forked from NVIDIA/tacotron2
-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
245 lines (199 loc) · 9.41 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
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
import os
import time
import argparse
import math
from numpy import finfo
import torch
from torch.utils.data import DataLoader
from model.tacotron import Tacotron2, Tacotron2Loss
from utils.dataset import TextMelDataset, TextMelCollate
from utils.logger import Tacotron2Logger
from utils.utils import data_parallel_workaround
from hparams import create_hparams
def prepare_dataloaders(hparams):
# Get data, data loaders and collate function ready
trainset = TextMelDataset(hparams.training_files, hparams)
valset = TextMelDataset(hparams.validation_files, hparams)
collate_fn = TextMelCollate(hparams.n_frames_per_step)
train_loader = DataLoader(trainset, num_workers=1, shuffle=True,
sampler=None,
batch_size=hparams.batch_size, pin_memory=False,
drop_last=True, collate_fn=collate_fn)
return train_loader, valset, collate_fn
def prepare_directories_and_logger(output_directory, log_directory, rank):
if rank == 0:
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o775)
logger = Tacotron2Logger(os.path.join(output_directory, log_directory))
else:
logger = None
return logger
def load_model(hparams):
model = Tacotron2(hparams).cuda()
if hparams.fp16_run:
model.decoder.attention_layer.score_mask_value = finfo('float16').min
return model
def warm_start_model(checkpoint_path, model, ignore_layers):
assert os.path.isfile(checkpoint_path)
print("Warm starting model from checkpoint '{}'".format(checkpoint_path))
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
model_dict = checkpoint_dict['state_dict']
if len(ignore_layers) > 0:
model_dict = {k: v for k, v in model_dict.items()
if k not in ignore_layers}
dummy_dict = model.state_dict()
dummy_dict.update(model_dict)
model_dict = dummy_dict
model.load_state_dict(model_dict)
return model
def load_checkpoint(checkpoint_path, model, optimizer):
assert os.path.isfile(checkpoint_path)
print("Loading checkpoint '{}'".format(checkpoint_path))
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint_dict['state_dict'])
optimizer.load_state_dict(checkpoint_dict['optimizer'])
learning_rate = checkpoint_dict['learning_rate']
iteration = checkpoint_dict['iteration']
print("Loaded checkpoint '{}' from iteration {}" .format(
checkpoint_path, iteration))
return model, optimizer, learning_rate, iteration
def save_checkpoint(model, optimizer, learning_rate, iteration, filepath):
print("Saving model and optimizer state at iteration {} to {}".format(
iteration, filepath))
torch.save({'iteration': iteration,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'learning_rate': learning_rate}, filepath)
def validate(model, criterion, valset, iteration, batch_size, n_gpus,
collate_fn, logger, distributed_run, rank):
"""Handles all the validation scoring and printing"""
model.eval()
with torch.no_grad():
val_loader = DataLoader(valset, sampler=None, num_workers=1,
shuffle=False, batch_size=batch_size,
pin_memory=False, collate_fn=collate_fn)
val_loss = 0.0
for i, batch in enumerate(val_loader):
x, y = model.parse_batch(batch)
y_pred = model(x)
loss = criterion(y_pred, y)
val_loss += loss.item()
val_loss = val_loss / (i + 1)
model.train()
if rank == 0:
print("Validation loss {}: {:9f} ".format(iteration, val_loss))
logger.log_validation(val_loss, model, y, y_pred, iteration)
def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
rank, group_name, hparams):
"""Training and validation logging results to tensorboard and stdout
Params
------
output_directory (string): directory to save checkpoints
log_directory (string) directory to save tensorboard logs
checkpoint_path(string): checkpoint path
n_gpus (int): number of gpus
rank (int): rank of current gpu
hparams (object): comma separated list of "name=value" pairs.
"""
torch.manual_seed(hparams.seed)
torch.cuda.manual_seed(hparams.seed)
model = load_model(hparams)
learning_rate = hparams.learning_rate
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate,
weight_decay=hparams.weight_decay)
if hparams.fp16_run:
from apex import amp
model, optimizer = amp.initialize(
model, optimizer, opt_level='O2')
criterion = Tacotron2Loss()
logger = prepare_directories_and_logger(
output_directory, log_directory, rank)
train_loader, valset, collate_fn = prepare_dataloaders(hparams)
# Load checkpoint if one exists
iteration = 0
epoch_offset = 0
if checkpoint_path is not None:
if warm_start:
model = warm_start_model(
checkpoint_path, model, hparams.ignore_layers)
else:
model, optimizer, _learning_rate, iteration = load_checkpoint(
checkpoint_path, model, optimizer)
if hparams.use_saved_learning_rate:
learning_rate = _learning_rate
iteration += 1 # next iteration is iteration + 1
epoch_offset = max(0, int(iteration / len(train_loader)))
model.train()
is_overflow = False
# ================ MAIN TRAINNIG LOOP! ===================
for epoch in range(epoch_offset, hparams.epochs):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
start = time.perf_counter()
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
model.zero_grad()
x, y = model.parse_batch(batch)
if hparams.distributed_run and torch.cuda.device_count() > 1:
y_pred = data_parallel_workaround(model, x)
else:
y_pred = model(x)
loss = criterion(y_pred, y)
reduced_loss = loss.item()
if hparams.fp16_run:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if hparams.fp16_run:
grad_norm = torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), hparams.grad_clip_thresh)
is_overflow = math.isnan(grad_norm)
else:
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), hparams.grad_clip_thresh)
optimizer.step()
if not is_overflow and rank == 0:
duration = time.perf_counter() - start
print("Train loss {} {:.6f} Grad Norm {:.6f} {:.2f}s/it".format(
iteration, reduced_loss, grad_norm, duration))
logger.log_training(
reduced_loss, grad_norm, learning_rate, duration, iteration)
if not is_overflow and (iteration % hparams.iters_per_checkpoint == 0):
validate(model, criterion, valset, iteration,
hparams.batch_size, n_gpus, collate_fn, logger,
hparams.distributed_run, rank)
if rank == 0:
checkpoint_path = os.path.join(
output_directory, "checkpoint_{}".format(iteration))
save_checkpoint(model, optimizer, learning_rate, iteration,
checkpoint_path)
iteration += 1
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-o', '--output_directory', type=str, default='out',
help='directory to save checkpoints')
parser.add_argument('-l', '--log_directory', type=str, default='log',
help='directory to save tensorboard logs')
parser.add_argument('-c', '--checkpoint_path', type=str, default=None,
required=False, help='checkpoint path')
parser.add_argument('--warm_start', action='store_true',
help='load model weights only, ignore specified layers')
parser.add_argument('--n_gpus', type=int, default=1,
required=False, help='number of gpus')
parser.add_argument('--rank', type=int, default=0,
required=False, help='rank of current gpu')
parser.add_argument('--group_name', type=str, default='group_name',
required=False, help='Distributed group name')
args = parser.parse_args()
hparams = create_hparams()
torch.backends.cudnn.enabled = hparams.cudnn_enabled
torch.backends.cudnn.benchmark = hparams.cudnn_benchmark
print("FP16 Run:", hparams.fp16_run)
print("Dynamic Loss Scaling:", hparams.dynamic_loss_scaling)
print("Distributed Run:", hparams.distributed_run)
print("cuDNN Enabled:", hparams.cudnn_enabled)
print("cuDNN Benchmark:", hparams.cudnn_benchmark)
train(args.output_directory, args.log_directory, args.checkpoint_path,
args.warm_start, args.n_gpus, args.rank, args.group_name, hparams)