X-Codec-2.0/train.py

import os
 
import pytorch_lightning as pl
import hydra
import torch
import random
import time
from os.path import join, basename, exists
from pytorch_lightning import seed_everything
from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from pytorch_lightning.strategies import DDPStrategy,FSDPStrategy
from torch.utils.data import DataLoader
from data_module import DataModule
from lightning_module import CodecLightningModule
from pytorch_lightning.loggers import WandbLogger, TensorBoardLogger
from omegaconf import OmegaConf
 
seed = 1024
seed_everything(seed)
 
@hydra.main(config_path='config', config_name='default', version_base=None)
def train(cfg):
    checkpoint_callback = ModelCheckpoint(dirpath=cfg.log_dir, 
                            save_top_k=-1, save_last=True,
                            every_n_train_steps=6000, monitor='mel_loss', mode='min')

    lr_monitor = LearningRateMonitor(logging_interval='step')
    callbacks = [checkpoint_callback, lr_monitor]

    datamodule = DataModule(cfg)
    lightning_module = CodecLightningModule(cfg)
    log_dir_name = os.path.basename(os.path.normpath(cfg.log_dir))
    if cfg.get('use_wandb', True):
        logger = WandbLogger(
            project='wavvae_debug',
            name=log_dir_name,
            config=OmegaConf.to_container(cfg, resolve=True)
        )
    else:
        # 用 TensorBoard，日志保存在 cfg.log_dir/tb_logs/
        logger = TensorBoardLogger(save_dir=cfg.log_dir, name='tb_logs')

    ckpt_path = None
    last_ckpt = os.path.join(cfg.log_dir, 'last.ckpt')
    if os.path.exists(last_ckpt):
        ckpt_path = last_ckpt
        print(f"Resuming from checkpoint: {ckpt_path}")
    else:
        print("No checkpoint found, starting training from scratch.")

    trainer = pl.Trainer(
        **cfg.train.trainer,
        strategy=DDPStrategy(find_unused_parameters=True),
        callbacks=callbacks,
        logger=logger,
        profiler="simple",  # 启用 Profiler
        limit_train_batches=1.0 if not cfg.debug else 0.001
    )
    torch.backends.cudnn.benchmark = True  
    # lightning_module.strict_loading = False
    # LightningModule.strict_loading = True
    trainer.fit(lightning_module, datamodule=datamodule,ckpt_path=ckpt_path )
    print(f'Training ends, best score: {checkpoint_callback.best_model_score}, ckpt path: {checkpoint_callback.best_model_path}')

if __name__ == '__main__':
    train()