revert to mlp voice model

.gitignore

@@ -156,10 +156,16 @@ npm-debug.log
 *.tmp
 *.temp
 
-# Model files - will be trained at runtime in Docker
-models/*.joblib
+# Model files - CNN training artifacts (ignored)
 models/*.h5
 models/*.keras
+models/*.pkl
+models/ravdess_cnn_model.*
+models/label_encoder.joblib
+
+# Keep MLP model files (needed for Docker build)
+# models/mlp_emotion_model.joblib
+# models/scaler.joblib
 
 # Keep FFmpeg binaries but ignore temporary files
 ffmpeg/bin/*.log

Dockerfile

@@ -10,7 +10,7 @@ RUN corepack enable && pnpm install --frozen-lockfile
 COPY frontend/ .
 RUN pnpm build
 
-# Stage 2: Python backend with model training at runtime
+# Stage 2: Python backend
 FROM python:3.11-slim
 
 WORKDIR /app
@@ -18,7 +18,6 @@ WORKDIR /app
 # Install system dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
     ffmpeg \
-    git \
     && rm -rf /var/lib/apt/lists/*
 
 # Copy and install Python dependencies
@@ -26,7 +25,7 @@ COPY backend/pyproject.toml backend/uv.lock* ./
 # Grab the uv binary from the official image
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
 
-# Install Python dependencies (includes tensorflow and kagglehub for model training)
+# Install Python dependencies
 RUN uv sync --frozen --no-dev
 
 # Create necessary directories
@@ -35,8 +34,9 @@ RUN mkdir -p ./uploads ./models
 # Copy backend code
 COPY backend/ .
 
-# Copy training script
-COPY models/train_cnn.py ./models/train_cnn.py
+# Copy model files (MLP model and scaler)
+COPY models/mlp_emotion_model.joblib ./models/
+COPY models/scaler.joblib ./models/
 
 # Copy frontend build
 COPY --from=frontend-build /app/frontend/dist ./static
@@ -45,5 +45,4 @@ ENV PORT=7860
 EXPOSE 7860
 
 # Use PORT environment variable (defaults to 7860 for Hugging Face Space compatibility)
-# Model training happens at first startup if models don't exist (see app/main.py)
 CMD ["sh", "-c", "uv run uvicorn app.main:app --host 0.0.0.0 --port ${PORT}"]

backend/app/main.py

@@ -5,7 +5,6 @@ FastAPI Main Application with LangGraph Multi-Agent System
 
 import os
 import re
-import logging
 from dotenv import load_dotenv
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
@@ -14,47 +13,6 @@ from fastapi.staticfiles import StaticFiles
 # Load environment variables
 load_dotenv()
 
-logger = logging.getLogger(__name__)
-
-def train_model_if_missing():
-    """Train the CNN model at startup if model files don't exist (Docker/HF only)."""
-    # Only train in Docker environment, not locally
-    # In Docker, the backend directory is at /app, locally it's at C:\...\backend
-    backend_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
-    is_docker = backend_dir == "/app"
-    
-    if not is_docker:
-        logger.info("Running locally, skipping model training. Using existing model files.")
-        return
-    
-    # In Docker: check if model exists, if not train
-    model_path = os.path.join(backend_dir, "models", "ravdess_cnn_model.h5")
-    
-    if not os.path.exists(model_path):
-        logger.info("Model not found in Docker. Starting model training...")
-        try:
-            import subprocess
-            env = os.environ.copy()
-            env["PYTHONPATH"] = backend_dir
-            result = subprocess.run(
-                ["python", "models/train_cnn.py"],
-                capture_output=True,
-                text=True,
-                env=env,
-                cwd=backend_dir
-            )
-            if result.returncode == 0:
-                logger.info("Model training completed successfully")
-            else:
-                logger.error(f"Model training failed: {result.stderr}")
-        except Exception as e:
-            logger.error(f"Error during model training: {e}")
-    else:
-        logger.info("Model already exists in Docker, skipping training")
-
-# Train model at startup if missing
-train_model_if_missing()
-
 # Import routes
 from app.routes import router
 

backend/app/voice_analysis.py

@@ -1,6 +1,6 @@
 """
 Voice Emotion Analysis API for ShantiView
-Uses the pre-trained CNN model with MFCC sequence features from the RAVDESS dataset
+Uses the pre-trained MLP model with MFCC features from the RAVDESS dataset
 """
 
 import os
@@ -14,63 +14,29 @@ warnings.filterwarnings("ignore")
 
 logger = logging.getLogger(__name__)
 
