-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathtrain.py
192 lines (172 loc) · 7.23 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
import os
import sys
import time
import shutil
import pickle
import argparse
import traceback
import torch
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from generic_utils import (remove_experiment_folder,
create_experiment_folder, save_checkpoint,
save_best_model, load_config, lr_decay,
count_parameters, check_update, get_commit_hash)
from model import FFTNetModel
from model import MaskedCrossEntropyLoss
from model import EMA
from dataset import LJSpeechDataset
def train(epoch):
avg_loss = 0.0
epoch_time = 0
# progbar = Progbar(len(train_loader.dataset) // c.batch_size)
num_iter_epoch = len(train_loader.dataset) // c.batch_size
if c.ema_decay > 0:
ema = EMA(c.ema_decay)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param)
else:
ema = None
model.train()
for num_iter, batch in enumerate(train_loader):
start_time = time.time()
wav = batch[0].unsqueeze(1)
mel = batch[1].transpose(1, 2)
lens = batch[2]
target = batch[3]
if use_cuda:
wav = wav.cuda()
mel = mel.cuda()
target = target.cuda()
current_step = num_iter + epoch * len(train_loader) + 1
lr = lr_decay(c.lr, current_step, c.warmup_steps)
for params_group in optimizer.param_groups:
params_group['lr'] = lr
optimizer.zero_grad()
out = torch.nn.parallel.data_parallel(model, (wav, mel))
# out = model(wav, mel)
loss, fp, tp = criterion(out, target, lens)
loss.backward()
grad_norm, skip_flag = check_update(model, c.grad_clip, c.grad_top)
if skip_flag:
optimizer.zero_grad()
print(" | > Iteration skipped!!")
continue
optimizer.step()
# model ema
if ema is not None:
for name, param in model.named_parameters():
if name in ema.shadow:
ema.update(name, param.data)
step_time = time.time() - start_time
epoch_time += step_time
if current_step % c.print_iter == 0:
print(" | > step:{}/{}\tgloba_step:{}\tloss:{:.4f}\tgrad_norm:{:.4f}\t\
fp:{}\ttp:{}\tlr:{:.5f}\t".format(num_iter, num_iter_epoch,
current_step,
loss.item(), grad_norm, fp,
tp, params_group['lr']))
avg_loss += loss.item()
avg_loss /= num_iter
return ema, avg_loss
def evaluate(epoch, ema):
avg_loss = 0.0
epoch_time = 0
# progbar = Progbar(len(val_loader.dataset) // c.eval_batch_size)
ema_model = FFTNetModel(hid_channels=256, out_channels=256, n_layers=c.num_quant,
cond_channels=80)
ema_model = ema.assign_ema_model(model, ema_model, use_cuda)
ema_model.eval()
with torch.no_grad():
for num_iter, batch in enumerate(train_loader):
start_time = time.time()
wav = batch[0].unsqueeze(1)
mel = batch[1].transpose(1, 2)
lens = batch[2]
target = batch[3]
if use_cuda:
wav = wav.cuda()
mel = mel.cuda()
target = target.cuda()
current_step = num_iter + epoch * len(train_loader) + 1
out = ema_model(wav, mel)
loss, fp, tp = criterion(out, target, lens)
step_time = time.time() - start_time
epoch_time += step_time
avg_loss += loss.item()
avg_loss /= num_iter
return avg_loss
def main(args):
for epoch in range(c.epochs):
print(" > Epoch:{}/{}".format(epoch, c.epochs))
ema, avg_loss = train(epoch)
avg_val_loss = evaluate(epoch, ema)
print(" -- loss:{:.5f}\tval_loss:{:.5f}".format(avg_loss, avg_val_loss))
if __name__ == "__main__":
torch.manual_seed(1)
use_cuda = torch.cuda.is_available()
if use_cuda:
torch.backends.cudnn.benchmark = False
parser = argparse.ArgumentParser()
parser.add_argument('--config_path', type=str,
help='path to config file for training',)
parser.add_argument('--debug', type=bool, default=False,
help='Stop asking for git hash before the run.')
parser.add_argument('--finetune_path', type=str)
args = parser.parse_args()
c = load_config(args.config_path)
# setup output paths and read configs
_ = os.path.dirname(os.path.realpath(__file__))
OUT_PATH = os.path.join(_, c.output_path)
OUT_PATH = create_experiment_folder(OUT_PATH, c.model_name, True)
CHECKPOINT_PATH = os.path.join(OUT_PATH, 'checkpoints')
shutil.copyfile(args.config_path, os.path.join(OUT_PATH, 'config.json'))
# setup tensorboard
tb = SummaryWriter(OUT_PATH)
model = FFTNetModel(hid_channels=256, out_channels=256, n_layers=c.num_quant,
cond_channels=80)
criterion = MaskedCrossEntropyLoss()
# criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=c.lr)
num_params = count_parameters(model)
print(" > Models has {} parameters".format(num_params))
if use_cuda:
model.cuda()
criterion.cuda()
train_dataset = LJSpeechDataset(os.path.join(c.data_path, "mels",
"meta_fftnet_train.csv"),
os.path.join(c.data_path, "mels"),
c.sample_rate,
c.num_mels, c.num_freq,
c.min_level_db, c.frame_shift_ms,
c.frame_length_ms, c.preemphasis, c.ref_level_db,
c.num_quant, c.min_wav_len, c.max_wav_len, False)
val_dataset = LJSpeechDataset(os.path.join(c.data_path, "mels",
"meta_fftnet_val.csv"),
os.path.join(c.data_path, "mels"),
c.sample_rate,
c.num_mels, c.num_freq,
c.min_level_db, c.frame_shift_ms,
c.frame_length_ms, c.preemphasis,
c.ref_level_db, c.num_quant, c.min_wav_len,
c.max_wav_len, False)
train_loader = DataLoader(train_dataset, batch_size=c.batch_size,
shuffle=False, collate_fn=train_dataset.collate_fn,
drop_last=True, num_workers=c.num_loader_workers)
val_loader = DataLoader(val_dataset, batch_size=c.eval_batch_size,
shuffle=False, collate_fn=train_dataset.collate_fn,
drop_last=True, num_workers=4)
try:
main(args)
remove_experiment_folder(OUT_PATH)
except KeyboardInterrupt:
remove_experiment_folder(OUT_PATH)
try:
sys.exit(0)
except SystemExit:
os._exit(0)
except Exception:
remove_experiment_folder(OUT_PATH)
traceback.print_exc()
sys.exit(1)