model/network.py

import torch
import torch.nn as nn
import math

class PositionalEncoding(nn.Module):
    """
    Injects information about the relative or absolute position of the tokens
    in the sequence. The model needs this because it has no recurrence.
    """
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        
        # Register buffer allows us to save this with state_dict but not train it
        self.register_buffer('pe', pe.unsqueeze(0))

    def forward(self, x):
        # x shape: [batch_size, seq_len, d_model]
        x = x + self.pe[:, :x.size(1)]
        return self.dropout(x)

class MiniTTS(nn.Module):
    def __init__(self, num_chars, num_mels, d_model=256, nhead=4, num_layers=4):
        super(MiniTTS, self).__init__()
        
        # 1. Text Encoder Layers
        self.embedding = nn.Embedding(num_chars, d_model)
        self.pos_encoder = PositionalEncoding(d_model)
        
        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, batch_first=True)
        self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        
        # 2. Spectrogram Decoder Layers
        # We process the mel spectrogram frames (Standard Transformers use teacher forcing during training)
        self.mel_embedding = nn.Linear(num_mels, d_model) # Project mel dimension to model dimension
        self.pos_decoder = PositionalEncoding(d_model)
        
        decoder_layer = nn.TransformerDecoderLayer(d_model=d_model, nhead=nhead, batch_first=True)
        self.transformer_decoder = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
        
        # 3. Final Projection
        # Project back from model dimension to Mel Spectrogram dimension (usually 80 channels)
        self.output_layer = nn.Linear(d_model, num_mels)
        
        # 4. Post-Net (Optional but recommended for TTS quality)
        # Simple convolutional network to refine the output
        self.post_net = nn.Sequential(
            nn.Conv1d(num_mels, 512, kernel_size=5, padding=2),
            nn.BatchNorm1d(512),
            nn.Tanh(),
            nn.Dropout(0.5),
            nn.Conv1d(512, num_mels, kernel_size=5, padding=2)
        )

    def forward(self, text_tokens, mel_target=None):
        """
        text_tokens: [batch, text_len] (Integers representing phonemes)
        mel_target: [batch, mel_len, num_mels] (The target spectrogram for training)
        """
        # --- ENCODING ---
        # [batch, text_len] -> [batch, text_len, d_model]
        src = self.embedding(text_tokens)
        src = self.pos_encoder(src)
        
        # Memory is the output of the encoder that the decoder attends to
        memory = self.transformer_encoder(src)
        
        # --- DECODING ---
        if mel_target is not None:
            # TRAINING MODE (Teacher Forcing)
            # We feed the real spectrogram (shifted) into the decoder
            tgt = self.mel_embedding(mel_target)
            tgt = self.pos_decoder(tgt)
            
            # Create a casual mask (prevent decoder from peeking at future frames)
            batch_size, tgt_len, _ = tgt.shape
            tgt_mask = self.generate_square_subsequent_mask(tgt_len).to(tgt.device)
            
            output = self.transformer_decoder(tgt, memory, tgt_mask=tgt_mask)
            output_mel = self.output_layer(output)
            
            # Post-net refinement
            # Conv1d expects [batch, channels, time], so we transpose
            output_mel_post = output_mel.transpose(1, 2)
            output_mel_post = self.post_net(output_mel_post)
            output_mel_post = output_mel_post.transpose(1, 2)
            
            # Combine raw output + residual
            final_output = output_mel + output_mel_post
            
            return final_output
        else:
            # INFERENCE MODE (Greedy Decoding)
            # We will handle this loop inside inference.py later
            # For now, we just return the encoder memory so we can debug shapes
            return memory

    def generate_square_subsequent_mask(self, sz):
        """Generates an upper-triangular matrix of -inf, with zeros on diag."""
        mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
        mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
        return mask

# --- SANITY CHECK ---
# Run this file directly to check if dimensions work!
if __name__ == "__main__":
    print("Testing Model Dimensions...")
    
    # Dummy Config
    num_chars = 50   # Size of vocabulary (phonemes)
    num_mels = 80    # Standard Mel Spectrogram channels
    batch_size = 2
    text_len = 10
    mel_len = 100
    
    # Instantiate Model
    model = MiniTTS(num_chars, num_mels)
    
    # Create Dummy Data
    dummy_text = torch.randint(0, num_chars, (batch_size, text_len))
    dummy_mel = torch.randn(batch_size, mel_len, num_mels)
    
    # Forward Pass
    try:
        output = model(dummy_text, dummy_mel)
        print(f"Input Text Shape: {dummy_text.shape}")
        print(f"Input Mel Shape:  {dummy_mel.shape}")
        print(f"Output Shape:     {output.shape}")
        print("\nSUCCESS: The architecture is valid!")
    except Exception as e:
        print(f"\nERROR: {e}")