-
Notifications
You must be signed in to change notification settings - Fork 12
/
train.py
206 lines (170 loc) · 9.04 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
import os, argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
from modules.model import Model
from modules.loss import MDNLoss
import hparams
from text import *
from utils.utils import *
from utils.writer import get_writer
from torch.utils.tensorboard import SummaryWriter
import math
import matplotlib.pyplot as plt
from datetime import datetime
from glob import glob
def validate(model, criterion, val_loader, iteration, writer, stage):
model.eval()
with torch.no_grad():
n_data, val_loss = 0, 0
for i, batch in enumerate(val_loader):
n_data += len(batch[0])
if stage==0:
text_padded, mel_padded, text_lengths, mel_lengths = [
reorder_batch(x, hparams.n_gpus).cuda() for x in batch
]
else:
text_padded, mel_padded, align_padded, text_lengths, mel_lengths = [
reorder_batch(x, hparams.n_gpus).cuda() for x in batch
]
if stage !=3:
encoder_input = model.module.Prenet(text_padded)
hidden_states, _ = model.module.FFT_lower(encoder_input, text_lengths)
if stage==0:
mu_sigma = model.module.get_mu_sigma(hidden_states)
loss, log_prob_matrix = criterion(mu_sigma, mel_padded, text_lengths, mel_lengths)
elif stage==1:
mel_out = model.module.get_melspec(hidden_states, align_padded, mel_lengths)
mel_mask = ~get_mask_from_lengths(mel_lengths)
mel_padded_selected = mel_padded.masked_select(mel_mask.unsqueeze(1))
mel_out_selected = mel_out.masked_select(mel_mask.unsqueeze(1))
loss = nn.L1Loss()(mel_out_selected, mel_padded_selected)
elif stage==2:
mu_sigma = model.module.get_mu_sigma(hidden_states)
mdn_loss, log_prob_matrix = criterion(mu_sigma, mel_padded, text_lengths, mel_lengths)
align = model.module.viterbi(log_prob_matrix, text_lengths, mel_lengths)
mel_out = model.module.get_melspec(hidden_states, align, mel_lengths)
mel_mask = ~get_mask_from_lengths(mel_lengths)
mel_padded_selected = mel_padded.masked_select(mel_mask.unsqueeze(1))
mel_out_selected = mel_out.masked_select(mel_mask.unsqueeze(1))
fft_loss = nn.L1Loss()(mel_out_selected, mel_padded_selected)
loss = mdn_loss + fft_loss
elif stage==3:
duration_out = model.module.get_duration(text_padded, text_lengths) # gradient cut
duration_target = align_padded.sum(-1)
duration_mask = ~get_mask_from_lengths(text_lengths)
duration_out = duration_out.masked_select(duration_mask)
duration_target = duration_target.masked_select(duration_mask)
loss = nn.MSELoss()(torch.log(duration_out), torch.log(duration_target))
val_loss += loss.item() * len(batch[0])
val_loss /= n_data
if stage==0:
writer.add_scalar('Validation loss', val_loss, iteration//hparams.accumulation)
align = model.module.viterbi(log_prob_matrix[0:1], text_lengths[0:1], mel_lengths[0:1]) # 1, L, T
mel_out = torch.matmul(align[0].t(), mu_sigma[0, :, :hparams.n_mel_channels]).t() # F, T
writer.add_image('Validation_alignments', align.detach().cpu(), iteration//hparams.accumulation)
writer.add_specs(mel_padded[0].detach().cpu(),
mel_out.detach().cpu(),
iteration//hparams.accumulation, 'Validation')
elif stage==1:
writer.add_scalar('Validation loss', val_loss, iteration//hparams.accumulation)
writer.add_specs(mel_padded[0].detach().cpu(),
mel_out[0].detach().cpu(),
iteration//hparams.accumulation, 'Validation')
elif stage==2:
writer.add_scalar('Validation mdn_loss', mdn_loss.item(), iteration//hparams.accumulation)
writer.add_scalar('Validation fft_loss', fft_loss.item(), iteration//hparams.accumulation)
