app_moe.py

import os
import sys
import torch
import numpy as np
import gradio as gr
import torchaudio
import torchvision
import spaces
import json

# Add parent directory to path to import preprocess functions
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

# Import functions from preprocess and model definitions
from preprocess import process_image_data
from evaluate_backbones import WatermelonModelModular, IMAGE_BACKBONES, AUDIO_BACKBONES

# Define the top-performing models based on evaluation
TOP_MODELS = [
    {"image_backbone": "efficientnet_b3", "audio_backbone": "transformer"},
    {"image_backbone": "efficientnet_b0", "audio_backbone": "transformer"},
    {"image_backbone": "resnet50", "audio_backbone": "transformer"}
]

# Define the MoE Model
class WatermelonMoEModel(torch.nn.Module):
    def __init__(self, model_configs, model_dir="models", weights=None):
        """
        Mixture of Experts model that combines multiple backbone models.
        
        Args:
            model_configs: List of dictionaries with 'image_backbone' and 'audio_backbone' keys
            model_dir: Directory where model checkpoints are stored
            weights: Optional list of weights for each model (None for equal weighting)
        """
        super(WatermelonMoEModel, self).__init__()
        self.models = torch.nn.ModuleList()  # Use ModuleList instead of regular list
        self.model_configs = model_configs
        
        # Load each model
        loaded_count = 0
        for config in model_configs:
            img_backbone = config["image_backbone"]
            audio_backbone = config["audio_backbone"]
            
            # Initialize model
            model = WatermelonModelModular(img_backbone, audio_backbone)
            
            # Load weights
            model_path = os.path.join(model_dir, f"{img_backbone}_{audio_backbone}_model.pt")
            if os.path.exists(model_path):
                print(f"\033[92mINFO\033[0m: Loading model {img_backbone}_{audio_backbone} from {model_path}")
                try:
                    model.load_state_dict(torch.load(model_path, map_location='cpu'))
                    model.eval()  # Set to evaluation mode
                    self.models.append(model)
                    loaded_count += 1
                except Exception as e:
                    print(f"\033[91mERR!\033[0m: Failed to load model from {model_path}: {e}")
                    continue
            else:
                print(f"\033[91mERR!\033[0m: Model checkpoint not found at {model_path}")
                continue
        
        # Add a dummy parameter if no models were loaded to prevent StopIteration
        if loaded_count == 0:
            print(f"\033[91mERR!\033[0m: No models were successfully loaded!")
            self.dummy_param = torch.nn.Parameter(torch.zeros(1))
            
        # Set model weights (uniform by default)
        if weights and loaded_count > 0:
            assert len(weights) == len(self.models), "Number of weights must match number of models"
            self.weights = weights
        else:
            self.weights = [1.0 / max(loaded_count, 1)] * max(loaded_count, 1)
            
        print(f"\033[92mINFO\033[0m: Loaded {loaded_count} models for MoE ensemble")
        print(f"\033[92mINFO\033[0m: Model weights: {self.weights}")

    def to(self, device):
        """
        Override to() method to ensure all sub-models are moved to the same device
        """
        for model in self.models:
            model.to(device)
        return super(WatermelonMoEModel, self).to(device)

    def forward(self, mfcc, image):
        """
        Forward pass through the MoE model.
        Returns the weighted average of all model outputs.
        """
        # Check if we have models loaded
        if not self.models:
            print(f"\033[91mERR!\033[0m: No models available for inference!")
            return torch.tensor([0.0], device=mfcc.device)  # Return a default value
            
        outputs = []
        
        # Get outputs from each model
        with torch.no_grad():
            for i, model in enumerate(self.models):
                output = model(mfcc, image)
                outputs.append(output * self.weights[i])
        
        # Return weighted average
        return torch.sum(torch.stack(outputs), dim=0)

# Modified version of process_audio_data specifically for the app to handle various tensor shapes
def app_process_audio_data(waveform, sample_rate):
    """Modified version of process_audio_data for the app that handles different tensor dimensions"""
    try:
        print(f"\033[92mDEBUG\033[0m: Processing audio - Initial shape: {waveform.shape}, Sample rate: {sample_rate}")
        
        # Handle different tensor dimensions
        if waveform.dim() == 3:
            print(f"\033[92mDEBUG\033[0m: Found 3D tensor, converting to 2D")
            # For 3D tensor, take the first item (batch dimension)
            waveform = waveform[0]
            
        if waveform.dim() == 2:
            # Use the first channel for stereo audio
            waveform = waveform[0]
            print(f"\033[92mDEBUG\033[0m: Using first channel, new shape: {waveform.shape}")
        
        # Resample to 16kHz if needed
        resample_rate = 16000
        if sample_rate != resample_rate:
            print(f"\033[92mDEBUG\033[0m: Resampling from {sample_rate}Hz to {resample_rate}Hz")
            waveform = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=resample_rate)(waveform)
        
