decoding

游雁

2024-06-13 f97f3e8dd5b40be6faaccd9c089aea52ceba29ef

 funasr/bin/train_ds.py

@@ -66,7 +66,6 @@
    # open tf32
    torch.backends.cuda.matmul.allow_tf32 = kwargs.get("enable_tf32", True)

    rank = int(os.environ.get("RANK", 0))
    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    world_size = int(os.environ.get("WORLD_SIZE", 1))

@@ -83,6 +82,8 @@
        logging.info(f"use_ddp: {use_ddp}, use_fsdp: {use_fsdp}")
        dist.init_process_group(backend=kwargs.get("backend", "nccl"), init_method="env://")
        torch.cuda.set_device(local_rank)

    rank = dist.get_rank()

    logging.info("Build model, frontend, tokenizer")
    device = kwargs.get("device", "cuda")
@@ -124,6 +125,7 @@
        use_ddp=use_ddp,
        use_fsdp=use_fsdp,
        device=kwargs["device"],
        excludes=kwargs.get("excludes", None),
        output_dir=kwargs.get("output_dir", "./exp"),
        **kwargs.get("train_conf"),
    )
@@ -158,6 +160,8 @@
        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
            )
@@ -178,6 +182,14 @@

            torch.cuda.empty_cache()

            time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
            logging.info(
                f"\n\nrank: {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} slices, {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours, "
                f"epoch: {trainer.max_epoch - epoch} epochs, {((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)
        scheduler.step()
@@ -189,7 +201,7 @@
        time2 = time.perf_counter()
        time_escaped = (time2 - time1) / 3600.0
        logging.info(
            f"rank: {local_rank}, "
            f"\n\nrank: {local_rank}, "
            f"time_escaped_epoch: {time_escaped:.3f} hours, "
            f"estimated to finish {trainer.max_epoch} "
            f"epoch: {(trainer.max_epoch - epoch) * time_escaped:.3f} hours\n"