"""
+det04 — Find the first op in det_500m.onnx (via onnx2torch) where B=1 and
+        frame[0] of B=2 outputs diverge.
+
+Registers forward hooks on all submodules, runs B=1 and B=2 on the same image,
+then reports the first module whose output differs — the culprit causing cross-frame
+contamination.
+"""
+import argparse
+import cv2
+import torch
+import onnx2torch
+
+DET_SIZE = (640, 640)
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--model', required=True, help='Path to det_500m.onnx')
+parser.add_argument('--image', required=True, help='Path to any face image (jpg/png)')
+args = parser.parse_args()
+
+# ── Load image ────────────────────────────────────────────────────────────────
+import cv2, numpy as np
+img = cv2.imread(args.image)
+assert img is not None, f'Could not read image: {args.image}'
+blob = cv2.dnn.blobFromImage(
+    cv2.resize(img, DET_SIZE), 1.0 / 128.0, DET_SIZE,
+    (127.5, 127.5, 127.5), swapRB=True,
+)
+t = torch.from_numpy(blob)
+
+# ── Load model ────────────────────────────────────────────────────────────────
+print('Loading onnx2torch model ...')
+pt = onnx2torch.convert(args.model)
+pt.eval()
+
+# ── Hook: capture one output tensor per named module ─────────────────────────
+acts_b1 = {}
+acts_b2 = {}
+
+def make_hook(store, name):
+    def hook(module, input, output):
+        out = output[0] if isinstance(output, (tuple, list)) else output
+        store[name] = out.detach()
+    return hook
+
+handles = []
+for name, module in pt.named_modules():
+    if name == '':
+        continue
+    handles.append(module.register_forward_hook(make_hook(acts_b1, name)))
+
+# ── B=1 forward ───────────────────────────────────────────────────────────────
+with torch.no_grad():
+    pt(t)
+
+for h in handles: h.remove()
+handles.clear()
+
+for name, module in pt.named_modules():
+    if name == '':
+        continue
+    handles.append(module.register_forward_hook(make_hook(acts_b2, name)))
+
+# ── B=2 forward ───────────────────────────────────────────────────────────────
+t2 = torch.cat([t, t], dim=0)
+with torch.no_grad():
+    pt(t2)
+
+for h in handles: h.remove()
+
+# ── Compare: find first divergence ───────────────────────────────────────────
+print(f'\nComparing {len(acts_b1)} intermediate activations ...')
+print(f'{"Module":<55} {"B1 shape":<22} {"B2 shape":<22} {"max_diff":>10}  status')
+print('-' * 120)
+
+first_culprit = None
+for name in acts_b1:
+    if name not in acts_b2:
+        continue
+    a1 = acts_b1[name]
+    a2 = acts_b2[name]
+
+    try:
+        if a2.shape[0] == 2:
+            a2_f0  = a2[0]
+            a1_cmp = a1[0] if a1.shape[0] == 1 else a1
+        elif a1.numel() > 0 and a2.numel() == 2 * a1.numel():
+            a2_f0  = a2.flatten()[:a1.numel()].reshape(a1.shape)
+            a1_cmp = a1
+        else:
+            continue
+    except Exception:
+        continue
+
+    try:
+        diff = (a1_cmp.float() - a2_f0.float()).abs().max().item()
+    except Exception:
+        continue
+
+    ok = diff < 1e-4
+    if not ok and first_culprit is None:
+        first_culprit = name
+        status = '<<< FIRST MISMATCH'
+    elif not ok:
+        status = 'MISMATCH'
+    else:
+        status = 'OK'
+
+    if not ok:
+        print(f'{name:<55} {str(tuple(a1.shape)):<22} {str(tuple(a2.shape)):<22} {diff:>10.5f}  {status}')
+
+if first_culprit:
+    print(f'\n>>> Culprit: {first_culprit}')
+else:
+    print('\nNo divergence found — model is batch-independent.')
+
+print('\nDone.')