scripts/train_rvc_v2_fixed.py

#!/usr/bin/env python3
"""
RVC v2 训练脚本 (Fixed) - 使用 torchaudio 替代 librosa
NumberBlocks One 音色克隆

修复内容:
- librosa.load → torchaudio.load (避免 numba 兼容问题)
- librosa.feature.melspectrogram → torchaudio.transforms.MelSpectrogram
- librosa.piptrack → torch-based pitch estimation
- 支持 soundfile / sox_backend 双后端
"""

import os
import sys
import yaml
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import torchaudio
import torchaudio.transforms as T
import numpy as np
from pathlib import Path
import json
import logging
import traceback

# 配置日志
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)

# 检查 torchaudio backend
logger.info(f"torchaudio version: {torchaudio.__version__}")
logger.info(f"torchaudio backends: {torchaudio.list_audio_backends()}")

class VoiceDataset(Dataset):
    """语音数据集 - 使用 torchaudio 加载"""

    def __init__(self, audio_dir, config, max_samples=None):
        self.audio_dir = Path(audio_dir)
        self.config = config
        self.sample_rate = config['data']['sample_rate']
        self.target_duration = config['data']['duration']
        self.target_samples = int(self.sample_rate * self.target_duration)

        # mel 频谱转换器
        n_mels = config['model'].get('spec_n_mels', 128)
        fmin = config['model'].get('spec_fmin', 0)
        fmax = config['model'].get('spec_fmax', self.sample_rate // 2)
        self.mel_transform = T.MelSpectrogram(
            sample_rate=self.sample_rate,
            n_mels=n_mels,
            f_min=fmin,
            f_max=fmax,
            n_fft=1024,
            hop_length=256,
        )
        self.amp_to_db = T.AmplitudeToDB(stype="power", top_db=80)

        # 获取音频文件
        extensions = ["*.wav", "*.mp3", "*.m4a", "*.flac", "*.ogg"]
        audio_files = []
        for ext in extensions:
            audio_files.extend(self.audio_dir.glob(ext))

        if max_samples:
            audio_files = audio_files[:max_samples]

        self.audio_files = sorted(audio_files)
        logger.info(f"加载了 {len(self.audio_files)} 个音频文件")

    def __len__(self):
        return len(self.audio_files)

    def _load_audio(self, audio_file):
        """使用 torchaudio 加载音频，带 fallback"""
        # 尝试 soundfile backend
        try:
            waveform, sr = torchaudio.load(str(audio_file), backend="soundfile")
        except Exception:
            pass

        # 尝试默认 backend
        try:
            waveform, sr = torchaudio.load(str(audio_file))
        except Exception as e:
            # 最后尝试 ffmpeg 后端
            try:
                waveform, sr = torchaudio.load(str(audio_file), backend="ffmpeg")
            except Exception:
                logger.error(f"无法加载 {audio_file}: {e}")
                return None, sr

        return waveform, sr

    def _load_audio_robust(self, audio_file):
        """鲁棒的音频加载：torchaudio → ffmpeg subprocess → zeros"""
        # Method 1: torchaudio 直接加载
        try:
            waveform, sr = torchaudio.load(str(audio_file))
            if waveform.numel() > 0:
                return waveform, sr
        except Exception:
            pass

        # Method 2: ffmpeg subprocess 转 WAV 到临时文件再加载
        try:
            import tempfile
            import subprocess as sp
            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
                tmp_path = tmp.name
            sp.run(
                ["ffmpeg", "-y", "-i", str(audio_file), "-ar", str(self.sample_rate),
                 "-ac", "1", "-f", "wav", tmp_path],
                capture_output=True, timeout=30
            )
            waveform, sr = torchaudio.load(tmp_path)
            os.unlink(tmp_path)
            if waveform.numel() > 0:
                return waveform, sr
        except Exception:
            pass

        # Method 3: 返回静音
        logger.warning(f"所有加载方式失败: {audio_file.name}，返回静音")
        return torch.zeros(1, self.target_samples), self.sample_rate

    def __getitem__(self, idx):
        audio_file = self.audio_files[idx]

        try:
            waveform, sr = self._load_audio_robust(audio_file)

            # 单声道
            if waveform.dim() > 1 and waveform.shape[0] > 1:
                waveform = waveform.mean(dim=0, keepdim=True)
            elif waveform.dim() == 1:
                waveform = waveform.unsqueeze(0)

            # 重采样
            if sr != self.sample_rate:
                resampler = T.Resample(orig_freq=sr, new_freq=self.sample_rate)
                waveform = resampler(waveform)

            # 裁剪或填充到目标长度
            if waveform.shape[1] > self.target_samples:
                start = torch.randint(0, waveform.shape[1] - self.target_samples, (1,)).item()
                waveform = waveform[:, start:start + self.target_samples]
            elif waveform.shape[1] < self.target_samples:
                padding = self.target_samples - waveform.shape[1]
                waveform = torch.nn.functional.pad(waveform, (0, padding))

