Dev gzf exp (#1670)

zhifu gao

2024-04-28 93ef505e2d426b6aa1e58c0b4721999de789ff8e

 funasr/bin/train.py

@@ -13,7 +13,7 @@

from contextlib import nullcontext
import torch.distributed as dist
from collections.abc import Sequence

from omegaconf import DictConfig, OmegaConf
from torch.cuda.amp import autocast, GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
@@ -99,7 +99,7 @@
    if freeze_param is not None:
        if "," in freeze_param:
            freeze_param = eval(freeze_param)
        if not isinstance(freeze_param, Sequence):
        if not isinstance(freeze_param, (list, tuple)):
            freeze_param = (freeze_param,)
        logging.info("freeze_param is not None: %s", freeze_param)
        for t in freeze_param:
@@ -193,7 +193,7 @@
    try:
        from tensorboardX import SummaryWriter

        writer = SummaryWriter(tensorboard_dir) if trainer.rank == 0 else None
        writer = SummaryWriter(tensorboard_dir)  # if trainer.rank == 0 else None
    except:
        writer = None

@@ -206,6 +206,7 @@
                epoch, data_split_i=data_split_i, start_step=trainer.start_step
            )
            trainer.start_step = 0

            trainer.train_epoch(
                model=model,
                optim=optim,
@@ -225,7 +226,7 @@
            model=model, dataloader_val=dataloader_val, epoch=epoch, writer=writer
        )
        scheduler.step()

        trainer.step_cur_in_epoch = 0
        trainer.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler)

        time2 = time.perf_counter()