src/trainers/meta_trainer.py

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from typing import Dict, Tuple
import os
from tqdm import tqdm
import wandb

from ..models.encoder import SpeakerEncoder
from ..configs.config import Config, TrainingConfig

class MetaTrainer:
    """元学习训练器：实现少样本语音克隆的训练过程"""
    
    def __init__(
        self,
        model: SpeakerEncoder,
        config: Config,
        use_wandb: bool = True
    ):
        self.model = model
        self.config = config
        self.use_wandb = use_wandb
        
        self.device = torch.device(config.training.device)
        self.model = self.model.to(self.device)
        
        self.optimizer = optim.Adam(
            self.model.parameters(),
            lr=config.training.learning_rate
        )
        
        self.criterion = nn.CrossEntropyLoss()
        
        if use_wandb:
            wandb.init(project="voice-cloning", config=config)
            
    def compute_loss(
        self,
        support_data: Dict[str, torch.Tensor],
        query_data: Dict[str, torch.Tensor]
    ) -> Tuple[torch.Tensor, float]:
        """
        计算元学习损失
        
        Args:
            support_data:
                - mel_spec: [n_way*k_shot, n_mels, time]
                - speaker_ids: [n_way*k_shot]
            query_data:
                - mel_spec: [n_way*k_query, n_mels, time]
                - speaker_ids: [n_way*k_query]
                
        Returns:
            loss: 标量损失值
            acc: 准确率
        """
        # 获取支持集和查询集的嵌入向量
        support_mel = support_data['mel_spec'].to(self.device)  # [n_way*k_shot, n_mels, time]
        query_mel = query_data['mel_spec'].to(self.device)      # [n_way*k_query, n_mels, time]
        
        # 获取嵌入向量
        support_embeds = self.model(support_mel)  # [n_way*k_shot, embedding_dim]
        query_embeds = self.model(query_mel)      # [n_way*k_query, embedding_dim]
        
        # 计算支持集的质心
        centroids = []  # 将存储每个说话人的质心
        for speaker_idx in range(self.config.meta_learning.n_way):
            speaker_mask = (support_data['speaker_ids'] == speaker_idx).to(self.device)
            speaker_embeds = support_embeds[speaker_mask]  # [k_shot, embedding_dim]
            centroid = speaker_embeds.mean(dim=0)         # [embedding_dim]
            centroids.append(centroid)
        centroids = torch.stack(centroids)  # [n_way, embedding_dim]
        
        # 计算查询集样本与各个质心的相似度
        similarities = torch.matmul(query_embeds, centroids.T)  # [n_way*k_query, n_way]
        
        # 计算分类损失
        target = query_data['speaker_ids'].to(self.device)  # [n_way*k_query]
        loss = self.criterion(similarities, target)
        
        # 计算准确率
        pred = similarities.argmax(dim=1)  # [n_way*k_query]
        acc = (pred == target).float().mean().item()
        
        return loss, acc
        
    def train_epoch(self, dataloader: DataLoader) -> Tuple[float, float]:
        """训练一个epoch"""
        self.model.train()
        total_loss = 0
        total_acc = 0
        
        with tqdm(dataloader, desc="Training") as pbar:
            for batch_idx, (support_batch, query_batch) in enumerate(pbar):
                self.optimizer.zero_grad()
                
                loss, acc = self.compute_loss(support_batch, query_batch)
                loss.backward()
                
                # 梯度裁剪
                torch.nn.utils.clip_grad_norm_(self.model.parameters(), 3.0)
                
                self.optimizer.step()
                
                total_loss += loss.item()
                total_acc += acc
                
                pbar.set_postfix({
                    'loss': total_loss / (batch_idx + 1),
                    'acc': total_acc / (batch_idx + 1)
                })
                
                if self.use_wandb:
                    wandb.log({
                        'batch_loss': loss.item(),
                        'batch_acc': acc
                    })
                    
        avg_loss = total_loss / len(dataloader)
        avg_acc = total_acc / len(dataloader)
        
        return avg_loss, avg_acc
        
    def validate(self, dataloader: DataLoader) -> Tuple[float, float]:
        """验证模型"""
        self.model.eval()
        total_loss = 0
        total_acc = 0
        
        with torch.no_grad():
            for support_batch, query_batch in dataloader:
                loss, acc = self.compute_loss(support_batch, query_batch)
                total_loss += loss.item()
                total_acc += acc
                
        avg_loss = total_loss / len(dataloader)
        avg_acc = total_acc / len(dataloader)
        
        return avg_loss, avg_acc
        
    def save_checkpoint(self, epoch: int, loss: float, acc: float):
        """保存检查点"""
        checkpoint = {
            'epoch': epoch,
            'model_state_dict': self.model.state_dict(),
            'optimizer_state_dict': self.optimizer.state_dict(),
            'loss': loss,
            'acc': acc
        }
        
        checkpoint_path = os.path.join(
            self.config.training.checkpoint_dir,
            f'checkpoint_epoch_{epoch}.pt'
        )
        
        os.makedirs(self.config.training.checkpoint_dir, exist_ok=True)
        torch.save(checkpoint, checkpoint_path)
        
    def load_checkpoint(self, checkpoint_path: str):
        """加载检查点"""
        checkpoint = torch.load(checkpoint_path)
        self.model.load_state_dict(checkpoint['model_state_dict'])
        self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
        return checkpoint['epoch'], checkpoint['loss'], checkpoint['acc']