emotion_model.py

import tensorflow as tf
import torch
import numpy as np
import os

# Load the Keras model
keras_model = tf.keras.models.load_model('wav2vec_model.h5')

# Create a PyTorch model with the same architecture
class EmotionClassifier(torch.nn.Module):
    def __init__(self, input_shape, num_classes):
        super().__init__()
        # Adjust this architecture to match your Keras model
        self.flatten = torch.nn.Flatten()
        self.layers = torch.nn.Sequential(
            torch.nn.Linear(input_shape, 128),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.3),
            torch.nn.Linear(128, 64),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.3),
            torch.nn.Linear(64, num_classes)
        )
        
    def forward(self, x):
        x = self.flatten(x)
        return self.layers(x)

# Create PyTorch model
# Adjust these parameters based on your Keras model
input_shape = 13 * 128  # n_mfcc * max_length
num_classes = 7  # Number of emotions
pytorch_model = EmotionClassifier(input_shape, num_classes)

# Copy weights from Keras to PyTorch
# This would need to be adjusted based on your exact architecture
for i, layer in enumerate(keras_model.layers):
    if isinstance(layer, tf.keras.layers.Dense):
        # Get Keras weights and bias
        keras_weights = layer.get_weights()[0]
        keras_bias = layer.get_weights()[1]
        
        # Find the corresponding PyTorch layer
        # This is simplified; you'd need to match layers properly
        pytorch_layer = pytorch_model.layers[i * 2]
        
        # Copy weights and bias
        pytorch_layer.weight.data = torch.tensor(keras_weights.T, dtype=torch.float32)
        pytorch_layer.bias.data = torch.tensor(keras_bias, dtype=torch.float32)

# Save the PyTorch model
torch.save(pytorch_model.state_dict(), 'emotion_model.pt')