/
waveglow.py
279 lines (234 loc) · 9.35 KB
/
waveglow.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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
# Copyright (c) 2019 NVIDIA Corporation
import argparse
import copy
from functools import partial
import os
from ruamel.yaml import YAML
import nemo
import nemo.utils.argparse as nm_argparse
import nemo_asr
import nemo_tts
from nemo_tts import (waveglow_log_to_tb_func,
waveglow_process_eval_batch,
waveglow_eval_log_to_tb_func)
def parse_args():
parser = argparse.ArgumentParser(
parents=[nm_argparse.NemoArgParser()],
description='Waveglow',
conflict_handler='resolve')
parser.set_defaults(
checkpoint_dir=None,
optimizer="adam",
batch_size=12,
eval_batch_size=12,
lr=0.0001,
amp_opt_level="O1",
create_tb_writer=True,
lr_policy=None,
weight_decay=1e-6
)
# Overwrite default args
parser.add_argument("--max_steps", type=int, default=None, required=False,
help="max number of steps to train")
parser.add_argument("--num_epochs", type=int, default=None, required=False,
help="number of epochs to train")
parser.add_argument("--model_config", type=str, required=True,
help="model configuration file: model.yaml")
# Create new args
parser.add_argument("--exp_name", default="Waveglow", type=str)
args = parser.parse_args()
if args.lr_policy:
raise NotImplementedError("Waveglow does not support lr policy arg")
if args.max_steps is not None and args.num_epochs is not None:
raise ValueError("Either max_steps or num_epochs should be provided.")
if args.eval_freq % 25 != 0:
raise ValueError("eval_freq should be a multiple of 25.")
exp_directory = [f"{args.exp_name}-lr_{args.lr}-bs_{args.batch_size}",
"",
(f"-wd_{args.weight_decay}-opt_{args.optimizer}"
f"-ips_{args.iter_per_step}")]
if args.max_steps:
exp_directory[1] = f"-s_{args.max_steps}"
elif args.num_epochs:
exp_directory[1] = f"-e_{args.num_epochs}"
else:
raise ValueError("Both max_steps and num_epochs were None.")
return args, "".join(exp_directory)
def create_NMs(waveglow_params, logger=None):
data_preprocessor = nemo_asr.AudioPreprocessing(
**waveglow_params["AudioPreprocessing"])
waveglow = nemo_tts.WaveGlowNM(**waveglow_params["WaveGlowNM"])
waveglow_loss = nemo_tts.WaveGlowLoss()
if logger:
logger.info('================================')
logger.info(f"Total number of parameters: {waveglow.num_weights}")
logger.info('================================')
return (data_preprocessor, waveglow, waveglow_loss)
def create_train_dag(neural_factory,
neural_modules,
waveglow_params,
train_dataset,
batch_size,
checkpoint_save_freq,
cpu_per_dl=1):
data_preprocessor, waveglow, waveglow_loss = neural_modules
train_dl_params = copy.deepcopy(waveglow_params["AudioDataLayer"])
train_dl_params.update(waveglow_params["AudioDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
data_layer = nemo_tts.AudioDataLayer(
manifest_filepath=train_dataset,
batch_size=batch_size,
num_workers=cpu_per_dl,
**train_dl_params,
)
N = len(data_layer)
steps_per_epoch = int(N / (batch_size * neural_factory.world_size))
neural_factory.logger.info('Have {0} examples to train on.'.format(N))
# Train DAG
audio, audio_len, = data_layer()
spec_target, spec_target_len = data_preprocessor(
input_signal=audio,
length=audio_len)
z, log_s_list, log_det_W_list = waveglow(
mel_spectrogram=spec_target, audio=audio)
loss_t = waveglow_loss(
z=z,
log_s_list=log_s_list,
log_det_W_list=log_det_W_list)
# Callbacks needed to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss_t, z, spec_target, spec_target_len],
print_func=lambda x: print(f"Loss: {x[0].data}"),
log_to_tb_func=partial(
waveglow_log_to_tb_func,
log_images=False),
tb_writer=neural_factory.tb_writer,
)
chpt_callback = nemo.core.CheckpointCallback(
folder=neural_factory.checkpoint_dir,
step_freq=checkpoint_save_freq)
callbacks = [train_callback, chpt_callback]
