File size: 5,212 Bytes
be9690e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import print_function
import argparse
import datetime
import logging
logging.getLogger('matplotlib').setLevel(logging.WARNING)
from copy import deepcopy
import torch
import torch.distributed as dist
import deepspeed

from hyperpyyaml import load_hyperpyyaml

from torch.distributed.elastic.multiprocessing.errors import record

from cosyvoice.utils.executor import Executor
from cosyvoice.utils.train_utils import (
    init_distributed,
    init_dataset_and_dataloader,
    init_optimizer_and_scheduler,
    init_summarywriter, save_model,
    wrap_cuda_model, check_modify_and_save_config)


def get_args():
    parser = argparse.ArgumentParser(description='training your network')
    parser.add_argument('--train_engine',
                        default='torch_ddp',
                        choices=['torch_ddp', 'deepspeed'],
                        help='Engine for paralleled training')
    parser.add_argument('--model', required=True, help='model which will be trained')
    parser.add_argument('--config', required=True, help='config file')
    parser.add_argument('--train_data', required=True, help='train data file')
    parser.add_argument('--cv_data', required=True, help='cv data file')
    parser.add_argument('--checkpoint', help='checkpoint model')
    parser.add_argument('--model_dir', required=True, help='save model dir')
    parser.add_argument('--tensorboard_dir',
                        default='tensorboard',
                        help='tensorboard log dir')
    parser.add_argument('--ddp.dist_backend',
                        dest='dist_backend',
                        default='nccl',
                        choices=['nccl', 'gloo'],
                        help='distributed backend')
    parser.add_argument('--num_workers',
                        default=0,
                        type=int,
                        help='num of subprocess workers for reading')
    parser.add_argument('--prefetch',
                        default=100,
                        type=int,
                        help='prefetch number')
    parser.add_argument('--pin_memory',
                        action='store_true',
                        default=False,
                        help='Use pinned memory buffers used for reading')
    parser.add_argument('--deepspeed.save_states',
                        dest='save_states',
                        default='model_only',
                        choices=['model_only', 'model+optimizer'],
                        help='save model/optimizer states')
    parser.add_argument('--timeout',
                        default=30,
                        type=int,
                        help='timeout (in seconds) of cosyvoice_join.')
    parser = deepspeed.add_config_arguments(parser)
    args = parser.parse_args()
    return args


@record
def main():
    args = get_args()
    logging.basicConfig(level=logging.DEBUG,
                        format='%(asctime)s %(levelname)s %(message)s')

    override_dict = {k: None for k in ['llm', 'flow', 'hift'] if k != args.model}
    with open(args.config, 'r') as f:
        configs = load_hyperpyyaml(f, overrides=override_dict)
    configs['train_conf'].update(vars(args))

    # Init env for ddp
    init_distributed(args)

    # Get dataset & dataloader
    train_dataset, cv_dataset, train_data_loader, cv_data_loader = \
        init_dataset_and_dataloader(args, configs)

    # Do some sanity checks and save config to arsg.model_dir
    configs = check_modify_and_save_config(args, configs)

    # Tensorboard summary
    writer = init_summarywriter(args)

    # load checkpoint
    model = configs[args.model]
    if args.checkpoint is not None:
        model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'))

    # Dispatch model from cpu to gpu
    model = wrap_cuda_model(args, model)

    # Get optimizer & scheduler
    model, optimizer, scheduler = init_optimizer_and_scheduler(args, configs, model)

    # Save init checkpoints
    info_dict = deepcopy(configs['train_conf'])
    save_model(model, 'init', info_dict)

    # Get executor
    executor = Executor()

    # Start training loop
    for epoch in range(info_dict['max_epoch']):
        executor.epoch = epoch
        train_dataset.set_epoch(epoch)
        dist.barrier()
        group_join = dist.new_group(backend="gloo", timeout=datetime.timedelta(seconds=args.timeout))
        executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join)
        dist.destroy_process_group(group_join)

if __name__ == '__main__':
    main()