-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain.py
329 lines (287 loc) · 11.8 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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
#----------------description----------------#
# Author : Lei yuan
# E-mail : [email protected]
# Company : Fudan University
# Date : 2020-10-10 17:40:40
# LastEditors : Zihao Zhao
# LastEditTime : 2021-05-08 12:32:21
# FilePath : /pytorch-asr-wavenet/train.py
# Description : 0.001 0-5, 0.0001
#-------------------------------------------#
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
import torch.optim as optim
import torch.nn.functional as F
import torch.nn.utils.rnn as rnn_utils
import deepdish as dd
import config_train as cfg
from dataset import VCTK
import dataset
from wavenet import WaveNet
from sparsity import *
import utils
import visualize as vis
from ctcdecode import CTCBeamDecoder
from tensorboardX import SummaryWriter
import os
import numpy as np
import time
import argparse
from write_excel import *
import torch.onnx
def parse_args():
'''
Parse input arguments
'''
parser = argparse.ArgumentParser(description='WaveNet for speech recognition.')
parser.add_argument('--exp', type=str, help='exp dir', default="default")
parser.add_argument('--resume', action='store_true', help='resume from exp_name/best.pth', default=False)
parser.add_argument('--batch_size', type=int, help='1, 16, 32', default=16)
parser.add_argument('--lr', type=float, help='0.001 for tensorflow', default=0.001)
parser.add_argument('--load_from', type=str, help='.pth', default="/z")
parser.add_argument('--skip_exist', action='store_true', help='if exist', default=False)
parser.add_argument('--save_excel', type=str, help='exp.xls', default="default.xls")
args = parser.parse_args()
return args
def train(train_loader, scheduler, model, loss_fn, val_loader, writer=None):
decoder_vocabulary = utils.Data.decoder_vocabulary
vocabulary = utils.Data.vocabulary
decoder = CTCBeamDecoder(
decoder_vocabulary,
#"_abcdefghijklmopqrstuvwxyz_",
model_path=None,
alpha=0,
beta=0,
cutoff_top_n=40,
cutoff_prob=1.0,
beam_width=100,
num_processes=4,
blank_id=27,
log_probs_input=True
)
train_loss_list = list()
val_loss_list = list()
best_loss = float('inf')
for epoch in range(cfg.epochs):
print(f'Training epoch {epoch}')
_loss = 0.0
step_cnt = 0
_tp, _pred, _pos = 0, 0, 0
for data in train_loader:
wave = data['wave'].cuda() # [1, 128, 109]
if epoch == 0 and step_cnt == 0:
# print("test3")
loss_val = validate(val_loader, model, loss_fn)
writer.add_scalar('val/loss', loss_val, epoch)
best_loss = loss_val
not_better_cnt = 0
torch.save(model.state_dict(), cfg.workdir+'/weights/best.pth')
print("saved", cfg.workdir+'/weights/best.pth', not_better_cnt)
val_loss_list.append(float(loss_val))
model.train()
logits = model(wave)
mask = torch.zeros_like(logits)
for n in range(len(data['length_wave'])):
mask[:, :, :data['length_wave'][n]] = 1
logits *= mask
logits = logits.permute(2, 0, 1)
logits = F.log_softmax(logits, dim=2)
if data['text'].size(0) == cfg.batch_size:
for i in range(cfg.batch_size):
if i == 0:
text = data['text'][i][0:data['length_text'][i]].cuda()
else:
text = torch.cat([text,