return loss_t, callbacks, steps_per_epoch
def create_eval_dags(neural_factory,
neural_modules,
waveglow_params,
eval_datasets,
eval_batch_size,
eval_freq,
cpu_per_dl=1):
data_preprocessor, waveglow, _ = neural_modules
eval_dl_params = copy.deepcopy(waveglow_params["AudioDataLayer"])
eval_dl_params.update(waveglow_params["AudioDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
callbacks = []
# assemble eval DAGs
for eval_dataset in eval_datasets:
data_layer_eval = nemo_tts.AudioDataLayer(
manifest_filepath=eval_dataset,
batch_size=eval_batch_size,
num_workers=cpu_per_dl,
**eval_dl_params,
)
audio, audio_len, = data_layer_eval()
spec_target, spec_target_len = data_preprocessor(
input_signal=audio,
length=audio_len)
audio_pred, log_s_list, log_det_W_list = waveglow(
mel_spectrogram=spec_target, audio=audio)
# create corresponding eval callback
tagname = os.path.basename(eval_dataset).split(".")[0]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[audio_pred, spec_target, spec_target_len],
user_iter_callback=waveglow_process_eval_batch,
user_epochs_done_callback=lambda x: x,
tb_writer_func=partial(
waveglow_eval_log_to_tb_func,
tag=tagname,
mel_fb=data_preprocessor.filter_banks),
eval_step=eval_freq,
tb_writer=neural_factory.tb_writer)
callbacks.append(eval_callback)
return callbacks
def create_all_dags(neural_factory,
neural_modules,
waveglow_params,
train_dataset,
batch_size,
checkpoint_save_freq,
eval_datasets=None,
eval_batch_size=None,
eval_freq=None):
# Calculate num_workers for dataloader
cpu_per_dl = max(int(os.cpu_count() / neural_factory.world_size), 1)
training_loss, training_callbacks, steps_per_epoch = create_train_dag(
neural_factory=neural_factory,
neural_modules=neural_modules,
waveglow_params=waveglow_params,
train_dataset=train_dataset,
batch_size=batch_size,
checkpoint_save_freq=checkpoint_save_freq,
cpu_per_dl=cpu_per_dl)
eval_callbacks = []
if eval_datasets:
eval_callbacks = create_eval_dags(
neural_factory=neural_factory,
neural_modules=neural_modules,
waveglow_params=waveglow_params,
eval_datasets=eval_datasets,
eval_batch_size=eval_batch_size,
eval_freq=eval_freq,
cpu_per_dl=cpu_per_dl)
else:
neural_factory.logger.info("There were no val datasets passed")
callbacks = training_callbacks + eval_callbacks
return training_loss, callbacks, steps_per_epoch
def main():
args, name = parse_args()
log_dir = None
if args.work_dir:
log_dir = os.path.join(args.work_dir, name)
# instantiate Neural Factory with supported backend
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=args.local_rank,
optimization_level=args.amp_opt_level,
log_dir=log_dir,
checkpoint_dir=args.checkpoint_dir,
create_tb_writer=args.create_tb_writer,
files_to_copy=[args.model_config, __file__],
cudnn_benchmark=args.cudnn_benchmark,
tensorboard_dir=args.tensorboard_dir)
if args.local_rank is not None:
neural_factory.logger.info('Doing ALL GPU')
yaml = YAML(typ="safe")
with open(args.model_config) as file:
waveglow_params = yaml.load(file)
# instantiate neural modules
neural_modules = create_NMs(waveglow_params, neural_factory.logger)
# build dags
train_loss, callbacks, steps_per_epoch = create_all_dags(
neural_factory=neural_factory,
neural_modules=neural_modules,
waveglow_params=waveglow_params,
train_dataset=args.train_dataset,
batch_size=args.batch_size,
checkpoint_save_freq=args.checkpoint_save_freq,
eval_datasets=args.eval_datasets,
eval_batch_size=args.eval_batch_size,
eval_freq=args.eval_freq)
# train model
neural_factory.train(
tensors_to_optimize=[train_loss],
callbacks=callbacks,
optimizer=args.optimizer,
optimization_params={
"num_epochs": args.num_epochs,
"max_steps": args.max_steps,
"lr": args.lr,
"weight_decay": args.weight_decay,
"grad_norm_clip": None},
batches_per_step=args.iter_per_step)
if __name__ == '__main__':
main()