-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathtrain.py
198 lines (179 loc) · 9.36 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
import os
import sys
import torch
import argparse
from tqdm import tqdm
from model.neufa import NeuFA_base, NeuFA_TeP, NeuFA_MeP, NeuFA_TeMP
from data.common import Collate
from save import Save
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--name', default=None)
parser.add_argument('--load_model', default=None)
parser.add_argument('--train_path', default=os.path.expanduser('~/LibriSpeech'))
parser.add_argument('--dev_path', default=os.path.expanduser('~/BuckeyeTrain'))
parser.add_argument('--valid_path', default=os.path.expanduser('~/BuckeyeTest'))
parser.add_argument('--model', default='temp', choices=['base', 'tep', 'mep', 'temp'])
args = parser.parse_args()
device = "cuda:%d" % args.gpu
if args.model == 'base':
from hparams import base as hparams
model = NeuFA_base(hparams)
elif args.model == 'tep':
from hparams import base as hparams
model = NeuFA_TeP(hparams)
elif args.model == 'mep':
from hparams import base as hparams
model = NeuFA_MeP(hparams)
elif args.model == 'temp':
from hparams import temp as hparams
model = NeuFA_TeMP(hparams)
if hparams.strategy != 'finetune':
if 'LJSpeech' in args.train_path:
from data.ljspeech import LJSpeech
train_dataset = LJSpeech(args.train_path, reduction=hparams.reduction_rate)
elif 'LibriSpeech' in args.train_path:
from data.librispeech import LibriSpeech
train_dataset = LibriSpeech(args.train_path, reduction=hparams.reduction_rate)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=hparams.batch_size, shuffle=True, collate_fn=Collate(device), drop_last=True)
if hparams.strategy != 'pretrain':
from data.buckeye import Buckeye, BuckeyePhoneme
dev_dataset = Buckeye(args.dev_path, reduction=hparams.reduction_rate)
dev_dataloader = torch.utils.data.DataLoader(dev_dataset, batch_size=hparams.batch_size, shuffle=True, collate_fn=Collate(device), drop_last=True)
valid_dataset = Buckeye(args.valid_path, reduction=hparams.reduction_rate)
valid_dataloader = torch.utils.data.DataLoader(valid_dataset, batch_size=hparams.batch_size, shuffle=True, collate_fn=Collate(device), drop_last=True)
if hparams.strategy == 'semi2':
dev_dataset2 = BuckeyePhoneme(args.dev_path, reduction=hparams.reduction_rate)
dev_dataloader2 = torch.utils.data.DataLoader(dev_dataset, batch_size=hparams.batch_size, shuffle=True, collate_fn=Collate(device), drop_last=True)
valid_dataset2 = BuckeyePhoneme(args.valid_path, reduction=hparams.reduction_rate)
valid_dataloader2 = torch.utils.data.DataLoader(valid_dataset, batch_size=hparams.batch_size, shuffle=True, collate_fn=Collate(device), drop_last=True)
if args.load_model:
#model_dict = model.state_dict()
#state_dict = torch.load(args.load_model)
#state_dict = {k: v for k, v in state_dict.items() if not k.startswith('aligner.')}
#model_dict.update(state_dict)
#model.load_state_dict(model_dict)
model.load_state_dict(torch.load(args.load_model, map_location='cpu'))
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=hparams.learning_rate, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)
if args.name is None:
args.name = args.model
#else:
# args.name = args.model + '_' + args.name
save = Save(args.name)
save.save_parameters(hparams)
def process(model, stage, data, step, batch):
if isinstance(model, NeuFA_base):
predicted = model(*data[:2])
text_loss = model.text_loss(predicted[0], data[0])
save.writer.add_scalar(f'{stage}/text loss', text_loss, step)
speech_loss = model.mfcc_loss(predicted[1], data[1])
save.writer.add_scalar(f'{stage}/speech loss', speech_loss, step)
loss = hparams.text_loss * text_loss + hparams.speech_loss * speech_loss
if stage == 'training':
attention_loss = model.attention_loss(*predicted[2:4], hparams.attention_loss_alpha)
save.writer.add_scalar(f'{stage}/attention loss', attention_loss, step)
loss += hparams.attention_loss * attention_loss
else:
boundary_loss = model.boundary_loss(predicted[-1], data[2])
save.writer.add_scalar(f'{stage}/boundary loss', boundary_loss, step)
boundaries = model.extract_boundary(predicted[-1])
boundary_mae = model.boundary_mae(boundaries, data[2])
