docs

游雁

2023-05-25 b18f7d121f2f17df8bf2d0c2bbb223bc5ddbcc0f

 funasr/bin/train.py

@@ -16,6 +16,7 @@
from funasr.build_utils.build_scheduler import build_scheduler
from funasr.build_utils.build_trainer import build_trainer
from funasr.text.phoneme_tokenizer import g2p_choices
from funasr.torch_utils.load_pretrained_model import load_pretrained_model
from funasr.torch_utils.model_summary import model_summary
from funasr.torch_utils.pytorch_version import pytorch_cudnn_version
from funasr.torch_utils.set_all_random_seed import set_all_random_seed
@@ -271,8 +272,8 @@
    parser.add_argument(
        "--init_param",
        type=str,
        action="append",
        default=[],
        nargs="*",
        help="Specify the file path used for initialization of parameters. "
             "The format is '<file_path>:<src_key>:<dst_key>:<exclude_keys>', "
             "where file_path is the model file path, "
@@ -333,7 +334,19 @@
        default="validation",
        help="dev dataset",
    )

    parser.add_argument(
        "--data_file_names",
        type=str,
        default="wav.scp,text",
        help="input data files",
    )
    parser.add_argument(
        "--speed_perturb",
        type=float,
        nargs="+",
        default=None,
        help="speed perturb",
    )
    parser.add_argument(
        "--use_preprocessor",
        type=str2bool,
@@ -506,6 +519,12 @@
        dtype=getattr(torch, args.train_dtype),
        device="cuda" if args.ngpu > 0 else "cpu",
    )
    for t in args.freeze_param:
        for k, p in model.named_parameters():
            if k.startswith(t + ".") or k == t:
                logging.info(f"Setting {k}.requires_grad = False")
                p.requires_grad = False

    optimizers = build_optimizer(args, model=model)
    schedulers = build_scheduler(args, optimizers)

@@ -530,6 +549,18 @@
            else:
                yaml_no_alias_safe_dump(vars(args), f, indent=4, sort_keys=False)

    for p in args.init_param:
        logging.info(f"Loading pretrained params from {p}")
        load_pretrained_model(
            model=model,
            init_param=p,
            ignore_init_mismatch=args.ignore_init_mismatch,
            map_location=f"cuda:{torch.cuda.current_device()}"
            if args.ngpu > 0
            else "cpu",
            oss_bucket=args.oss_bucket,
        )

    # dataloader for training/validation
    train_dataloader, valid_dataloader = build_dataloader(args)