        # Ensure 3 seconds of audio
        if waveform.size(0) < 3 * resample_rate:
            print(f"\033[92mDEBUG\033[0m: Padding audio from {waveform.size(0)} to {3 * resample_rate} samples")
            waveform = torch.nn.functional.pad(waveform, (0, 3 * resample_rate - waveform.size(0)))
        else:
            print(f"\033[92mDEBUG\033[0m: Trimming audio from {waveform.size(0)} to {3 * resample_rate} samples")
            waveform = waveform[: 3 * resample_rate]
        
        # Apply MFCC transformation
        print(f"\033[92mDEBUG\033[0m: Applying MFCC transformation")
        mfcc_transform = torchaudio.transforms.MFCC(
            sample_rate=resample_rate,
            n_mfcc=13,
            melkwargs={
                "n_fft": 256,
                "win_length": 256,
                "hop_length": 128,
                "n_mels": 40,
            }
        )
        
        mfcc = mfcc_transform(waveform)
        print(f"\033[92mDEBUG\033[0m: MFCC output shape: {mfcc.shape}")
        
        return mfcc
    except Exception as e:
        import traceback
        print(f"\033[91mERR!\033[0m: Error in audio processing: {e}")
        print(traceback.format_exc())
        return None

# Using the decorator for GPU acceleration
@spaces.GPU
def predict_sugar_content(audio, image, model_dir="models", weights=None):
    """Function with GPU acceleration to predict watermelon sugar content in Brix using MoE model"""
    try:
        # Check CUDA availability inside the GPU-decorated function
        if torch.cuda.is_available():
            device = torch.device("cuda")
            print(f"\033[92mINFO\033[0m: CUDA is available. Using device: {device}")
        else:
            device = torch.device("cpu")
            print(f"\033[92mINFO\033[0m: CUDA is not available. Using device: {device}")
        
        # Load MoE model
        moe_model = WatermelonMoEModel(TOP_MODELS, model_dir, weights)
        # Explicitly move the entire model to device
        moe_model = moe_model.to(device)
        moe_model.eval()
        print(f"\033[92mINFO\033[0m: Loaded MoE model with {len(moe_model.models)} backbone models")
        
        # Debug information about input types
        print(f"\033[92mDEBUG\033[0m: Audio input type: {type(audio)}")
        print(f"\033[92mDEBUG\033[0m: Audio input shape/length: {len(audio)}")
        print(f"\033[92mDEBUG\033[0m: Image input type: {type(image)}")
        if isinstance(image, np.ndarray):
            print(f"\033[92mDEBUG\033[0m: Image input shape: {image.shape}")
        
        # Handle different audio input formats
        if isinstance(audio, tuple) and len(audio) == 2:
            # Standard Gradio format: (sample_rate, audio_data)
            sample_rate, audio_data = audio
            print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
            print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
        elif isinstance(audio, tuple) and len(audio) > 2:
            # Sometimes Gradio returns (sample_rate, audio_data, other_info...)
            sample_rate, audio_data = audio[0], audio[-1]
            print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
            print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
        elif isinstance(audio, str):
            # Direct path to audio file
            audio_data, sample_rate = torchaudio.load(audio)
            print(f"\033[92mDEBUG\033[0m: Loaded audio from path with shape: {audio_data.shape}")
        else:
            return f"Error: Unsupported audio format. Got {type(audio)}"
        
        # Create a temporary file path for the audio and image
        temp_dir = "temp"
        os.makedirs(temp_dir, exist_ok=True)
        
        temp_audio_path = os.path.join(temp_dir, "temp_audio.wav")
        temp_image_path = os.path.join(temp_dir, "temp_image.jpg")
        
        # Import necessary libraries
        from PIL import Image
        
        # Audio handling - direct processing from the data in memory
        if isinstance(audio_data, np.ndarray):
            # Convert numpy array to tensor
            print(f"\033[92mDEBUG\033[0m: Converting numpy audio with shape {audio_data.shape} to tensor")
            audio_tensor = torch.tensor(audio_data).float()
            
            # Handle different audio dimensions
            if audio_data.ndim == 1:
                # Single channel audio
                audio_tensor = audio_tensor.unsqueeze(0)
            elif audio_data.ndim == 2:
                # Ensure channels are first dimension
                if audio_data.shape[0] > audio_data.shape[1]:
                    # More rows than columns, probably (samples, channels)
                    audio_tensor = torch.tensor(audio_data.T).float()
        else:
            # Already a tensor
            audio_tensor = audio_data.float()
        
        print(f"\033[92mDEBUG\033[0m: Audio tensor shape before processing: {audio_tensor.shape}")
        
