update

嘉渊

2023-04-24 6427c834dfd97b1f05c6659cdc7ccf010bf82fe1

 funasr/bin/train.py

old mode 100644
new mode 100755
@@ -1,18 +1,33 @@
#!/usr/bin/env python3

import argparse
import logging
import os
import sys
from io import BytesIO

import torch

from funasr.build_utils.build_args import build_args
from funasr.build_utils.build_dataloader import build_dataloader
from funasr.build_utils.build_distributed import build_distributed
from funasr.build_utils.build_model import build_model
from funasr.build_utils.build_optimizer import build_optimizer
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.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
from funasr.utils import config_argparse
from funasr.utils.build_distributed import build_distributed
from funasr.utils.nested_dict_action import NestedDictAction
from funasr.utils.prepare_data import prepare_data
from funasr.utils.types import str2bool
from funasr.utils.types import str_or_none
from funasr.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump


def get_parser():
    parser = config_argparse.ArgumentParser(
    parser = argparse.ArgumentParser(
        description="FunASR Common Training Parser",
    )

@@ -25,6 +40,7 @@
        help="The number of gpus. 0 indicates CPU mode",
    )
    parser.add_argument("--seed", type=int, default=0, help="Random seed")
    parser.add_argument("--task_name", type=str, default="asr", help="Name for different tasks")

    # ddp related
    parser.add_argument(
@@ -61,6 +77,12 @@
        default=False,
        help="Whether to use the find_unused_parameters in "
             "torch.nn.parallel.DistributedDataParallel ",
    )
    parser.add_argument(
        "--gpu_id",
        type=int,
        default=0,
        help="local gpu id.",
    )

    # cudnn related
@@ -263,6 +285,98 @@
        action="append",
        default=[],
    )
    parser.add_argument(
        "--train_shape_file",
        type=str, action="append",
        default=[],
    )
    parser.add_argument(
        "--valid_shape_file",
        type=str,
        action="append",
        default=[],
    )
    parser.add_argument(
        "--use_preprocessor",
        type=str2bool,
        default=True,
        help="Apply preprocessing to data or not",
    )

    # optimization related
    parser.add_argument(
        "--optim",
        type=lambda x: x.lower(),
        default="adam",
        help="The optimizer type",
    )
    parser.add_argument(
        "--optim_conf",
        action=NestedDictAction,
        default=dict(),
        help="The keyword arguments for optimizer",
    )
    parser.add_argument(
        "--scheduler",
        type=lambda x: str_or_none(x.lower()),
        default=None,
        help="The lr scheduler type",
    )
    parser.add_argument(
        "--scheduler_conf",
        action=NestedDictAction,
        default=dict(),
        help="The keyword arguments for lr scheduler",
    )

    # most task related
    parser.add_argument(
        "--init",
        type=lambda x: str_or_none(x.lower()),
        default=None,
        help="The initialization method",
        choices=[
            "chainer",
            "xavier_uniform",
            "xavier_normal",
            "kaiming_uniform",
            "kaiming_normal",
            None,
        ],
    )
    parser.add_argument(
        "--token_list",
        type=str_or_none,
        default=None,
        help="A text mapping int-id to token",
    )
    parser.add_argument(
        "--token_type",
        type=str,
        default="bpe",
        choices=["bpe", "char", "word"],
        help="",
    )
    parser.add_argument(
        "--bpemodel",
        type=str_or_none,
        default=None,
        help="The model file fo sentencepiece",
    )
    parser.add_argument(
        "--cleaner",
        type=str_or_none,
        choices=[None, "tacotron", "jaconv", "vietnamese"],
        default=None,
        help="Apply text cleaning",
    )
    parser.add_argument(
        "--g2p",
        type=str_or_none,
        choices=g2p_choices,
        default=None,
        help="Specify g2p method if --token_type=phn",
    )

    # pai related
    parser.add_argument(
@@ -312,19 +426,28 @@
        help="oss bucket.",
    )

    # task related
    parser.add_argument("--task_name", help="for different task")

    return parser


if __name__ == '__main__':
    parser = get_parser()
    args = parser.parse_args()
    args, extra_task_params = parser.parse_known_args()
    if extra_task_params:
        args = build_args(args, parser, extra_task_params)
        # args = argparse.Namespace(**vars(args), **vars(task_args))

    # set random seed
    set_all_random_seed(args.seed)
    torch.backends.cudnn.enabled = args.cudnn_enabled
    torch.backends.cudnn.benchmark = args.cudnn_benchmark
    torch.backends.cudnn.deterministic = args.cudnn_deterministic

    # ddp init
    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
    args.distributed = args.dist_world_size > 1
    distributed_option = build_distributed(args)

    # for logging
    if not distributed_option.distributed or distributed_option.dist_rank == 0:
        logging.basicConfig(
            level="INFO",
@@ -337,14 +460,46 @@
            format=f"[{os.uname()[1].split('.')[0]}]"
                   f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
        )
    logging.info("world size: {}, rank: {}, local_rank: {}".format(distributed_option.dist_world_size,
                                                                   distributed_option.dist_rank,
                                                                   distributed_option.local_rank))

    # prepare files for dataloader
    prepare_data(args, distributed_option)

    set_all_random_seed(args.seed)
    torch.backends.cudnn.enabled = args.cudnn_enabled
    torch.backends.cudnn.benchmark = args.cudnn_benchmark
    torch.backends.cudnn.deterministic = args.cudnn_deterministic
    model = build_model(args)
    optimizers = build_optimizer(args, model=model)
    schedulers = build_scheduler(args, optimizers)

    logging.info("world size: {}, rank: {}, local_rank: {}".format(distributed_option.dist_world_size,
                                                                   distributed_option.dist_rank,
                                                                   distributed_option.local_rank))
    logging.info(pytorch_cudnn_version())
    logging.info(model_summary(model))
    logging.info("Optimizer: {}".format(optimizers))
    logging.info("Scheduler: {}".format(schedulers))

    # dump args to config.yaml
    if not distributed_option.distributed or distributed_option.dist_rank == 0:
        os.makedirs(args.output_dir, exist_ok=True)
        with open(os.path.join(args.output_dir, "config.yaml"), "w") as f:
            logging.info("Saving the configuration in {}/{}".format(args.output_dir, "config.yaml"))
            if args.use_pai:
                buffer = BytesIO()
                torch.save({"config": vars(args)}, buffer)
                args.oss_bucket.put_object(os.path.join(args.output_dir, "config.dict"), buffer.getvalue())
            else:
                yaml_no_alias_safe_dump(vars(args), f, indent=4, sort_keys=False)

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

    # Trainer, including model, optimizers, etc.
    trainer = build_trainer(
        args=args,
        model=model,
        optimizers=optimizers,
        schedulers=schedulers,
        train_dataloader=train_dataloader,
        valid_dataloader=valid_dataloader,
        distributed_option=distributed_option
    )

    trainer.run()