forked from igormq/speech2text
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
199 lines (144 loc) · 7.35 KB
/
train.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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
""" Command to train a model
It requires a configuration file and a directory in which to write the results.
Heavily inspired on SeanNaren DeepSpeech2, AllenNLP, and NVIDIA/OpenSeq2Seq
"""
import argparse
import logging
import os
import re
from distutils.dir_util import copy_tree
import torch
import torchaudio
from asr import losses, models
from asr.common.params import Params
from asr.data import Alphabet
from asr.engine import Trainer
from asr.exceptions import ConfigurationError
from asr.utils.dataset import loaders_from_params
from asr.utils.exp_utils import create_serialization_dir, prepare_environment
from asr.utils.io_utils import prepare_global_logging
from asr.utils.model import freeze_params
import torch.distributed as dist
logger = logging.getLogger('asr')
CONFIG_NAME = 'config.yaml'
def train_model_from_args(args):
if args.local_rank == 0 and args.prev_output_dir is not None:
logger.info('Copying results from {} to {}...'.format(args.prev_output_dir, args.serialization_dir))
copy_tree(args.prev_output_dir, args.serialization_dir, update=True, verbose=True)
if not os.path.isfile(args.param_path):
raise ConfigurationError(f'Parameters file {args.param_path} not found.')
logger.info(f'Loading experiment from {args.param_path} with overrides `{args.overrides}`.')
params = Params.load(args.param_path, args.overrides)
prepare_environment(params)
logger.info(args.local_rank)
if args.local_rank == 0:
create_serialization_dir(params, args.serialization_dir, args.reset)
if args.distributed:
logger.info(f'World size: {dist.get_world_size()} | Rank {dist.get_rank()} | ' f'Local Rank {args.local_rank}')
dist.barrier()
prepare_global_logging(args.serialization_dir, local_rank=args.local_rank, verbosity=args.verbosity)
if args.local_rank == 0:
params.save(os.path.join(args.serialization_dir, CONFIG_NAME))
loaders = loaders_from_params(params,
distributed=args.distributed,
world_size=args.world_size,
first_epoch=args.first_epoch)
if os.path.exists(os.path.join(args.serialization_dir, "alphabet")):
alphabet = Alphabet.from_file(os.path.join(args.serialization_dir, "alphabet"))
else:
alphabet = Alphabet.from_params(params.pop("alphabet", {}))
alphabet.save_to_files(os.path.join(args.serialization_dir, "alphabet"))
loss = losses.from_params(params.pop('loss'))
model = models.from_params(alphabet=alphabet, params=params.pop('model'))
trainer_params = params.pop("trainer")
if args.fine_tune:
_, archive_weight_file = models.load_archive(args.fine_tune)
archive_weights = torch.load(archive_weight_file, map_location=lambda storage, loc: storage)['model']
# Avoiding initializing from archive some weights
no_ft_regex = trainer_params.pop("no_ft", ())
finetune_weights = {}
random_weights = []
for name, parameter in archive_weights.items():
if any(re.search(regex, name) for regex in no_ft_regex):
random_weights.append(name)
continue
finetune_weights[name] = parameter
logger.info(f'Loading the following weights from archive {args.fine_tune}:')
logger.info(','.join(finetune_weights.keys()))
logger.info(f'The following weights are at random:')
logger.info(','.join(random_weights))
model.load_state_dict(finetune_weights, strict=False)
# Freezing some parameters
freeze_params(model, trainer_params.pop('no_grad', ()))
trainer = Trainer(args.serialization_dir,
trainer_params,
model,
loss,
alphabet,
local_rank=args.local_rank,
world_size=args.world_size,
sync_bn=args.sync_bn,
opt_level=args.opt_level,
keep_batchnorm_fp32=args.keep_batchnorm_fp32,
loss_scale=args.loss_scale)
try:
trainer.run(loaders['train'], val_loader=loaders.get('val'), num_epochs=trainer_params['num_epochs'])
except KeyboardInterrupt:
# if we have completed an epoch, try to create a model archive.
if os.path.exists(os.path.join(args.serialization_dir, models.DEFAULT_WEIGHTS)):
logging.info("Training interrupted by the user. Attempting to create "
"a model archive using the current best epoch weights.")
raise
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('param_path', type=str, help='path to parameter file describing the model to be trained')
parser.add_argument('-s',
'--serialization-dir',
'--output-dir',
'--output_dir',
default=os.environ.get('PT_OUTPUT_DIR', os.path.join('results')),
type=str,
help='directory in which to save the model and its logs')
parser.add_argument('--prev-output-dir', '--prevModelDir', default=os.environ.get('PT_PREV_OUTPUT_DIR', None))
parser.add_argument('-r', '--reset', '--restart', action='store_true', default=False, help='Restart training')
parser.add_argument('--fine-tune', default=None, type=str, help='Path to an archive model to fine-tune from')
parser.add_argument('--first-epoch',
default='asc',
choices=['asc', 'desc', None],
help='First epoch data loader behavior')
parser.add_argument('-o',
'--overrides',
type=str,
default="{}",
help='a JSON structure used to override the experiment configuration')
# Deterministic
parser.add_argument('--deterministic', '-d', action='store_true', default=False)
# Distributed training
parser.add_argument('--backend', default='nccl', type=str)
parser.add_argument('--init-method', default='env://', type=str)
parser.add_argument('--local-rank', '--local_rank', '--gpu', default=0, type=int)
parser.add_argument('--sync-bn', action='store_true', default=False)
# F16 training
parser.add_argument('--opt-level', default='O0', type=str, choices=['O0', 'O1', 'O2', 'O3'])
parser.add_argument('--keep-batchnorm-fp32', default=None, action='store_true')
parser.add_argument('--loss-scale', type=str, default=None)
parser.add_argument('--verbosity', '-v', action='count', default=0)
args = parser.parse_args()
# Initialize sox
torchaudio.initialize_sox()
args.world_size = 1
# Pin GPU to be used to process local rank (one GPU per process)
torch.cuda.set_device(args.local_rank)
if args.deterministic:
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed:
torch.distributed.init_process_group(backend=args.backend, init_method=args.init_method)
args.world_size = torch.distributed.get_world_size()
else:
args.local_rank = 0
train_model_from_args(args)
torchaudio.shutdown_sox()