-# Model type configuration
-# Set to "cnn" or "mlp" - CNN is preferred, MLP is fallback
-MODEL_TYPE = "cnn"  # Can be "cnn" or "mlp"
-
-# Constants for CNN
-MAX_SEQ_LENGTH = 130
+# Constants for MLP
 N_MFCC = 40
-EMOTION_LABELS = ['neutral', 'calm', 'happy', 'sad', 'angry', 'fearful', 'disgust', 'surprised']
 
 # Paths to models
 # __file__ is /app/app/voice_analysis.py in Docker, or C:\...\backend\app\voice_analysis.py locally
-# - os.path.dirname(__file__) = .../backend/app
-# - os.path.dirname(os.path.dirname(__file__)) = .../backend (project root)
 BACKEND_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
-# Check if we're in Docker (/app) or local (C:\...\backend)
+# Check if we're in Docker (/app) or local
 if BACKEND_DIR == "/app":
     MODEL_DIR = os.path.join(BACKEND_DIR, "models")
 else:
     # Local: models are in parent directory of backend
     MODEL_DIR = os.path.join(os.path.dirname(BACKEND_DIR), "models")
 
-# CNN model paths
-CNN_MODEL_PATH = os.path.join(MODEL_DIR, "ravdess_cnn_model.h5")
-LABEL_ENCODER_PATH = os.path.join(MODEL_DIR, "label_encoder.joblib")
-CNN_SCALER_PATH = os.path.join(MODEL_DIR, "scaler.joblib")
-
-# MLP model paths (fallback)
+# MLP model paths
 MLP_MODEL_PATH = os.path.join(MODEL_DIR, "mlp_emotion_model.joblib")
 MLP_SCALER_PATH = os.path.join(MODEL_DIR, "scaler.joblib")
 
 # Model and scaler cache
-_cnn_model = None
-_mlp_model = None
+_model = None
 _scaler = None
-_label_encoder = None
-
-
-def _extract_mfcc_features(file_path, max_len=MAX_SEQ_LENGTH, n_mfcc=N_MFCC):
-    """Extract MFCC sequence features for CNN model."""
-    try:
-        y, sr = librosa.load(file_path, duration=3, offset=0.5, sr=22050)
-        mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=n_mfcc)
-        mfcc = mfcc.T
-        
-        if mfcc.shape[0] < max_len:
-            pad_width = max_len - mfcc.shape[0]
-            mfcc = np.pad(mfcc, pad_width=((0, pad_width), (0, 0)), mode='constant')
-        else:
-            mfcc = mfcc[:max_len]
-        
-        return mfcc
-    except Exception as e:
-        logger.error(f"Error extracting MFCC features from {file_path}: {e}")
-        return None
 
 
-def _extract_mfcc_mean(file_path, n_mfcc=40):
+def _extract_mfcc_mean(file_path, n_mfcc=N_MFCC):
     """Extract mean MFCC features for MLP model."""
     try:
         y, sr = librosa.load(file_path, duration=3, offset=0.5, sr=22050)
@@ -81,48 +47,12 @@ def _extract_mfcc_mean(file_path, n_mfcc=40):
         return None
 
