-
Notifications
You must be signed in to change notification settings - Fork 21
/
Copy pathtrain.py
191 lines (158 loc) · 7.77 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
import warnings
warnings.filterwarnings('ignore')
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
import numpy as np
from tqdm import tqdm
import copy
import torch
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import params
from model.tts import Comospeech
from data import TextMelDataset, TextMelBatchCollate
from utils import plot_tensor, save_plot,load_teacher_model
from text.symbols import symbols
train_filelist_path = params.train_filelist_path
valid_filelist_path = params.test_filelist_path
cmudict_path = params.cmudict_path
add_blank = params.add_blank
log_dir = params.log_dir
n_epochs = params.n_epochs
batch_size = params.batch_size
out_size = params.out_size
learning_rate = params.learning_rate
random_seed = params.seed
nsymbols = len(symbols) + 1 if add_blank else len(symbols)
n_enc_channels = params.n_enc_channels
filter_channels = params.filter_channels
filter_channels_dp = params.filter_channels_dp
n_enc_layers = params.n_enc_layers
enc_kernel = params.enc_kernel
enc_dropout = params.enc_dropout
n_heads = params.n_heads
window_size = params.window_size
n_feats = params.n_feats
n_fft = params.n_fft
sample_rate = params.sample_rate
hop_length = params.hop_length
win_length = params.win_length
f_min = params.f_min
f_max = params.f_max
teacher = params.teacher
if __name__ == "__main__":
torch.manual_seed(random_seed)
np.random.seed(random_seed)
print('Initializing logger...')
logger = SummaryWriter(log_dir=log_dir)
print('Initializing data loaders...')
train_dataset = TextMelDataset(train_filelist_path, cmudict_path, add_blank,
n_fft, n_feats, sample_rate, hop_length,
win_length, f_min, f_max)
batch_collate = TextMelBatchCollate()
loader = DataLoader(dataset=train_dataset, batch_size=batch_size,
collate_fn=batch_collate, drop_last=True,pin_memory=True,prefetch_factor=8,
num_workers=16, shuffle=True,persistent_workers=True)
test_dataset = TextMelDataset(valid_filelist_path, cmudict_path, add_blank,
n_fft, n_feats, sample_rate, hop_length,
win_length, f_min, f_max)
print('Initializing model...')
if teacher:
print('comospeech_teacher')
else:
print('comospeech')
if teacher:
model = Comospeech(nsymbols, 1, None, n_enc_channels, filter_channels, filter_channels_dp,
n_heads, n_enc_layers, enc_kernel, enc_dropout, window_size,
n_feats ).cuda()
optimizer = torch.optim.Adam(params=model.parameters(), lr=learning_rate)
else:
model = Comospeech(nsymbols, 1, None, n_enc_channels, filter_channels, filter_channels_dp,
n_heads, n_enc_layers, enc_kernel, enc_dropout, window_size,
n_feats,teacher=False).cuda()
model = load_teacher_model(model,checkpoint_dir='') # teacher model path
optimizer = torch.optim.Adam(params=model.decoder.denoise_fn.parameters(), lr=learning_rate)
print('Logging test batch...')
test_batch = test_dataset.sample_test_batch(size=params.test_size)
for i, item in enumerate(test_batch):
mel = item['y']
logger.add_image(f'image_{i}/ground_truth', plot_tensor(mel.squeeze()),
global_step=0, dataformats='HWC')
save_plot(mel.squeeze(), f'{log_dir}/original_{i}.png')
print('Start training...')
iteration = 0
model.train()
for epoch in range(0, n_epochs + 1):
dur_losses = []
prior_losses = []
diff_losses = []
with tqdm(loader, total=len(train_dataset)//batch_size) as progress_bar:
for batch_idx, batch in enumerate(progress_bar):
model.zero_grad()
x, x_lengths = batch['x'].cuda(), batch['x_lengths'].cuda()
y, y_lengths = batch['y'].cuda(), batch['y_lengths'].cuda()
dur_loss, prior_loss, diff_loss = model.compute_loss(x, x_lengths,
y, y_lengths,
out_size=out_size)
if teacher:
loss = sum([dur_loss, prior_loss, diff_loss])
else:
loss = diff_loss
loss.backward()
enc_grad_norm = torch.nn.utils.clip_grad_norm_(model.encoder.parameters(),
max_norm=1)
dec_grad_norm = torch.nn.utils.clip_grad_norm_(model.decoder.parameters(),
max_norm=1)
optimizer.step()
logger.add_scalar('training/duration_loss', dur_loss.item(),
global_step=iteration)
logger.add_scalar('training/prior_loss', prior_loss.item(),
global_step=iteration)
logger.add_scalar('training/diffusion_loss', diff_loss.item(),
global_step=iteration)
logger.add_scalar('training/encoder_grad_norm', enc_grad_norm,
global_step=iteration)
logger.add_scalar('training/decoder_grad_norm', dec_grad_norm,
global_step=iteration)
dur_losses.append(dur_loss.item())
prior_losses.append(prior_loss.item())
diff_losses.append(diff_loss.item())
if batch_idx % 5 == 0:
msg = f'Epoch: {epoch}, iteration: {iteration} | dur_loss: {dur_loss.item()}, prior_loss: {prior_loss.item()}, diff_loss: {diff_loss.item() }'
progress_bar.set_description(msg)
iteration += 1
log_msg = 'Epoch %d: duration loss = %.3f ' % (epoch, np.mean(dur_losses))
log_msg += '| prior loss = %.3f ' % np.mean(prior_losses)
log_msg += '| diffusion loss = %.3f\n' % np.mean(diff_losses)
with open(f'{log_dir}/train.log', 'a') as f:
f.write(log_msg)
if epoch % params.save_every > 0:
continue
model.eval()
print('Synthesis...')
with torch.no_grad():
for i, item in enumerate(test_batch):
x = item['x'].to(torch.long).unsqueeze(0).cuda()
x_lengths = torch.LongTensor([x.shape[-1]]).cuda()
if teacher:
y_enc, y_dec, attn = model(x, x_lengths, n_timesteps=20)
else:
y_enc, y_dec, attn = model(x, x_lengths, n_timesteps=1)
logger.add_image(f'image_{i}/generated_enc',
plot_tensor(y_enc.squeeze().cpu()),
global_step=iteration, dataformats='HWC')
logger.add_image(f'image_{i}/generated_dec',
plot_tensor(y_dec.squeeze().cpu()),
global_step=iteration, dataformats='HWC')
logger.add_image(f'image_{i}/alignment',
plot_tensor(attn.squeeze().cpu()),
global_step=iteration, dataformats='HWC')
save_plot(y_enc.squeeze().cpu(),
f'{log_dir}/generated_enc_{i}.png')
save_plot(y_dec.squeeze().cpu(),
f'{log_dir}/generated_dec_{i}.png')
save_plot(attn.squeeze().cpu(),
f'{log_dir}/alignment_{i}.png')
ckpt = model.state_dict()
torch.save(ckpt, f=f"{log_dir}/model_{epoch}.pt")
model.train()