save.writer.add_scalar(f'{stage}/boundary mae', boundary_mae, step)
loss += hparams.boundary_loss * boundary_loss
if isinstance(model, NeuFA_TeP):
tep_loss = model.length_loss(*predicted[4:6])
tep_mse = model.length_loss(*predicted[4:6], normalize=False)
save.writer.add_scalar(f'{stage}/tep loss', tep_loss, step)
save.writer.add_scalar(f'{stage}/tep rmse', torch.sqrt(tep_mse), step)
loss += hparams.tep_loss * tep_loss
if isinstance(model, NeuFA_MeP):
if isinstance(model, NeuFA_TeMP):
mep_loss = model.length_loss(*predicted[6:8])
mep_mse = model.length_loss(*predicted[6:8], normalize=False)
else:
mep_loss = model.length_loss(*predicted[4:6])
mep_mse = model.length_loss(*predicted[4:6], normalize=False)
save.writer.add_scalar(f'{stage}/mep loss', mep_loss, step)
save.writer.add_scalar(f'{stage}/mep rmse', torch.sqrt(mep_mse), step)
loss += hparams.mep_loss * mep_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
save.save_log(stage, epoch, batch, step, loss)
if step % 100 == 0:
save.save_attention(stage, step, predicted[2][0], predicted[3][0])
if stage != 'training':
save.save_boundary(stage, step, boundaries[0], data[2][0], predicted[2][0].shape)
if step % 20000 == 0:
save.save_model(model, f'{step // 1000}k')
step = 1
for epoch in range(hparams.max_epochs):
save.logger.info('Epoch %d', epoch)
batch = 1
if hparams.strategy == 'pretrain':
for data in train_dataloader:
process(model, 'training', data, step, batch)
step += 1
batch += 1
continue
for data in dev_dataloader:
if hparams.strategy == 'finetune':
process(model, 'dev', data, step, batch)
if hparams.strategy == 'semi':
training_data = next(iter(train_dataloader))
process(model, 'training', training_data, step, batch)
process(model, 'dev', data, step, batch)
if hparams.strategy == 'semi2':
training_data = next(iter(train_dataloader))
dev_data2 = next(iter(dev_dataloader2))
process(model, 'training', training_data, step, batch)
process(model, 'dev', data, step, batch)
process(model, 'dev2', dev_data2, step, batch)
batch += 1
step += 1
with torch.no_grad():
predicted = []
all_data = []
for data in tqdm(valid_dataloader):
all_data.append(data)
predicted.append(model(*data[:2]))
data = [i for i in zip(*all_data)]
predicted = [i for i in zip(*predicted)]
for i in range(len(data)):
data[i] = [k for j in data[i] for k in j]
for i in range(len(predicted)):
predicted[i] = [k for j in predicted[i] for k in j]
if isinstance(model, NeuFA_base):
text_loss = model.text_loss(predicted[0], data[0])
save.writer.add_scalar('test/text loss', text_loss, epoch)
speech_loss = model.mfcc_loss(predicted[1], data[1])
save.writer.add_scalar('test/speech loss', speech_loss, epoch)
attention_loss = model.attention_loss(*predicted[2:4], 1)
save.writer.add_scalar('test/attention loss', attention_loss, epoch)
loss = hparams.text_loss * text_loss + hparams.speech_loss * speech_loss + hparams.attention_loss * attention_loss
boundary_loss = model.boundary_loss(predicted[-1], data[2])
save.writer.add_scalar('test/boundary loss', boundary_loss, epoch)
boundaries = model.extract_boundary(predicted[-1])
boundary_mae = model.boundary_mae(boundaries, data[2])
save.writer.add_scalar(f'test/boundary mae', boundary_mae, epoch)
loss += hparams.boundary_loss * boundary_loss
if isinstance(model, NeuFA_TeP):
tep_loss = model.length_loss(*predicted[4:6])
tep_mse = model.length_loss(*predicted[4:6], normalize=False)
save.writer.add_scalar('test/tep loss', tep_loss, epoch)
save.writer.add_scalar('test/tep rmse', torch.sqrt(tep_mse), epoch)
loss += hparams.tep_loss * tep_loss
if isinstance(model, NeuFA_MeP):
if isinstance(model, NeuFA_TeMP):
mep_loss = model.length_loss(*predicted[6:8])
mep_mse = model.length_loss(*predicted[6:8], normalize=False)
else:
mep_loss = model.length_loss(*predicted[4:6])
mep_mse = model.length_loss(*predicted[4:6], normalize=False)
save.writer.add_scalar('test/mep loss', mep_loss, epoch)
save.writer.add_scalar('test/mep rmse', torch.sqrt(mep_mse), epoch)
loss += hparams.mep_loss * mep_loss
save.save_log('test', epoch, batch, epoch, loss)
save.save_attention('test', epoch, predicted[2][0], predicted[3][0])
save.save_boundary('test', epoch, boundaries[0], data[2][0], predicted[2][0].shape)