-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
275 lines (235 loc) · 10.5 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
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
import torch
import torch.nn as nn
from torch import optim
from network import *
from gen_training_loader import DataLoader, collate_fn, SpeechData
from pytorch_pretrained_bert import BertTokenizer, BertModel, BertForMaskedLM
from multiprocessing import cpu_count
import numpy as np
import argparse
import os
import time
def get_embedding_input(em_infos, length):
# print(length)
output = list()
for em_info in em_infos:
embeddings = em_info[0]
sep_list = em_info[1]
one_batch = list()
for j in range(len(sep_list)-1):
temp = torch.stack([embeddings[j]
for i in range(sep_list[j+1]-sep_list[j])])
one_batch.append(temp)
zero_pad = length - sep_list[len(sep_list)-1]
# print(zero_pad)
if zero_pad > 0:
# print(zero_pad)
zeros = torch.stack([torch.zeros(768) for i in range(zero_pad)])
# print(zeros.size())
one_batch.append(zeros)
cat_temp = torch.cat(one_batch, 0)
cat_temp = cat_temp[0:length, :]
if cat_temp.size(0) < length:
# print("###############")
cat_temp = torch.cat([cat_temp, torch.stack(
[torch.zeros(768) for i in range(length-cat_temp.size(0))])], 0)
# print(cat_temp.size())
# cat_temp = 0
# print(cat_temp.size())
cat_temp = cat_temp[0:length, :]
output.append(cat_temp)
output = torch.stack(output)
# print(output.size())
# output = output[:, 0:length, :]
# print(output.size())
return output
def main(args):
# Get device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Define model
model = nn.DataParallel(Tacotron()).to(device)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model_bert = BertModel.from_pretrained('bert-base-uncased')
print("Models Have Been Defined")
# Get dataset
dataset = SpeechData(args.dataset_path, tokenizer, model_bert)
# print(type(args.dataset_path))
# print(len(dataset))
# Optimizer
optimizer = optim.Adam(model.parameters(), lr=hp.lr)
# # Loss for frequency of human register
# n_priority_freq = int(3000 / (hp.sample_rate * 0.5) * hp.num_freq)
# Get training loader
print("Get Training Loader")
training_loader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True,
collate_fn=collate_fn, drop_last=True, num_workers=cpu_count())
# print(len(training_loader))
# Load checkpoint if exists
try:
checkpoint = torch.load(os.path.join(
hp.checkpoint_path, 'checkpoint_%d.pth.tar' % args.restore_step))
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("---Model Restored at Step %d---\n" % args.restore_step)
except:
print("---Start New Training---\n")
if not os.path.exists(hp.checkpoint_path):
os.mkdir(hp.checkpoint_path)
# Training
model = model.train()
total_step = hp.epochs * len(training_loader)
# print(total_step)
# Loss = []
Time = np.array([])
Start = time.clock()
for epoch in range(hp.epochs):
# print("########")
for i, data_batch in enumerate(training_loader):
start_time = time.clock()
# print("in")
# Count step
current_step = i + args.restore_step + \
epoch * len(training_loader) + 1
# print(current_step)
# Init
optimizer.zero_grad()
# {"text": texts, "mel": mels, "spec": specs}
texts = data_batch["text"]
em_infos = data_batch["em_infos"]
# print(len(em_info))
# print(texts)
mels = trans(data_batch["mel"])
# mels = trans(mels)
# print(np.shape(mels))
specs = trans(data_batch["spec"])
# print(np.shape(specs))
# mel_input = mels[:, :-1, :]
# print(np.shape(mel_input))
# mel_input = mel_input[:, :, -hp.num_mels:]
# print(np.shape(mel_input))
# print(np.shape(mels))
frame_arr = np.zeros(
[args.batch_size, hp.num_mels, 1], dtype=np.float32)
# print(np.shape(frame_arr))
# print(np.shape(mels[:, :, 1:]))
mel_input = np.concatenate((frame_arr, mels[:, :, 1:]), axis=2)
# print(np.shape(mel_input))
# print(mels)
if torch.cuda.is_available():
texts = torch.from_numpy(texts).type(
torch.cuda.LongTensor).to(device)
else:
texts = torch.from_numpy(texts).type(
torch.LongTensor).to(device)
# print(texts.size())
embeddings = get_embedding_input(em_infos, texts.size(1))
embeddings = embeddings.to(device)
mels = torch.from_numpy(mels).to(device)
specs = torch.from_numpy(specs).to(device)
mel_input = torch.from_numpy(mel_input).to(device)
# Forward
mel_output, linear_output = model.forward(
texts, mel_input, embeddings)
# print("#####################")
# print(np.shape(mel_output))
# print(np.shape(linear_output))
# print()
# print(np.shape(mels[:, :, 1:]))
# print(np.shape(np.transpose(mels.cpu().numpy())))
# print(np.shape(specs))
# print(np.shape(np.transpose(mels)))
# Calculate loss
# st = time.clock()
mel_loss = torch.abs(
mel_output - compare(mel_output, mels[:, :, 1:], device))
mel_loss = torch.mean(mel_loss)
linear_loss = torch.abs(
linear_output - compare(linear_output, specs, device))
linear_loss = torch.mean(linear_loss)
loss = mel_loss + hp.loss_weight * linear_loss
loss = loss.to(device)
# Loss.append(loss)
# et = time.clock()
# print(et - st)
# print(loss)
# Backward
loss.backward()
# Clipping gradients to avoid gradient explosion
nn.utils.clip_grad_norm_(model.parameters(), 1.)