        # Skip saving/loading and process directly
        mfcc = app_process_audio_data(audio_tensor, sample_rate)
        print(f"\033[92mDEBUG\033[0m: MFCC tensor shape after processing: {mfcc.shape if mfcc is not None else None}")
        
        # Image handling
        if isinstance(image, np.ndarray):
            print(f"\033[92mDEBUG\033[0m: Converting numpy image with shape {image.shape} to PIL")
            pil_image = Image.fromarray(image)
            pil_image.save(temp_image_path)
            print(f"\033[92mDEBUG\033[0m: Saved image to {temp_image_path}")
        elif isinstance(image, str):
            # If image is already a path
            temp_image_path = image
            print(f"\033[92mDEBUG\033[0m: Using provided image path: {temp_image_path}")
        else:
            return f"Error: Unsupported image format. Got {type(image)}"
        
        # Process image
        print(f"\033[92mDEBUG\033[0m: Loading and preprocessing image from {temp_image_path}")
        image_tensor = torchvision.io.read_image(temp_image_path)
        print(f"\033[92mDEBUG\033[0m: Loaded image shape: {image_tensor.shape}")
        image_tensor = image_tensor.float()
        processed_image = process_image_data(image_tensor)
        print(f"\033[92mDEBUG\033[0m: Processed image shape: {processed_image.shape if processed_image is not None else None}")
        
        # Add batch dimension for inference and move to device
        if mfcc is not None:
            # Ensure mfcc is on the same device as the model
            mfcc = mfcc.unsqueeze(0).to(device)
            print(f"\033[92mDEBUG\033[0m: Final MFCC shape with batch dimension: {mfcc.shape}, device: {mfcc.device}")
        
        if processed_image is not None:
            # Ensure processed_image is on the same device as the model
            processed_image = processed_image.unsqueeze(0).to(device)
            print(f"\033[92mDEBUG\033[0m: Final image shape with batch dimension: {processed_image.shape}, device: {processed_image.device}")
        
        # Double-check model is on the correct device
        print(f"\033[92mDEBUG\033[0m: MoE model device: {next(moe_model.parameters()).device}")
        for i, model in enumerate(moe_model.models):
            print(f"\033[92mDEBUG\033[0m: Model {i} device: {next(model.parameters()).device}")
        
        # Run inference with MoE model
        print(f"\033[92mDEBUG\033[0m: Running inference with MoE model on device: {device}")
        if mfcc is not None and processed_image is not None:
            with torch.no_grad():
                brix_value = moe_model(mfcc, processed_image)
                print(f"\033[92mDEBUG\033[0m: Prediction successful: {brix_value.item()}")
        else:
            return "Error: Failed to process inputs. Please check the debug logs."
        
        # Format the result with a range display
        if brix_value is not None:
            brix_score = brix_value.item()
            
            # Create a header with the numerical result
            result = f"🍉 Predicted Sugar Content: {brix_score:.1f}° Brix 🍉\n\n"
            
            # Add extra info about the MoE model
            result += "Using Ensemble of Top-3 Models:\n"
            result += "- EfficientNet-B3 + Transformer\n"
            result += "- EfficientNet-B0 + Transformer\n"
            result += "- ResNet-50 + Transformer\n\n"
            
            # Add Brix scale visualization
            result += "Sugar Content Scale (in °Brix):\n"
            result += "──────────────────────────────────\n"
            
            # Create the scale display with Brix ranges
            scale_ranges = [
                (0, 8, "Low Sugar (< 8° Brix)"),
                (8, 9, "Mild Sweetness (8-9° Brix)"),
                (9, 10, "Medium Sweetness (9-10° Brix)"),
                (10, 11, "Sweet (10-11° Brix)"),
                (11, 13, "Very Sweet (11-13° Brix)")
            ]
            
            # Find which category the prediction falls into
            user_category = None
            for min_val, max_val, category_name in scale_ranges:
                if min_val <= brix_score < max_val:
                    user_category = category_name
                    break
            if brix_score >= scale_ranges[-1][0]:  # Handle edge case
                user_category = scale_ranges[-1][2]
            
            # Display the scale with the user's result highlighted
            for min_val, max_val, category_name in scale_ranges:
                if category_name == user_category:
                    result += f"▶ {min_val}-{max_val}: {category_name} ◀ (YOUR WATERMELON)\n"
                else:
                    result += f"  {min_val}-{max_val}: {category_name}\n"
            
            result += "──────────────────────────────────\n\n"
            
            # Add assessment of the watermelon's sugar content
            if brix_score < 8:
                result += "Assessment: This watermelon has low sugar content. It may taste bland or slightly bitter."
            elif brix_score < 9:
                result += "Assessment: This watermelon has mild sweetness. Acceptable flavor but not very sweet."
            elif brix_score < 10:
                result += "Assessment: This watermelon has moderate sugar content. It should have pleasant sweetness."
            elif brix_score < 11:
                result += "Assessment: This watermelon has good sugar content! It should be sweet and juicy."
            else:
                result += "Assessment: This watermelon has excellent sugar content! Perfect choice for maximum sweetness and flavor."
                