data['text'][i][0: data['length_text'][i]].cuda()])
else:
continue
loss = 0.0
for i in range(logits.size(1)):
loss += loss_fn(logits[:data['length_wave'][i], i:i+1, :], data['text'][i][0:data['length_text'][i]].cuda(), data['length_wave'][i], data['length_text'][i])
loss /= logits.size(1)
scheduler.zero_grad()
loss.backward()
scheduler.step()
_loss += loss.data
step_cnt += 1
_loss /= len(train_loader)
writer.add_scalar('train/loss', _loss, epoch)
train_loss_list.append(float(_loss))
torch.cuda.empty_cache()
loss_val = validate(val_loader, model, loss_fn)
writer.add_scalar('val/loss', loss_val, epoch)
val_loss_list.append(float(loss_val))
model.train()
if loss_val < best_loss:
not_better_cnt = 0
torch.save(model.state_dict(), cfg.workdir+f'/weights/best.pth')
print("saved", cfg.workdir+f'/weights/best.pth', not_better_cnt)
best_loss = loss_val
else:
not_better_cnt += 1
if not_better_cnt > 3:
write_excel(os.path.join(cfg.work_root, cfg.save_excel),
cfg.exp_name, train_loss_list, val_loss_list)
# exit()
def validate(val_loader, model, loss_fn):
decoder_vocabulary = utils.Data.decoder_vocabulary
vocabulary = utils.Data.vocabulary
decoder = CTCBeamDecoder(
decoder_vocabulary,
#"_abcdefghijklmopqrstuvwxyz_",
model_path=None,
alpha=0,
beta=0,
cutoff_top_n=40,
cutoff_prob=1.0,
beam_width=100,
num_processes=4,
blank_id=27,
log_probs_input=True
)
model.eval()
_loss = 0.0
step_cnt = 0
_tp, _pred, _pos = 0, 0, 0
with torch.no_grad():
for data in val_loader:
wave = data['wave'].cuda() # [1, 128, 109]
logits = model(wave)
logits = logits.permute(2, 0, 1)
logits = F.log_softmax(logits + 1e-10, dim=2)
if data['text'].size(0) == cfg.batch_size:
for i in range(cfg.batch_size):
if i == 0:
text = data['text'][i][0:data['length_text'][i]].cuda()
# print(data['text'].size())
# print(data['length_text'][i])
else:
text = torch.cat([text,
data['text'][i][0: data['length_text'][i]].cuda()])
else:
continue
loss = 0.0
for i in range(logits.size(1)):
loss += loss_fn(logits[:data['length_wave'][i], i:i+1, :], data['text'][i][0:data['length_text'][i]].cuda(), data['length_wave'][i], data['length_text'][i])
loss /= logits.size(1)
_loss += loss.data
# beam_results, beam_scores, timesteps, out_lens = decoder.decode(logits.permute(1, 0, 2))
# voc = np.tile(vocabulary, (cfg.batch_size, 1))
# pred = np.take(voc, beam_results[0][0][:out_lens[0][0]].data.numpy())
# text_np = np.take(voc, data['text'][0][0:data['length_text'][0]].cpu().numpy().astype(int))
# tp, pred, pos = utils.evalutes(utils.cvt_np2string(pred), utils.cvt_np2string(text_np))
# _tp += tp
# _pred += pred
# _pos += pos
# f1 = 2 * _tp / (_pred + _pos + 1e-10)
step_cnt += 1
print("Val step", step_cnt, "/", len(val_loader),
", loss: ", round(float(_loss/len(val_loader)), 5))
return _loss/len(val_loader)
def test_acc(val_loader, model, loss_fn):
decoder_vocabulary = utils.Data.decoder_vocabulary
vocabulary = utils.Data.vocabulary
decoder = CTCBeamDecoder(
decoder_vocabulary,
#"_abcdefghijklmopqrstuvwxyz_",
model_path=None,
alpha=0,
beta=0,
cutoff_top_n=40,
cutoff_prob=1.0,
beam_width=100,
num_processes=4,
blank_id=27,
log_probs_input=True
)
model.eval()
_loss = 0.0
step_cnt = 0
tps, preds, poses = 0, 0, 0
