deepspeed

游雁

2024-05-17 f68ae892be3e59c20a033f18d5f61db7f633801f

 funasr/bin/train_ds.py

@@ -133,29 +133,7 @@
    kwargs["device"] = next(model.parameters()).device
    trainer.device = kwargs["device"]

    # optim
    logging.info("Build optim")
    optim = kwargs.get("optim", "adam")
    assert optim in optim_classes
    optim_class = optim_classes.get(optim)
    optim = optim_class(model.parameters(), **kwargs.get("optim_conf"))

    # scheduler
    logging.info("Build scheduler")
    scheduler = kwargs.get("scheduler", "warmuplr")
    assert scheduler in scheduler_classes
    scheduler_class = scheduler_classes.get(scheduler)
    scheduler = scheduler_class(optim, **kwargs.get("scheduler_conf"))

    if use_deepspeed:
        args = OmegaConf.create({"deepspeed_config": kwargs.get("deepspeed_config", "")})
        model, optimizer, _, scheduler = deepspeed.initialize(
            args=args,
            model=model,
            optimizer=optim,
            lr_scheduler=scheduler,
            model_parameters=model.parameters(),
        )
    model, optim, scheduler = trainer.warp_optim_scheduler(model, kwargs)

    # dataset
    logging.info("Build dataloader")
@@ -175,15 +153,6 @@
        scaler=scaler,
    )

    tensorboard_dir = os.path.join(kwargs.get("output_dir"), "tensorboard")
    os.makedirs(tensorboard_dir, exist_ok=True)
    try:
        from tensorboardX import SummaryWriter

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

    dataloader_tr, dataloader_val = None, None
    for epoch in range(trainer.start_epoch, trainer.max_epoch):
        time1 = time.perf_counter()
@@ -201,7 +170,6 @@
                dataloader_train=dataloader_tr,
                dataloader_val=dataloader_val,
                epoch=epoch,
                writer=writer,
                data_split_i=data_split_i,
                data_split_num=dataloader.data_split_num,
                start_step=trainer.start_step,
@@ -211,9 +179,7 @@
            torch.cuda.empty_cache()

        trainer.start_data_split_i = 0
        trainer.validate_epoch(
            model=model, dataloader_val=dataloader_val, epoch=epoch + 1, writer=writer
        )
        trainer.validate_epoch(model=model, dataloader_val=dataloader_val, epoch=epoch + 1)
        scheduler.step()
        trainer.step_in_epoch = 0
        trainer.save_checkpoint(