 
-def load_cnn_model():
-    """Load the CNN emotion model, scaler, and label encoder."""
-    global _cnn_model, _scaler, _label_encoder
-    
-    if _cnn_model is not None:
-        return _cnn_model, _scaler, _label_encoder
-    
-    if not os.path.exists(CNN_MODEL_PATH):
-        logger.error(f"CNN model file not found at {CNN_MODEL_PATH}")
-        return None, None, None
-    if not os.path.exists(CNN_SCALER_PATH):
-        logger.error(f"Scaler file not found at {CNN_SCALER_PATH}")
-        return None, None, None
-    if not os.path.exists(LABEL_ENCODER_PATH):
-        logger.error(f"Label encoder file not found at {LABEL_ENCODER_PATH}")
-        return None, None, None
-    
-    try:
-        import tensorflow as tf
-        # Try loading with custom_objects to handle compatibility issues
-        _cnn_model = tf.keras.models.load_model(
-            CNN_MODEL_PATH,
-            custom_objects=None,
-            compile=False  # Skip compilation to avoid deserialization issues
-        )
-        # Recompile the model since we skipped compilation
-        _cnn_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
-        _scaler = joblib.load(CNN_SCALER_PATH)
-        _label_encoder = joblib.load(LABEL_ENCODER_PATH)
-        logger.info("CNN voice emotion model loaded successfully")
-        return _cnn_model, _scaler, _label_encoder
-    except Exception as e:
-        logger.error(f"Error loading CNN model: {e}")
-        return None, None, None
-
-
-def load_mlp_model():
-    """Load the MLP emotion model and scaler (fallback)."""
-    global _mlp_model, _scaler
+def load_model():
+    """Load the MLP emotion model and scaler."""
+    global _model, _scaler
     
-    if _mlp_model is not None:
-        return _mlp_model, _scaler
+    if _model is not None:
+        return _model, _scaler
     
     if not os.path.exists(MLP_MODEL_PATH):
         logger.error(f"MLP model file not found at {MLP_MODEL_PATH}")
@@ -132,63 +62,18 @@ def load_mlp_model():
         return None, None
     
     try:
-        _mlp_model = joblib.load(MLP_MODEL_PATH)
+        _model = joblib.load(MLP_MODEL_PATH)
         _scaler = joblib.load(MLP_SCALER_PATH)
         logger.info("MLP voice emotion model loaded successfully")
-        return _mlp_model, _scaler
+        return _model, _scaler
     except Exception as e:
         logger.error(f"Error loading MLP model: {e}")
         return None, None
 
 
-def load_model():
-    """Load the appropriate model based on configuration."""
-    if MODEL_TYPE == "cnn":
-        model, scaler, le = load_cnn_model()
-        if model is not None:
-            return model, scaler, le, "cnn"
-        logger.warning("CNN model failed to load, falling back to MLP")
-    
-    # Fallback to MLP
-    model, scaler = load_mlp_model()
-    if model is not None:
-        return model, scaler, None, "mlp"
-    
-    logger.error("Both CNN and MLP models failed to load")
-    return None, None, None, None
-
-
-def extract_mfcc_features(file_path, n_mfcc=40):
-    """
-    Extract Mel-frequency cepstral coefficients (MFCCs) from an audio file.
-    
-    Args:
-        file_path: Path to the audio file
-        n_mfcc: Number of MFCCs to extract
-        
-    Returns:
-        numpy array of MFCC features, or None if extraction fails
-    """
-    try:
-        # Load the audio file with parameters matching the training
-        y, sr = librosa.load(file_path, duration=3, offset=0.5, sr=22050)
-        
-        # Extract MFCCs
-        mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=n_mfcc)
-        
-        # Calculate mean across time axis
-        mfccs_mean = np.mean(mfccs.T, axis=0)
-        
-        return mfccs_mean
-    except Exception as e:
-        logger.error(f"Error extracting MFCC features from {file_path}: {e}")
-        return None
-
-
 async def predict_voice_emotion(audio_file_path: str) -> dict:
     """
-    Predict the emotion of an audio file using the trained model.
-    Supports both CNN and MLP models.
+    Predict the emotion of an audio file using the trained MLP model.
     
     Args:
         audio_file_path: Path to the audio file
@@ -198,8 +83,7 @@ async def predict_voice_emotion(audio_file_path: str) -> dict:
     """
     try:
         # Load model
-        result = load_model()
-        model, scaler, label_encoder, model_type = result
+        model, scaler = load_model()
         
         if model is None:
             return {
@@ -208,75 +92,38 @@ async def predict_voice_emotion(audio_file_path: str) -> dict:
                 "message": "Voice emotion model is not loaded. Please ensure model files exist."
             }
         
