Upload app.py

app.py

@@ -215,36 +215,49 @@ def load_translation_model():
             st.warning("Keras or HuggingFace Hub not available. Model loading skipped.")
             return None
             
-        model = Sequential()
-        model.add(Input(shape=((20, 156))))
-        model.add(keras.layers.Masking(mask_value=0.))
-        model.add(BatchNormalization())
-        model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2, return_sequences=True)))
-        
-        model.add(Dropout(0.2))
-        model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2)))
-        
-        model.add(keras.layers.Activation('elu'))
-        model.add(Dense(32, use_bias=False, kernel_initializer='he_normal'))
-        
-        model.add(BatchNormalization())
-        model.add(Dropout(0.2))
-        model.add(keras.layers.Activation('elu'))
-        model.add(Dense(32, kernel_initializer='he_normal', use_bias=False))
-        
-        model.add(BatchNormalization())
-        model.add(keras.layers.Activation('elu'))
-        model.add(Dropout(0.2))
-        model.add(Dense(len(list(expression_mapping.keys())), activation='softmax'))
-        
-        # Download pre-trained model weights
+        # Download pre-trained model file
         model_file = hf_hub_download(
             repo_id="sunilsarolkar/isl-translation-model",
             filename="isl_model_final.keras"
         )
-        model.load_weights(model_file)
         
-        return model
+        # Try to load the complete model first
+        try:
+            model = keras.models.load_model(model_file)
+            st.success("✅ Model loaded successfully from saved file")
+            return model
+        except Exception as load_error:
+            st.warning(f"Failed to load complete model: {load_error}")
+            st.info("Attempting to build model architecture and load weights...")
+            
+            # Fallback: Build model architecture and load weights
+            model = Sequential()
+            model.add(Input(shape=((20, 156))))
+            model.add(keras.layers.Masking(mask_value=0.))
+            model.add(BatchNormalization())
+            model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2, return_sequences=True)))
+            
+            model.add(Dropout(0.2))
+            model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2)))
+            
+            model.add(keras.layers.Activation('elu'))
+            model.add(Dense(32, use_bias=False, kernel_initializer='he_normal'))
+            
+            model.add(BatchNormalization())
+            model.add(Dropout(0.2))
+            model.add(keras.layers.Activation('elu'))
+            model.add(Dense(32, kernel_initializer='he_normal', use_bias=False))
+            
+            model.add(BatchNormalization())
+            model.add(keras.layers.Activation('elu'))
+            model.add(Dropout(0.2))
+            model.add(Dense(len(list(expression_mapping.keys())), activation='softmax'))
+            
+            # Try to load weights
+            model.load_weights(model_file)
+            st.success("✅ Model architecture built and weights loaded successfully")
+            return model
+        
     except Exception as e:
         st.error(f"Failed to load translation model: {e}")
         return None
@@ -795,32 +808,47 @@ elif app_mode == 'Test Video Translation':
                         # Load translation model
                         translation_model = load_translation_model()
                         
-                        # Make prediction on current window
-                        sequence_idx = frame_idx - 20
-                        prediction_output = translation_model(
-                            X_test_processed[sequence_idx].reshape(
-                                1, X_test_processed[sequence_idx].shape[0], 
-                                X_test_processed[sequence_idx].shape[1]
+                        # Check if model loaded successfully
+                        sequence_idx = frame_idx - 20  # Define sequence_idx for both cases
+                        if translation_model is None:
+                            st.error("❌ Translation model failed to load. Cannot make predictions.")
+                            # Use dummy predictions to keep the visualization working
+                            current_predictions = {"model_not_available": 0.0}
+                            top_3_signs = ["model_not_available"]
+                            top_3_probabilities = [0.0]
+                        else:
+                            # Make prediction on current window
+                            prediction_output = translation_model(
+                                X_test_processed[sequence_idx].reshape(
+                                    1, X_test_processed[sequence_idx].shape[0], 
+                                    X_test_processed[sequence_idx].shape[1]
+                                )
                             )
-                        )
-                        prediction_output = prediction_output[0].cpu().detach().numpy()
-                        
-                        # Get top predictions
-                        top_prediction_idx = np.argmax(prediction_output)
-                        top_3_indices = prediction_output.argsort()[-3:][::-1]
-                        top_3_signs = [expression_mapping[i] for i in top_3_indices]
-                        top_3_probabilities = prediction_output[top_3_indices]
+                            
+                            # Handle both PyTorch and Keras/TensorFlow models
+                            try:
+                                # Try PyTorch tensor operations first
+                                prediction_output = prediction_output[0].cpu().detach().numpy()
+                            except AttributeError:
+                                # If it's a Keras model, it already returns NumPy arrays
+                                prediction_output = prediction_output[0]
+                            
+                            # Get top predictions
+                            top_prediction_idx = np.argmax(prediction_output)
+                            top_3_indices = prediction_output.argsort()[-3:][::-1]
+                            top_3_signs = [expression_mapping[i] for i in top_3_indices]
+                            top_3_probabilities = prediction_output[top_3_indices]
+                            
+                            # Current frame predictions
+                            current_predictions = {}
+                            for sign, prob in zip(top_3_signs, top_3_probabilities):
+                                current_predictions[sign] = prob
                         
                         # Update frame-wise predictions for weighted average
                         for sign, prob in zip(top_3_signs, top_3_probabilities):
                             if sign not in frame_predictions:
                                 frame_predictions[sign] = []
                             frame_predictions[sign].append(prob)
-                        
-                        # Current frame predictions
-                        current_predictions = {}
-                        for sign, prob in zip(top_3_signs, top_3_probabilities):
-                            current_predictions[sign] = prob
 
                         # Calculate weighted averages
                         for sign in frame_predictions:
@@ -868,8 +896,12 @@ elif app_mode == 'Test Video Translation':
                         video_writer.write_frame(canvas_with_predictions)
                         
                         # Get best prediction for display
-                        best_sign = max(weighted_predictions, key=weighted_predictions.get)
-                        best_confidence = weighted_predictions[best_sign]
+                        if weighted_predictions:
+                            best_sign = max(weighted_predictions, key=weighted_predictions.get)
+                            best_confidence = weighted_predictions[best_sign]
+                        else:
+                            best_sign = "no_predictions"
+                            best_confidence = 0.0
                         
                         # Update progress display
                         with progress_container.container():