            return result
        else:
            return "Error: Could not predict sugar content. Please try again with different inputs."
    
    except Exception as e:
        import traceback
        error_msg = f"Error: {str(e)}\n\n"
        error_msg += traceback.format_exc()
        print(f"\033[91mERR!\033[0m: {error_msg}")
        return error_msg

def create_app(model_dir="models", weights=None):
    """Create and launch the Gradio interface"""
    # Define the prediction function with model path
    def predict_fn(audio, image):
        return predict_sugar_content(audio, image, model_dir, weights)
    
    # Create Gradio interface
    with gr.Blocks(title="Watermelon Sugar Content Predictor (MoE)", theme=gr.themes.Soft()) as interface:
        gr.Markdown("# 🍉 Watermelon Sugar Content Predictor (Ensemble Model)")
        gr.Markdown("""
        This app predicts the sugar content (in °Brix) of a watermelon based on its sound and appearance.
        
        ## What's New
        This version uses a Mixture of Experts (MoE) ensemble model that combines the three best-performing models:
        - EfficientNet-B3 + Transformer
        - EfficientNet-B0 + Transformer
        - ResNet-50 + Transformer
        
        The ensemble approach provides more accurate predictions than any single model!
        
        ## Instructions:
        1. Upload or record an audio of tapping the watermelon
        2. Upload or capture an image of the watermelon
        3. Click 'Predict' to get the sugar content estimation
        """)
        
        with gr.Row():
            with gr.Column():
                audio_input = gr.Audio(label="Upload or Record Audio", type="numpy")
                image_input = gr.Image(label="Upload or Capture Image")
                submit_btn = gr.Button("Predict Sugar Content", variant="primary")
            
            with gr.Column():
                output = gr.Textbox(label="Prediction Results", lines=15)
                
        submit_btn.click(
            fn=predict_fn,
            inputs=[audio_input, image_input],
            outputs=output
        )
        
        gr.Markdown("""
        ## Tips for best results
        - For audio: Tap the watermelon with your knuckle and record the sound
        - For image: Take a clear photo of the whole watermelon in good lighting
        
        ## About Brix Measurement
        Brix (°Bx) is a measurement of sugar content in a solution. For watermelons, higher Brix values indicate sweeter fruit.
        The average ripe watermelon has a Brix value between 9-11°.
        
        ## About the Mixture of Experts Model
        This app uses a Mixture of Experts (MoE) model that combines predictions from multiple neural networks.
        Our testing shows the ensemble approach achieves a Mean Absolute Error (MAE) of ~0.22, which is significantly 
        better than any individual model (best individual model: ~0.36 MAE).
        """)
    
    return interface

if __name__ == "__main__":
    import argparse
    
    parser = argparse.ArgumentParser(description="Watermelon Sugar Content Prediction App (MoE)")
    parser.add_argument(
        "--model_dir", 
        type=str, 
        default="models", 
        help="Directory containing the model checkpoints"
    )
    parser.add_argument(
        "--share", 
        action="store_true", 
        help="Create a shareable link for the app"
    )
    parser.add_argument(
        "--debug", 
        action="store_true", 
        help="Enable verbose debug output"
    )
    parser.add_argument(
        "--weighting", 
        type=str, 
        choices=["uniform", "performance"], 
        default="uniform", 
        help="How to weight the models (uniform or based on performance)"
    )
    
    args = parser.parse_args()
    
    if args.debug:
        print(f"\033[92mINFO\033[0m: Debug mode enabled")
    
    # Check if model directory exists
    if not os.path.exists(args.model_dir):
        print(f"\033[91mERR!\033[0m: Model directory not found at {args.model_dir}")
        sys.exit(1)
    
    # Determine weights based on argument
    weights = None
    if args.weighting == "performance":
        # Weights inversely proportional to the MAE (better models get higher weights)
        # These are the MAE values from the evaluation results
        mae_values = [0.3635, 0.3765, 0.3959]  # efficientnet_b3+transformer, efficientnet_b0+transformer, resnet50+transformer
        
        # Convert to weights (inverse of MAE, normalized)
        inverse_mae = [1/mae for mae in mae_values]
        total = sum(inverse_mae)
        weights = [val/total for val in inverse_mae]
        
        print(f"\033[92mINFO\033[0m: Using performance-based weights: {weights}")
    else:
        print(f"\033[92mINFO\033[0m: Using uniform weights")
    
    # Create and launch the app
    app = create_app(args.model_dir, weights)
    app.launch(share=args.share)