            # 提取 mel 频谱
            mel_spec = self.mel_transform(waveform)
            mel_spec = self.amp_to_db(mel_spec)

            # 简单 pitch 特征 (用 energy 作为 proxy)
            frame_length = 256
            hop_length = 256
            energy = waveform.unfold(1, frame_length, hop_length).pow(2).mean(dim=2)
            pitch_feat = energy.squeeze(0)

            return {
                'audio': waveform.squeeze(0),
                'mel': mel_spec.squeeze(0),
                'pitch': pitch_feat,
                'filename': audio_file.name
            }

        except Exception as e:
            logger.error(f"处理 {audio_file.name} 失败: {e}")
            traceback.print_exc()
            return {
                'audio': torch.zeros(self.target_samples),
                'mel': torch.zeros(self.config['model'].get('spec_n_mels', 128), 100),
                'pitch': torch.zeros(100),
                'filename': audio_file.name
            }


class SimplifiedRVC(nn.Module):
    """简化版RVC模型"""

    def __init__(self, config):
        super().__init__()
        self.config = config

        # 特征提取器
        self.feature_extractor = nn.Sequential(
            nn.Conv1d(1, 64, kernel_size=7, stride=2, padding=3),
            nn.ReLU(),
            nn.Conv1d(64, 128, kernel_size=7, stride=2, padding=3),
            nn.ReLU(),
            nn.Conv1d(128, 256, kernel_size=7, stride=2, padding=3),
            nn.ReLU()
        )

        # 编码器
        self.encoder = nn.Sequential(
            nn.Conv1d(256, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv1d(128, 64, kernel_size=3, padding=1),
            nn.ReLU()
        )

        # 解码器
        self.decoder = nn.Sequential(
            nn.Conv1d(64, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv1d(128, 256, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.ConvTranspose1d(256, 1, kernel_size=7, stride=8, padding=3, output_padding=1)
        )

    def forward(self, x):
        # x: (batch, time)
        x = x.unsqueeze(1)  # (batch, 1, time)

        # 特征提取
        features = self.feature_extractor(x)

        # 编码
        encoded = self.encoder(features)

        # 解码
        decoded = self.decoder(encoded)

        # 输出 - 裁剪到和输入一致
        output = decoded.squeeze(1)
        if output.shape[1] > x.shape[1]:
            output = output[:, :x.shape[1]]
        elif output.shape[1] < x.shape[1]:
            output = torch.nn.functional.pad(output, (0, x.shape[1] - output.shape[1]))

        return output


def train_model(config):
    """训练模型"""
    logger.info("=" * 60)
    logger.info("🎤 开始RVC v2训练 (torchaudio版)")
    logger.info("=" * 60)

    # 设备
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    logger.info(f"📊 使用设备: {device}")

    # 创建数据集
    train_dir = config['data']['train_dir']
    logger.info(f"📂 加载数据集: {train_dir}")

    # 先测试能否加载至少一个音频
    test_dir = Path(train_dir)
    test_files = list(test_dir.glob("*.wav")) + list(test_dir.glob("*.mp3"))
    if test_files:
        logger.info(f"🔍 测试音频加载: {test_files[0].name}")
        try:
            wav, sr = torchaudio.load(str(test_files[0]))
            logger.info(f"  ✅ 成功! shape={wav.shape}, sr={sr}")
        except Exception as e:
            logger.warning(f"  ⚠️ torchaudio 直接加载失败: {e}")
            logger.info("  尝试 ffmpeg fallback...")
            import subprocess as sp
            import tempfile
            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
                tmp_path = tmp.name
            sp.run(
                ["ffmpeg", "-y", "-i", str(test_files[0]), "-ar", "40000",
                 "-ac", "1", "-f", "wav", tmp_path],
                capture_output=True, timeout=30
            )
            wav, sr = torchaudio.load(tmp_path)
            os.unlink(tmp_path)
            logger.info(f"  ✅ ffmpeg fallback 成功! shape={wav.shape}, sr={sr}")

    full_dataset = VoiceDataset(train_dir, config)

    if len(full_dataset) == 0:
        logger.error("❌ 没有找到任何音频文件！请检查数据目录。")
        return None, float('inf')

    # 分割训练集和验证集
    val_split = config['data'].get('val_split', 0.1)
    val_size = int(len(full_dataset) * val_split)
    train_size = len(full_dataset) - val_size

    train_dataset, val_dataset = torch.utils.data.random_split(
        full_dataset,
        [train_size, max(val_size, 1)]
    )

    logger.info(f"  训练集: {len(train_dataset)} 个样本")
    logger.info(f"  验证集: {len(val_dataset)} 个样本")