writer.add_image('Validation_alignments',
align[0:1, :text_lengths[0], :mel_lengths[0]].detach().cpu(),
iteration//hparams.accumulation)
writer.add_specs(mel_padded[0].detach().cpu(),
mel_out[0].detach().cpu(),
iteration//hparams.accumulation, 'Validation')
elif stage==3:
writer.add_scalar('Validation loss', val_loss, iteration//hparams.accumulation)
model.train()
def main(args):
train_loader, val_loader, collate_fn = prepare_dataloaders(hparams, stage=args.stage)
if args.stage!=0:
checkpoint_path = f"training_log/aligntts/stage{args.stage-1}/checkpoint_{hparams.train_steps[args.stage-1]}"
if not os.path.isfile(checkpoint_path):
print(f'{checkpoint_path} does not exist')
checkpoint_path = sorted(glob(f"training_log/aligntts/stage{args.stage-1}/checkpoint_*"))[-1]
print(f'Loading {checkpoint_path} instead')
state_dict = {}
for k, v in torch.load(checkpoint_path)['state_dict'].items():
state_dict[k[7:]]=v
model = Model(hparams).cuda()
model.load_state_dict(state_dict)
model = nn.DataParallel(model).cuda()
else:
model = nn.DataParallel(Model(hparams)).cuda()
criterion = MDNLoss()
writer = get_writer(hparams.output_directory, f'{hparams.log_directory}/stage{args.stage}')
optimizer = torch.optim.Adam(model.parameters(),
lr=hparams.lr,
betas=(0.9, 0.98),
eps=1e-09)
iteration, loss = 0, 0
model.train()
print(f'Stage{args.stage} Start!!! ({str(datetime.now())})')
while True:
for i, batch in enumerate(train_loader):
if args.stage==0:
text_padded, mel_padded, text_lengths, mel_lengths = [
reorder_batch(x, hparams.n_gpus).cuda() for x in batch
]
align_padded=None
else:
text_padded, mel_padded, align_padded, text_lengths, mel_lengths = [
reorder_batch(x, hparams.n_gpus).cuda() for x in batch
]
sub_loss = model(text_padded,
mel_padded,
align_padded,
text_lengths,
mel_lengths,
criterion,
stage=args.stage,
log_viterbi=args.log_viterbi,
cpu_viterbi=args.cpu_viterbi)
sub_loss = sub_loss.mean()/hparams.accumulation
sub_loss.backward()
loss = loss+sub_loss.item()
iteration += 1
print(f'[{str(datetime.now())}] Stage {args.stage} Iter {iteration:<6d} Loss {loss:<8.6f}')
if iteration%hparams.accumulation == 0:
lr_scheduling(optimizer, iteration//hparams.accumulation)
nn.utils.clip_grad_norm_(model.parameters(), hparams.grad_clip_thresh)
optimizer.step()
model.zero_grad()
writer.add_scalar('Train loss', loss, iteration//hparams.accumulation)
loss=0
if iteration%(hparams.iters_per_validation*hparams.accumulation)==0:
validate(model, criterion, val_loader, iteration, writer, args.stage)
if iteration%(hparams.iters_per_checkpoint*hparams.accumulation)==0:
save_checkpoint(model,
optimizer,
hparams.lr,
iteration//hparams.accumulation,
filepath=f'{hparams.output_directory}/{hparams.log_directory}/stage{args.stage}')
if iteration==(hparams.train_steps[args.stage]*hparams.accumulation):
break
if iteration==(hparams.train_steps[args.stage]*hparams.accumulation):
break
print(f'Stage{args.stage} End!!! ({str(datetime.now())})')
if __name__ == '__main__':
p = argparse.ArgumentParser()
p.add_argument('--gpu', type=str, default='0,1')
p.add_argument('-v', '--verbose', type=str, default='0')
p.add_argument('--stage', type=int, required=True)
p.add_argument('--log_viterbi', type=bool, default=False)
p.add_argument('--cpu_viterbi', type=bool, default=False)
args = p.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"]=args.gpu
torch.manual_seed(hparams.seed)
torch.cuda.manual_seed(hparams.seed)
if args.verbose=='0':
import warnings
warnings.filterwarnings("ignore")
main(args)