-
Notifications
You must be signed in to change notification settings - Fork 66
/
run.py
135 lines (108 loc) · 5.47 KB
/
run.py
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
# File : run.py
# Author : Zhengkun Tian
# Email : zhengkun.tian@outlook.com
import re
import os
import yaml
import random
import logging
import shutil
import numpy as np
import torch
import argparse
from otrans.model import End2EndModel, LanguageModel
from otrans.train.scheduler import BuildOptimizer, BuildScheduler
from otrans.train.trainer import Trainer
from otrans.utils import count_parameters
from otrans.data.loader import FeatureLoader
def main(args, params, expdir):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
model_type = params['model']['type']
if model_type[-2:] == 'lm':
model = LanguageModel[model_type](params['model'])
else:
model = End2EndModel[model_type](params['model'])
# Count total parameters
count_parameters(model.named_parameters())
if args.ngpu >= 1:
model.cuda()
logging.info(model)
optimizer = BuildOptimizer[params['train']['optimizer_type']](
filter(lambda p: p.requires_grad, model.parameters()), **params['train']['optimizer']
)
logger.info('[Optimizer] Build a %s optimizer!' % params['train']['optimizer_type'])
scheduler = BuildScheduler[params['train']['scheduler_type']](optimizer, **params['train']['scheduler'])
logger.info('[Scheduler] Build a %s scheduler!' % params['train']['scheduler_type'])
if args.continue_training and args.init_model:
chkpt = torch.load(args.init_model)
model.load_model(chkpt)
logger.info('[Continue Training] Load saved model %s' % args.init_model)
if args.continue_training and args.init_optim_state:
ochkpt = torch.load(args.init_optim_state)
optimizer.load_state_dict(ochkpt['optim'])
logger.info('[Continue Training] Load saved optimizer state dict!')
global_step = ochkpt['global_step'] if 'global_step' in ochkpt else args.from_step
scheduler.global_step = global_step
scheduler.set_lr()
logger.info('Set the global step to %d and init lr to %.6f' % (scheduler.global_step, scheduler.lr))
trainer = Trainer(params, model=model, optimizer=optimizer, scheduler=scheduler, expdir=expdir, ngpu=args.ngpu,
parallel_mode=args.parallel_mode, local_rank=args.local_rank, is_debug=args.debug,
keep_last_n_chkpt=args.keep_last_n_chkpt, from_epoch=args.from_epoch, mixspeech=args.mixspeech)
train_loader = FeatureLoader(params, 'train', ngpu=args.ngpu, mode=args.parallel_mode)
trainer.train(train_loader=train_loader)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, default='egs/aishell/conf/transformer_base.yaml')
parser.add_argument('-n', '--ngpu', type=int, default=1)
parser.add_argument('-g', '--gpus', type=str, default='0')
parser.add_argument('-se', '--seed', type=int, default=1234)
parser.add_argument('-p', '--parallel_mode', type=str, default='dp')
parser.add_argument('-r', '--local_rank', type=int, default=0)
parser.add_argument('-l', '--logging_level', type=str, default='info', choices=['info','debug'])
parser.add_argument('-lg', '--log_file', type=str, default=None)
parser.add_argument('-mp', '--mixed_precision', action='store_true', default=False)
parser.add_argument('-ct', '--continue_training', action='store_true', default=False)
parser.add_argument('-dir', '--expdir', type=str, default=None)
parser.add_argument('-im', '--init_model', type=str, default=None)
parser.add_argument('-ios', '--init_optim_state', type=str, default=None)
parser.add_argument('-debug', '--debug', action='store_true', default=False)
parser.add_argument('-knpt', '--keep_last_n_chkpt', type=int, default=30)
parser.add_argument('-tfs', '--from_step', type=int, default=0)
parser.add_argument('-tfe', '--from_epoch', type=int, default=0)
parser.add_argument('-vb', '--verbose', type=int, default=0)
parser.add_argument('-ol', '--opt_level', type=str, choices=['O0', 'O1', 'O2', 'O3'], default='O1')
parser.add_argument('-ms', '--mixspeech', type=int, default=0)
cmd_args = parser.parse_args()
with open(cmd_args.config, 'r') as f:
params = yaml.load(f, Loader=yaml.FullLoader)
if cmd_args.expdir is not None:
expdir = os.path.join(cmd_args.expdir, params['train']['save_name'])
else:
expdir = os.path.join('egs', params['data']['name'], 'exp', params['train']['save_name'])
if not os.path.exists(expdir):
os.makedirs(expdir)
shutil.copy(cmd_args.config, os.path.join(expdir, 'config.yaml'))
logging_level = {
'info': logging.INFO,
'debug': logging.DEBUG
}
if cmd_args.log_file is not None:
log_file = cmd_args.log_file
else:
log_file = cmd_args.config.split('/')[-1][:-5] + '.log'
LOG_FORMAT = "%(asctime)s - %(levelname)s - %(message)s"
logging.basicConfig(level=logging_level[cmd_args.logging_level], format=LOG_FORMAT)
logger = logging.getLogger(__name__)
if cmd_args.ngpu > 0:
os.environ['CUDA_VISIBLE_DEVICES'] = str(cmd_args.gpus)
logger.info('Set CUDA_VISIBLE_DEVICES as %s' % cmd_args.gpus)
if cmd_args.parallel_mode == 'ddp':
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355'
os.environ["OMP_NUM_THREADS"] = '1'
torch.cuda.set_device(cmd_args.local_rank)
main(cmd_args, params, expdir)