    # 创建数据加载器
    train_loader = DataLoader(
        train_dataset,
        batch_size=config['training']['batch_size'],
        shuffle=True,
        num_workers=0,
        drop_last=True
    )

    val_loader = DataLoader(
        val_dataset,
        batch_size=config['training']['batch_size'],
        shuffle=False,
        num_workers=0
    )

    # 创建模型
    logger.info(f"🏗️ 创建模型: {config['model']['name']}")
    model = SimplifiedRVC(config).to(device)

    total_params = sum(p.numel() for p in model.parameters())
    logger.info(f"  参数量: {total_params:,}")

    # 损失函数
    criterion = nn.MSELoss()

    # 优化器
    optimizer = optim.AdamW(
        model.parameters(),
        lr=config['training']['learning_rate'],
        weight_decay=config['training'].get('weight_decay', 1e-5)
    )

    # 学习率调度器
    scheduler = optim.lr_scheduler.StepLR(
        optimizer,
        step_size=config['training'].get('step_size', 100),
        gamma=config['training'].get('gamma', 0.5)
    )

    # 创建输出目录
    save_dir = Path(config['output']['save_dir'])
    save_dir.mkdir(parents=True, exist_ok=True)

    # 训练循环
    epochs = config['training']['epochs']
    best_val_loss = float('inf')

    logger.info(f"🚀 开始训练: {epochs} 个epoch")
    logger.info("=" * 60)

    for epoch in range(epochs):
        # 训练阶段
        model.train()
        train_loss = 0.0
        num_batches = 0

        for batch_idx, batch in enumerate(train_loader):
            audio = batch['audio'].to(device)

            # 前向传播
            optimizer.zero_grad()
            output = model(audio)

            # 确保输出和目标长度一致
            min_len = min(output.shape[1], audio.shape[1])
            loss = criterion(output[:, :min_len], audio[:, :min_len])

            # 反向传播
            loss.backward()
            optimizer.step()

            train_loss += loss.item()
            num_batches += 1

            if (batch_idx + 1) % 10 == 0:
                logger.info(f"Epoch {epoch+1}/{epochs} Batch {batch_idx+1}/{len(train_loader)} loss={loss.item():.6f}")

        train_loss /= max(num_batches, 1)

        # 验证阶段
        val_every = config['training'].get('val_every_n_epochs', 10)
        if (epoch + 1) % val_every == 0:
            model.eval()
            val_loss = 0.0
            val_batches = 0

            with torch.no_grad():
                for batch in val_loader:
                    audio = batch['audio'].to(device)
                    output = model(audio)
                    min_len = min(output.shape[1], audio.shape[1])
                    loss = criterion(output[:, :min_len], audio[:, :min_len])
                    val_loss += loss.item()
                    val_batches += 1

            val_loss /= max(val_batches, 1)

            logger.info(f"Epoch {epoch+1}/{epochs}: Train Loss = {train_loss:.6f}, Val Loss = {val_loss:.6f}")

            # 保存最佳模型
            if val_loss < best_val_loss:
                best_val_loss = val_loss
                save_path = save_dir / "best_model.pth"
                torch.save({
                    'epoch': epoch,
                    'model_state_dict': model.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                    'val_loss': val_loss,
                    'config': config,
                    'model_class': 'SimplifiedRVC',
                    'torchaudio_version': torchaudio.__version__,
                }, save_path)
                logger.info(f"  ✅ 保存最佳模型: {save_path} (Val Loss = {val_loss:.6f})")
        else:
            logger.info(f"Epoch {epoch+1}/{epochs}: Train Loss = {train_loss:.6f}")

        # 更新学习率
        scheduler.step()

    # 保存最终模型
    final_path = save_dir / "final_model.pth"
    torch.save({
        'epoch': epochs,
        'model_state_dict': model.state_dict(),
        'optimizer_state_dict': optimizer.state_dict(),
        'train_loss': train_loss,
        'config': config,
        'model_class': 'SimplifiedRVC',
        'torchaudio_version': torchaudio.__version__,
    }, final_path)

    logger.info("=" * 60)
    logger.info("✅ 训练完成！")
    logger.info(f"📊 最佳验证损失: {best_val_loss:.6f}")
    logger.info(f"📦 最终模型: {final_path}")
    logger.info("=" * 60)

    return model, best_val_loss


def main():
    """主函数"""
    # 加载配置
    config_file = "config_rvc_v2.yaml"
    if not Path(config_file).exists():
        logger.error(f"配置文件不存在: {config_file}")
        sys.exit(1)

    with open(config_file, 'r', encoding='utf-8') as f:
        config = yaml.safe_load(f)

    logger.info(f"📋 加载配置: {config_file}")

    # 训练模型
    model, best_val_loss = train_model(config)

    if model is not None:
        logger.info("🎉 训练成功完成！")
    else:
        logger.error("❌ 训练失败")
        sys.exit(1)


if __name__ == "__main__":
    main()