Update app.py

app.py

@@ -7,19 +7,19 @@ import os
 # ==================================================
 # CONFIGURATION
 # ==================================================
-MODEL_PATH = "./models/deepfake_detector_multi_input.h5"
+MODEL_PATH = "./models/deepfake_detector_multi_input_v3_robust.h5"
 IMG_SIZE = (224, 224)
-NUM_FRAMES = 20  # frames sampled per video
+NUM_FRAMES = 20  # number of frames sampled per video
 CASCADE_PATH = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
 
 # ==================================================
-# MODEL LOADING
+# LOAD MODEL
 # ==================================================
 print("[INFO] Loading TensorFlow model...")
 model = tf.keras.models.load_model(MODEL_PATH)
 print("[INFO] Model loaded successfully!")
 
-# Enable GPU memory growth
+# Enable GPU memory growth (optional, prevents OOM)
 gpus = tf.config.experimental.list_physical_devices('GPU')
 if gpus:
     for g in gpus:
@@ -29,10 +29,10 @@ if gpus:
 face_cascade = cv2.CascadeClassifier(CASCADE_PATH)
 
 # ==================================================
-# FACE EXTRACTION
+# HELPER: Extract faces
 # ==================================================
 def extract_faces(video_path, num_frames=NUM_FRAMES, size=IMG_SIZE):
-    """Extract faces (or full frames if none detected) from uploaded video."""
+    """Extract faces (or full frames if no faces detected) from a video."""
     cap = cv2.VideoCapture(video_path)
     total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
     step = max(1, total_frames // num_frames)
@@ -61,17 +61,25 @@ def extract_faces(video_path, num_frames=NUM_FRAMES, size=IMG_SIZE):
     return np.array(frames)
 
 # ==================================================
-# PREDICTION PIPELINE
+# PREDICTION FUNCTION
 # ==================================================
 def predict(video):
     if not video:
         return "⚠️ Please upload a video."
+
     try:
         frames = extract_faces(video)
-        preds = model.predict(frames, verbose=0)
+
+        # Dummy handcrafted feature vector (3 features per frame)
+        dummy_features = np.zeros((frames.shape[0], 3), dtype=np.float32)
+
+        # Run inference with both inputs
+        preds = model.predict([frames, dummy_features], verbose=0)
         score = float(np.mean(preds))
         label = "🧠 FAKE" if score > 0.5 else "✅ REAL"
+
         return f"**Prediction:** {label}\nConfidence: {score:.2f}"
+
     except Exception as e:
         return f"❌ Error processing video: {e}"
 
@@ -82,11 +90,11 @@ demo = gr.Interface(
     fn=predict,
     inputs=gr.Video(label="🎥 Upload a short video (≤ 20 s)"),
     outputs=gr.Markdown(),
-    title="Deepfake Detection Demo",
+    title="Deepfake Detection Demo (Video + Dummy Features)",
     description=(
-        "Upload a short video clip to test the model. "
-        "The app extracts faces, runs inference on each frame, "
-        "and reports the average confidence for REAL vs FAKE."
+        "Upload a short video to test the model. "
+        "This demo uses your trained multimodal deepfake detector, "
+        "with dummy blink features for quick inference."
     ),
 )