# Update weights
optimizer.step()
if current_step % hp.log_step == 0:
Now = time.clock()
# print("time per step: %.2f sec" % time_per_step)
# print("At timestep %d" % current_step)
# print("linear loss: %.4f" % linear_loss.data[0])
# print("mel loss: %.4f" % mel_loss.data[0])
# print("total loss: %.4f" % loss.data[0])
# print("Epoch [{}/{}], Step [{}/{}], Linear Loss: {:.4f}, Mel Loss: {:.4f}, Total Loss: {:.4f}.".format(
# epoch+1, hp.epochs, current_step, total_step, linear_loss.item(), mel_loss.item(), loss.item()))
str1 = "Epoch [{}/{}], Step [{}/{}], Linear Loss: {:.4f}, Mel Loss: {:.4f}, Total Loss: {:.4f}.".format(
epoch+1, hp.epochs, current_step, total_step, linear_loss.item(), mel_loss.item(), loss.item())
str2 = "Time Used: {:.3f}s, Estimated Time Remaining: {:.3f}s.".format(
(Now-Start), (total_step-current_step)*np.mean(Time))
# print("Time Used: {:.3f}s, Estimated Time Remaining: {:.3f}s.".format(
# (Now-Start), (total_step-current_step)*np.mean(Time)))
print(str1)
print(str2)
with open("logger.txt", "a")as f_logger:
f_logger.write(str1 + "\n")
f_logger.write(str2 + "\n")
f_logger.write("\n")
# print(current_step)
if current_step % hp.save_step == 0:
torch.save({'model': model.state_dict(), 'optimizer': optimizer.state_dict(
)}, os.path.join(hp.checkpoint_path, 'checkpoint_%d.pth.tar' % current_step))
print("save model at step %d ..." % current_step)
if current_step in hp.decay_step:
optimizer = adjust_learning_rate(optimizer, current_step)
end_time = time.clock()
Time = np.append(Time, end_time - start_time)
if len(Time) == hp.clear_Time:
temp_value = np.mean(Time)
Time = np.delete(
Time, [i for i in range(len(Time))], axis=None)
Time = np.append(Time, temp_value)
# print(Time)
def trans(arr):
return np.stack([np.transpose(ele) for ele in arr])
# for i, b in enumerate(arr):
# arr[i] = np.transpose(b)
def adjust_learning_rate(optimizer, step):
if step == 500000:
# if step == 20:
# print("update")
for param_group in optimizer.param_groups:
param_group['lr'] = 0.0005
elif step == 1000000:
for param_group in optimizer.param_groups:
param_group['lr'] = 0.0003
elif step == 2000000:
for param_group in optimizer.param_groups:
param_group['lr'] = 0.0001
return optimizer
def compare(out, stan, device):
# for batch_index in range(len(out)):
# for i in range(min([np.shape(out)[2], np.shape(stan)[2]])):
# torch.abs(out[batch_index][i], stan[batch_index][i])
# cnt = min([np.shape(out)[2], np.shape(stan)[2]])
if np.shape(stan)[2] >= np.shape(out)[2]:
return stan[:, :, :np.shape(out)[2]]
# return out[:,:,:cnt], stan[:,:,:cnt]
else:
frame_arr = np.zeros([np.shape(out)[0], np.shape(out)[1], np.shape(out)[
2]-np.shape(stan)[2]], dtype=np.float32)
return torch.Tensor(np.concatenate((stan.cpu(), frame_arr), axis=2)).to(device)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--dataset_path', type=str,
help='dataset path', default='dataset')
parser.add_argument('--restore_step', type=int,
help='Global step to restore checkpoint', default=0)
parser.add_argument('--batch_size', type=int,
help='Batch size', default=hp.batch_size)
args = parser.parse_args()
main(args)