fix bug

游雁

2024-06-09 b75d1e89bb2f513a79bb07e9100ba1cd2bbcf40c

 funasr/bin/train_ds.py

@@ -130,8 +130,8 @@

    model = trainer.warp_model(model)

    kwargs["device"] = next(model.parameters()).device
    trainer.device = kwargs["device"]
    kwargs["device"] = int(os.environ.get("LOCAL_RANK", 0))
    trainer.device = int(os.environ.get("LOCAL_RANK", 0))

    model, optim, scheduler = trainer.warp_optim_scheduler(model, **kwargs)

@@ -158,6 +158,8 @@
        time1 = time.perf_counter()

        for data_split_i in range(trainer.start_data_split_i, dataloader.data_split_num):
            time_slice_i = time.perf_counter()

            dataloader_tr, dataloader_val = dataloader.build_iter(
                epoch, data_split_i=data_split_i, start_step=trainer.start_step
            )
@@ -177,6 +179,14 @@
            trainer.start_step = 0

            torch.cuda.empty_cache()

            time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
            logging.info(
                f"rank: {local_rank}, "
                f"time_escaped_epoch: {time_escaped:.3f} hours, "
                f"estimated to finish {dataloader.data_split_num} data_slices, remaining: {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours"
                f"epoch: {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
            )

        trainer.start_data_split_i = 0
        trainer.validate_epoch(model=model, dataloader_val=dataloader_val, epoch=epoch + 1)
@@ -198,7 +208,9 @@
        trainer.train_loss_avg = 0.0

    if trainer.rank == 0:
        average_checkpoints(trainer.output_dir, trainer.avg_nbest_model)
        average_checkpoints(
            trainer.output_dir, trainer.avg_nbest_model, use_deepspeed=trainer.use_deepspeed
        )

    trainer.close()