Upload folder using huggingface_hub

config.json

@@ -0,0 +1,23 @@
+{
+  "type": "probability_stacking_fusion",
+  "name": "fusion-logreg-v2",
+  "version": "1.0.0",
+  "submodels": [
+    "DeepFakeDetector/cnn-transfer",
+    "DeepFakeDetector/vit-base",
+    "DeepFakeDetector/deit-distilled",
+    "DeepFakeDetector/gradfield-cnn"
+  ],
+  "submodel_order": [
+    "cnn-transfer",
+    "vit-base",
+    "deit-distilled",
+    "gradfield-cnn"
+  ],
+  "num_submodels": 4,
+  "threshold": 0.5,
+  "labels": {
+    "0": "real",
+    "1": "fake"
+  }
+}

fusion_logreg.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79acc3405b144acd008605109665956c1c3315def0ef99f906ea8511c2331680
+size 895

label_map.json

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

predict.py

@@ -0,0 +1,59 @@
+
+import numpy as np
+import os
+import json
+
+def predict(submodel_outputs: dict, **kwargs) -> dict:
+    '''
+    Standard Fusion Predict Interface
+    Args:
+        submodel_outputs: dict mapping model_name -> {"prob_fake": float}
+    '''
+    # Load config to get order
+    base_path = os.path.dirname(__file__)
+    with open(os.path.join(base_path, "config.json")) as f:
+        config = json.load(f)
+
+    order = config["submodel_order"]
+    probs = []
+    for name in order:
+        if name not in submodel_outputs:
+            raise ValueError(f"Missing output for {name}")
+        probs.append(submodel_outputs[name]["prob_fake"])
+
+    X = np.array([probs]) # (1, n_models)
+
+    # Load Model
+    # Detect if logreg or pytorch
+    if os.path.exists(os.path.join(base_path, "fusion_logreg.pkl")):
+        import joblib
+        model = joblib.load(os.path.join(base_path, "fusion_logreg.pkl"))
+        prob_fake = model.predict_proba(X)[0, 1]
+    elif os.path.exists(os.path.join(base_path, "fusion_model.pt")):
+        import torch
+        import torch.nn as nn
+        # Simple reconstruction of architecture (must match training)
+        # For robustness, one might pickle the whole model or save arch config.
+        # Here we assume the simple MLP structure used in notebook.
+        input_dim = len(order)
+        net = nn.Sequential(
+            nn.Linear(input_dim, 32),
+            nn.ReLU(),
+            nn.Dropout(0.3),
+            nn.Linear(32, 1),
+            nn.Sigmoid()
+        )
+        net.load_state_dict(torch.load(os.path.join(base_path, "fusion_model.pt")))
+        net.eval()
+        with torch.no_grad():
+            prob_fake = net(torch.tensor(X, dtype=torch.float32)).item()
+    else:
+        raise FileNotFoundError("No model file found")
+
+    return {
+        "pred": "fake" if prob_fake >= config.get("threshold", 0.5) else "real",
+        "pred_int": 1 if prob_fake >= config.get("threshold", 0.5) else 0,
+        "prob_fake": float(prob_fake),
+        "meta": {"min_prob": float(np.min(probs)), "max_prob": float(np.max(probs))}
+    }
+    

test_metrics.json

@@ -0,0 +1,12 @@
+{
+  "accuracy": 0.8840116682872621,
+  "auroc": 0.9519397363465161,
+  "f1": 0.8845568920226738,
+  "model_type": "Logistic Regression Fusion",
+  "submodel_order": [
+    "cnn-transfer",
+    "vit-base",
+    "deit-distilled",
+    "gradfield-cnn"
+  ]
+}