-        if model_type == "cnn":
-            # CNN prediction
-            features = _extract_mfcc_features(audio_file_path)
-            
-            if features is None:
-                return {
-                    "error": True,
-                    "emotion": "Feature extraction failed",
-                    "message": "Could not extract features from audio file"
-                }
-            
-            # Scale features (reshape for 2D scaler, then reshape back)
-            original_shape = features.shape
-            features_scaled = scaler.transform(features.reshape(-1, original_shape[1])).reshape(original_shape)
-            
-            # Add batch dimension and predict
-            features_batch = np.expand_dims(features_scaled, axis=0)
-            predictions = model.predict(features_batch, verbose=0)[0]
-            
-            # Get emotion from predictions
-            predicted_class = np.argmax(predictions)
-            confidence = float(predictions[predicted_class])
-            
-            # Use label encoder if available, otherwise use emotion labels
-            if label_encoder is not None:
-                emotion_display = label_encoder.inverse_transform([predicted_class])[0]
-                emotion_labels = label_encoder.classes_
-            else:
-                emotion_display = EMOTION_LABELS[predicted_class]
-                emotion_labels = EMOTION_LABELS
-            
-            # Build all_emotions dict using the correct label order (convert to native Python types)
-            emotion_probs = {str(label): float(predictions[i]) for i, label in enumerate(emotion_labels)}
-            
-        else:
-            # MLP prediction (fallback)
-            features = _extract_mfcc_mean(audio_file_path)
-            
-            if features is None:
-                return {
-                    "error": True,
-                    "emotion": "Feature extraction failed",
-                    "message": "Could not extract features from audio file"
-                }
-            
-            # Scale features and predict
-            features_scaled = scaler.transform(features.reshape(1, -1))
-            prediction = model.predict(features_scaled)[0]
-            
-            # Get probabilities if available
-            try:
-                probabilities = model.predict_proba(features_scaled)[0]
-                emotion_probs = {label: float(prob) for label, prob in zip(model.classes_, probabilities)}
-                confidence = float(max(probabilities))
-            except Exception:
-                emotion_probs = {}
-                confidence = 1.0
-            
-            emotion_display = prediction.capitalize()
+        # Extract features
+        features = _extract_mfcc_mean(audio_file_path)
+        
+        if features is None:
+            return {
+                "error": True,
+                "emotion": "Feature extraction failed",
+                "message": "Could not extract features from audio file"
+            }
+        
+        # Scale features and predict
+        features_scaled = scaler.transform(features.reshape(1, -1))
+        prediction = model.predict(features_scaled)[0]
+        
+        # Get probabilities if available
+        try:
+            probabilities = model.predict_proba(features_scaled)[0]
+            emotion_probs = {str(label): float(prob) for label, prob in zip(model.classes_, probabilities)}
+            confidence = float(max(probabilities))
+        except Exception:
+            emotion_probs = {}
+            confidence = 1.0
         
-        emotion_display = emotion_display.capitalize()
-        logger.info(f"Voice emotion prediction ({model_type}): {emotion_display} (confidence: {confidence:.3f})")
+        emotion_display = str(prediction).capitalize()
+        logger.info(f"Voice emotion prediction: {emotion_display} (confidence: {confidence:.3f})")
         
         return {
             "error": False,
             "emotion": emotion_display,
             "confidence": confidence,
             "all_emotions": emotion_probs,
-            "model_type": model_type
+            "model_type": "mlp"
         }
         
     except Exception as e:
@@ -285,4 +132,4 @@ async def predict_voice_emotion(audio_file_path: str) -> dict:
             "error": True,
             "emotion": "Error",
             "message": str(e)
-        }
+        }

backend/pyproject.toml

@@ -22,9 +22,6 @@ dependencies = [
     "Pillow>=11.0.0",
     "aiohttp>=3.10.0",
     "deepface>=0.0.93",
-    "tensorflow>=2.15.0",
-    "tf-keras>=2.18.0",
-    "kagglehub>=0.3.0",
 ]
 
 [dependency-groups]

models/train_cnn.py

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:615316b4c5ea006dcf082eb520c35275fcb6fdc48eb686110d8d66baf50d8495
-size 6752
+oid sha256:299eed1888b90f9425435d33b185544857a5ab53af4c9badbc79ad9bb460bf99
+size 6842