f_cnt = 0
with torch.no_grad():
for data in val_loader:
wave = data['wave'].cuda() # [1, 128, 109]
if 1:
print(data['wave'].size())
np.savetxt("/zhzhao/dataset/VCTK/c_model_input_txt/"+str(f_cnt)+".txt", data['wave'].flatten())
print(f_cnt)
f_cnt += 1
logits = model(wave)
logits = logits.permute(2, 0, 1)
logits = F.log_softmax(logits, dim=2)
if data['text'].size(0) == cfg.batch_size:
for i in range(cfg.batch_size):
if i == 0:
text = data['text'][i][0:data['length_text'][i]].cuda()
else:
text = torch.cat([text,
data['text'][i][0: data['length_text'][i]].cuda()])
else:
continue
loss = loss_fn(logits, text, data['length_wave'], data['length_text'])
_loss += loss.data
for i in range(logits.size(1)):
logit = logits[:data['length_wave'][i], i:i+1, :]
beam_results, beam_scores, timesteps, out_lens = decoder.decode(logit.permute(1, 0, 2))
voc = np.tile(vocabulary, (cfg.batch_size, 1))
pred = np.take(voc, beam_results[0][0][:out_lens[0][0]].data.numpy())
text_np = np.take(voc, data['text'][i][0:data['length_text'][i]].cpu().numpy().astype(int))
pred = [pred]
text_np = [text_np]
tp, pred, pos = utils.evalutes(utils.cvt_np2string(pred), utils.cvt_np2string(text_np))
tps += tp
preds += pred
poses += pos
f1 = 2 * tps / (preds + poses + 1e-10)
step_cnt += 1
# if cnt % 10 == 0:
print("Val step", step_cnt, "/", len(val_loader),
", loss: ", round(float(_loss.data/step_cnt), 5))
print("Val tps:", tps, ",preds:", preds, ",poses:", poses, ",f1:", f1)
return f1, _loss/len(val_loader), tps, preds, poses
def check_and_mkdir(dir):
if not os.path.exists(dir):
os.makedirs(dir)
def main():
args = parse_args()
cfg.resume = args.resume
cfg.exp_name = args.exp
cfg.work_root = '/zhzhao/code/wavenet_torch/torch_lyuan/exp_result/'
cfg.workdir = cfg.work_root + args.exp + '/debug'
cfg.sparse_mode = args.sparse_mode
cfg.batch_size = args.batch_size
cfg.lr = args.lr
cfg.load_from = args.load_from
cfg.save_excel = args.save_excel
weights_dir = os.path.join(cfg.workdir, 'weights')
check_and_mkdir(weights_dir)
print('initial training...')
print(f'work_dir:{cfg.workdir}, \n\
pretrained: {cfg.load_from}, \n\
batch_size: {cfg.batch_size}, \n\
lr : {cfg.lr}, \n\
epochs : {cfg.epochs}, \n\
sparse : {cfg.sparse_mode}')
writer = SummaryWriter(log_dir=cfg.workdir+'/runs')
# build train data
vctk_train = VCTK(cfg, 'train')
train_loader = DataLoader(vctk_train, batch_size=cfg.batch_size, num_workers=4, shuffle=True, pin_memory=True)
vctk_val = VCTK(cfg, 'val')
val_loader = DataLoader(vctk_val, batch_size=cfg.batch_size, num_workers=4, shuffle=False, pin_memory=True)
# build model
model = WaveNet(num_classes=28, channels_in=40, dilations=[1,2,4,8,16])
model = nn.DataParallel(model)
model.cuda()
model.train()
# build loss
loss_fn = nn.CTCLoss(blank=27)
if cfg.resume and os.path.exists(cfg.workdir + '/weights/best.pth'):
model.load_state_dict(torch.load(cfg.workdir + '/weights/best.pth'), strict=True)
print("loading", cfg.workdir + '/weights/best.pth')
cfg.load_from = cfg.workdir + '/weights/best.pth'
scheduler = optim.Adam(model.parameters(), lr=cfg.lr, eps=1e-4)
train(train_loader, scheduler, model, loss_fn, val_loader, writer)
if __name__ == '__main__':
main()