fix: regenerate model.safetensors from correct .pt checkpoint

The old safetensors was converted from different weights (intermediate_size=3072,
num_classes=2 classifier, different backbone weights) instead of the correct
trained checkpoint (intermediate_size=1536, num_classes=1, ViT-Small 6 heads).

Changes:
- model.safetensors: regenerated from pretrained_weights/model_v11_ViT_384_base_ckpt.pt
- config.json: num_classes=2→1, intermediate_size=3072→1536, correct for ViT-Small
- preprocessor_config.json: size changed to shortest_edge=440 (no more squashing)
- test_app: handle num_classes=1 with sigmoid output

.gitignore

@@ -1 +1,3 @@
 .agents/
+__pycache__/
+*.pyc

config.json

@@ -7,22 +7,20 @@
   "hidden_dropout_prob": 0.0,
   "hidden_size": 384,
   "id2label": {
-    "0": "real",
-    "1": "fake"
+    "0": "real"
   },
   "image_size": 384,
   "initializer_range": 0.02,
   "input_size": 384,
-  "intermediate_size": 3072,
+  "intermediate_size": 1536,
   "label2id": {
-    "fake": 1,
     "real": 0
   },
   "layer_norm_eps": 1e-06,
   "model_type": "vit",
   "num_attention_heads": 6,
   "num_channels": 3,
-  "num_classes": 2,
+  "num_classes": 1,
   "num_hidden_layers": 12,
   "patch_size": 16,
   "qkv_bias": true,

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91a3c4d1b7d453b4b74e377c8e079662a6cc2c97f8ca448daedda8864ec54605
-size 143969256
+oid sha256:275ba982236ddd6afddf7131f8133e89f537574b964cf8fa5825b4956d741692
+size 87270764

preprocessor_config.json

@@ -7,7 +7,6 @@
   "image_std": [0.26862954, 0.26130258, 0.27577711],
   "resample": 3,
   "size": {
-    "height": 440,
-    "width": 440
+    "shortest_edge": 440
   }
 }

test_app/app.py

@@ -86,9 +86,11 @@ def _get_onnx(variant):
     return _onnx_sessions[variant]
 
 def _onnx_preprocess(image):
-    img = image.resize((440, 440))
-    left = (440 - 384) // 2
-    top = (440 - 384) // 2
+    w, h = image.size
+    scale = 440 / min(w, h)
+    img = image.resize((int(w * scale), int(h * scale)))
+    left = (img.size[0] - 384) // 2
+    top = (img.size[1] - 384) // 2
     img = img.crop((left, top, left + 384, top + 384))
     arr = np.array(img, dtype=np.float32)
     arr *= np.float32(1.0 / 255.0)
@@ -109,15 +111,21 @@ def compare(image, onnx_variants):
     inputs = {k: v.to(pt_device) for k, v in inputs.items()}
     with torch.no_grad():
         logits = pt_model(**inputs).logits[0].cpu()
-        pt_probs = torch.softmax(logits, dim=-1)
-        pt_pred = pt_model.config.id2label[torch.argmax(pt_probs).item()]
+        if pt_model.config.num_classes == 1:
+            fake_prob = torch.sigmoid(logits).item()
+            pt_pred = "fake" if fake_prob > 0.5 else "real"
+            pt_probs = [1.0 - fake_prob, fake_prob]
+        else:
+            pt_probs_arr = torch.softmax(logits, dim=-1)
+            pt_pred = pt_model.config.id2label[torch.argmax(pt_probs_arr).item()]
+            pt_probs = [pt_probs_arr[0].item(), pt_probs_arr[1].item()]
     pt_ms = (time.perf_counter() - t0) * 1000
 
     pt = {
         "device": pt_device.upper(),
         "prediction": pt_pred,
-        "real": round(pt_probs[0].item(), 4),
-        "fake": round(pt_probs[1].item(), 4),
+        "real": round(pt_probs[0], 4),
+        "fake": round(pt_probs[1], 4),
         "time_ms": round(pt_ms, 1),
     }
 
@@ -236,15 +244,15 @@ def compare_old_new(image):
     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()]
+        fake_prob = torch.sigmoid(logits).item()
+        new_pred = "fake" if fake_prob > 0.5 else "real"
     new_ms = (time.perf_counter() - t0) * 1000
 
     new_result = {
-        "backend": "ViTForImageClassification (fixed config, July 2026)",
+        "backend": "ViTForImageClassification (fixed, July 2026)",
         "prediction": new_pred,
-        "real": round(new_probs[0].item(), 4),
-        "fake": round(new_probs[1].item(), 4),
+        "real": round(1.0 - fake_prob, 4),
+        "fake": round(fake_prob, 4),
         "time_ms": round(new_ms, 1),
     }