initial commit

app.py

@@ -126,30 +126,36 @@ def load_frames_from_video(video_path, transform, num_frames=10):
     frames = frames.unsqueeze(0)  # Add batch dimension
     return frames
 
-# Prediction function for a single video
 def predict_video(model, video_path, text_input, tokenizer, transform):
-    # Set model to evaluation mode
-    model.eval()
-    
-    # Tokenize the text input and move to device
-    encoding = tokenizer(
-        text_input, padding='max_length', truncation=True, max_length=128, return_tensors='pt'
-    )
-    encoding = {key: val.to(device) for key, val in encoding.items()}  # Ensure text input is on the device
-    
-    # Load frames from the video and move to device
-    frames = load_frames_from_video(video_path, transform)
-    frames = frames.to(device)  # Ensure frames are on the device
-    
-    # Perform forward pass through the model
-    with torch.no_grad():
-        output = model(encoding, frames)
-    
-    # Apply sigmoid to get probability, then threshold to get prediction
-    prediction = (output.squeeze(-1) > 0.5).float()
+    try:
+        # Set model to evaluation mode
+        model.eval()
+        
+        # Tokenize the text input
+        encoding = tokenizer(
+            text_input, padding='max_length', truncation=True, max_length=128, return_tensors='pt'
+        )
+        encoding = {key: val.to(device) for key, val in encoding.items()}
+        
+        # Load frames from the video
+        frames = load_frames_from_video(video_path, transform)
+        frames = frames.to(device)
+        
+        # Log input shapes and devices
+        print(f"Encoding device: {next(iter(encoding.values())).device}, Frames shape: {frames.shape}")
+        
+        # Perform forward pass through the model
+        with torch.no_grad():
+            output = model(encoding, frames)
+        
+        # Apply sigmoid to get probability, then threshold to get prediction
+        prediction = (output.squeeze(-1) > 0.5).float()
+        
+        return prediction.item()
     
-    # Return the predicted label (0 or 1)
-    return prediction.item()
+    except Exception as e:
+        print(f"Prediction error: {e}")
+        return "Error during prediction"