Upload 4 files

README.md

@@ -0,0 +1,22 @@
+---
+license: apache-2.0
+tags:
+- vision-transformer
+- deepfake-detection
+- image-classification
+---
+
+# DeiT Base Distilled – DeepFakeDetector
+
+Data-efficient Image Transformer (DeiT-B/16) fine-tuned for binary deepfake detection. This model utilizes distillation to improve performance and efficiency.
+
+- Retrained using a mix of OpenFake, WildFake, Dragon, and manual images (with a focus on classifying real images)
+
+## Labels
+
+- **0**: real
+- **1**: fake
+
+## Output
+
+- **prob_fake** ∈ [0, 1] (via softmax or sigmoid)

config.json

@@ -0,0 +1,44 @@
+{
+  "model_name": "deit_base_distilled_patch16_224",
+  "library": "timm",
+  "input_size": 224,
+  "num_classes": 2,
+  "class_mapping": {
+    "0": "real",
+    "1": "fake"
+  },
+  "distilled": true,
+  "logits_handling": {
+    "mode": "first",
+    "description": "During training, model returns tuple (logits, dist_logits). Use outputs[0] for predictions. During eval with model.eval(), single tensor is returned."
+  },
+  "normalization": {
+    "scheme": "imagenet",
+    "mean": [0.485, 0.456, 0.406],
+    "std": [0.229, 0.224, 0.225]
+  },
+  "head": {
+    "architecture": [
+      {"type": "LayerNorm", "features": 768},
+      {"type": "Linear", "in_features": 768, "out_features": 512},
+      {"type": "GELU"},
+      {"type": "Dropout", "p": 0.2},
+      {"type": "Linear", "in_features": 512, "out_features": 2}
+    ],
+    "note": "Custom MLP head replaces both model.head and model.head_dist (shared weights)"
+  },
+  "training": {
+    "optimizer": "AdamW",
+    "learning_rates": {
+      "backbone": 2e-5,
+      "head": 5e-5
+    },
+    "weight_decay": 1e-4,
+    "gradient_clipping": {
+      "max_norm": 1.0
+    },
+    "epochs": 2,
+    "batch_size": 16,
+    "criterion": "CrossEntropyLoss"
+  }
+}

label_map.json

@@ -0,0 +1,4 @@
+{
+    "0": "real",
+    "1": "fake"
+}

preprocess.json

@@ -0,0 +1,16 @@
+{
+    "input_size": 224,
+    "interpolation": "bicubic",
+    "normalize": {
+        "mean": [
+            0.485,
+            0.456,
+            0.406
+        ],
+        "std": [
+            0.229,
+            0.224,
+            0.225
+        ]
+    }
+}