feat: add Old vs New comparison tab

Compares the deprecated timm ViTClassifier against the fixed HF
ViTForImageClassification pipeline. Same weights, different loading
paths — proves the fix is transparent and correct.

test_app/app.py

@@ -44,7 +44,34 @@ print(f"[PyTorch]  heads={pt_model.config.num_attention_heads}  "
       f"classes={pt_model.config.num_classes}  "
       f"hidden={pt_model.config.hidden_size}")
 
-# ── ONNX ──────────────────────────────────────────────────────────────
+# ── Old timm-based model (deprecated path) ────────────────────────────
+
+import sys
+SCRIPTS_DIR = os.path.join(BASE_DIR, "scripts")
+if SCRIPTS_DIR not in sys.path:
+    sys.path.insert(0, SCRIPTS_DIR)
+from modeling_vit_classifier import ViTClassifier as OldViTClassifier
+
+def _load_old_model():
+    old_cfg = json.load(open(os.path.join(os.path.dirname(__file__), "old_config.json")))
+    device = old_cfg["device"]
+    if not torch.cuda.is_available():
+        device = "cpu"
+    model = OldViTClassifier(old_cfg, device=device)
+    ckpt = torch.load(old_cfg["checkpoint_path"], map_location=device, weights_only=False)
+    model.load_state_dict(ckpt["model"])
+    return model.to(device).eval()
+
+import json
+_old_model = None
+_old_device = None
+
+def _get_old_model():
+    global _old_model, _old_device
+    if _old_model is None:
+        _old_device = "cuda" if torch.cuda.is_available() else "cpu"
+        _old_model = _load_old_model()
+    return _old_model, _old_device
 
 _onnx_sessions = {}
 def _get_onnx(variant):
@@ -192,6 +219,50 @@ def benchmark():
     return (gr.update(value=markdown), guide)
 
 
+# ── Tab 3: Old vs New ────────────────────────────────────────────────
+
+def compare_old_new(image):
+    if image is None:
+        return (None, None)
+
+    # ── New (HF ViTForImageClassification, fixed config) ──
+    t0 = time.perf_counter()
+    inputs = pt_processor(image, return_tensors="pt")
+    inputs = {k: v.to(pt_device) for k, v in inputs.items()}
+    with torch.no_grad():
+        logits = pt_model(**inputs).logits[0].cpu()
+        new_probs = torch.softmax(logits, dim=-1)
+        new_pred = pt_model.config.id2label[torch.argmax(new_probs).item()]
+    new_ms = (time.perf_counter() - t0) * 1000
+
+    new_result = {
+        "backend": "ViTForImageClassification (fixed config, July 2026)",
+        "prediction": new_pred,
+        "real": round(new_probs[0].item(), 4),
+        "fake": round(new_probs[1].item(), 4),
+        "time_ms": round(new_ms, 1),
+    }
+
+    # ── Old (timm ViTClassifier, deprecated) ──
+    old_model, old_dev = _get_old_model()
+    t0 = time.perf_counter()
+    with torch.no_grad():
+        fake_prob = old_model.forward(image).item()
+    old_ms = (time.perf_counter() - t0) * 1000
+    old_pred = "fake" if fake_prob > 0.5 else "real"
+
+    old_result = {
+        "backend": "ViTClassifier (timm wrapper, deprecated)",
+        "prediction": old_pred,
+        "real": round(1.0 - fake_prob, 4),
+        "fake": round(fake_prob, 4),
+        "time_ms": round(old_ms, 1),
+        "note": "sigmoid single-class output",
+    }
+
+    return (old_result, new_result)
+
+
 # ── UI ────────────────────────────────────────────────────────────────
 
 with gr.Blocks(title="DeepfakeDet-ViT") as demo:
@@ -220,6 +291,16 @@ with gr.Blocks(title="DeepfakeDet-ViT") as demo:
             guide = gr.Markdown("")
             run_btn.click(fn=benchmark, inputs=[], outputs=[bench_md, guide])
 
+        with gr.TabItem("Old vs New"):
+            gr.Markdown("Compare the **deprecated timm wrapper** against the **fixed HF pipeline**. Same weights, different loading paths.")
+            with gr.Row():
+                with gr.Column(scale=1):
+                    cmp_img = gr.Image(type="pil", label="Upload Image")
+                with gr.Column(scale=1):
+                    old_out = gr.JSON(label="Old (timm, deprecated)")
+                    new_out = gr.JSON(label="New (HF, fixed)")
+            cmp_img.change(fn=compare_old_new, inputs=[cmp_img], outputs=[old_out, new_out])
+
         with gr.TabItem("Help"):
             gr.Markdown("""
             **About this app**
@@ -228,6 +309,7 @@ with gr.Blocks(title="DeepfakeDet-ViT") as demo:
 
             - **Compare tab**: Upload an image to see PyTorch and ONNX predictions side by side with timing.
             - **Benchmark tab**: Runs all 8 ONNX variants against all images in the `test_app/images/` directory. Shows predictions and average inference time per variant.
+            - **Old vs New tab**: Compares the deprecated timm wrapper against the fixed HF pipeline — proves the fix is correct.
 
             **Adding test images**
 

test_app/old_config.json

@@ -0,0 +1,29 @@
+{
+  "model": {
+    "variant": "vit_small_patch16_384.augreg_in21k_ft_in1k",
+    "input_size": 384,
+    "patch_size": 16,
+    "freeze_backbone": false,
+    "hidden_dropout_prob": 0.0,
+    "hidden_size": 384,
+    "num_attention_heads": 6,
+    "num_hidden_layers": 12,
+    "attention_probs_dropout_prob": 0.0,
+    "layer_norm_eps": 1e-6,
+    "num_classes": 1,
+    "head": {
+      "in_features": 384,
+      "out_features": 1,
+      "bias": true
+    }
+  },
+  "preprocessing": {
+    "norm_mean": [0.48145466, 0.4578275, 0.40821073],
+    "norm_std": [0.26862954, 0.26130258, 0.27577711],
+    "resize_size": 440,
+    "crop_size": 384
+  },
+  "device": "cuda",
+  "dtype": "float32",
+  "checkpoint_path": "pretrained_weights/model_v11_ViT_384_base_ckpt.pt"
+}