-
Notifications
You must be signed in to change notification settings - Fork 3
/
trainer_mul_g9_tfgrid_progressive_pitch.py
303 lines (229 loc) · 12.6 KB
/
trainer_mul_g9_tfgrid_progressive_pitch.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
'''
multi gpu version
'''
import os
import torch
import toml
from datetime import datetime
from tqdm import tqdm
from glob import glob
import soundfile as sf
import numpy as np
import einops
from torch.utils.tensorboard import SummaryWriter
from pystoi import stoi
class Trainer:
def __init__(self, config, model, optimizer, loss_func,
train_dataloader, validation_dataloader,train_sampler, args):
self.model_pret = model[0]
self.model_pitch = model[1]
self.model_comb = model[2]
self.model = model[-1]
self.model_pret.eval()
self.model_pitch.eval()
self.model_comb.eval()
self.optimizer = optimizer
#self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, 'min', factor=0.8, patience=5,verbose=True)
self.loss_func = loss_func
self.train_dataset = train_dataloader
self.validation_dataloader = validation_dataloader
self.train_sampler = train_sampler
self.rank = args.rank
self.device = args.device
#self.WINDOW = torch.sqrt(torch.hann_window(512) + 1e-8).to(self.device)
# training config
self.trainer_config = config['trainer']
self.epochs = self.trainer_config['epochs']
self.save_checkpoint_interval = self.trainer_config['save_checkpoint_interval']
self.clip_grad_norm_value = self.trainer_config['clip_grad_norm_value']
self.resume = self.trainer_config['resume']
if not self.resume:
self.exp_path = self.trainer_config['exp_path'] + '_' + datetime.now().strftime("%Y-%m-%d-%Hh%Mm")
else:
self.exp_path = self.trainer_config['exp_path'] + '_' + self.trainer_config['resume_datetime']
self.log_path = os.path.join(self.exp_path, 'logs')
self.checkpoint_path = os.path.join(self.exp_path, 'checkpoints')
self.sample_path = os.path.join(self.exp_path, 'val_samples')
os.makedirs(self.log_path, exist_ok=True)
os.makedirs(self.checkpoint_path, exist_ok=True)
os.makedirs(self.sample_path, exist_ok=True)
#os.makedirs(self.sample_path + '/spk1', exist_ok=True)
#os.makedirs(self.sample_path + '/spk2', exist_ok=True)
#os.makedirs(self.sample_path + '/spk3', exist_ok=True)
# save the config
if self.rank == 1:
with open(
os.path.join(
self.exp_path, 'config_g8.toml'.format(datetime.now().strftime("%Y-%m-%d-%Hh%Mm"))), 'w') as f:
toml.dump(config, f)
self.writer = SummaryWriter(self.log_path)
self.start_epoch = 1
self.best_score = 0
if self.resume:
self._resume_checkpoint()
self.sr = config['listener']['listener_sr']
self.loss_func = self.loss_func.to(self.device)
def _set_train_mode(self):
self.model.train()
#print(self.model.state_dict()['module.fc.bias'])
def _set_eval_mode(self):
self.model.eval()
#print(self.model.state_dict()['module.fc.bias'])
def _save_checkpoint(self, epoch, score):
state_dict = {'epoch': epoch,
'optimizer': self.optimizer.optimizer.state_dict(),
'model': self.model.state_dict()}
torch.save(state_dict, os.path.join(self.checkpoint_path, f'model_{str(epoch).zfill(4)}.tar'))
if score > self.best_score:
self.state_dict_best = state_dict.copy()
self.best_score = score
def _resume_checkpoint(self):
latest_checkpoints = sorted(glob(os.path.join(self.checkpoint_path, 'model_*.tar')))[-1]
map_location = self.device
checkpoint = torch.load(latest_checkpoints, map_location=map_location)
self.start_epoch = checkpoint['epoch'] + 1
self.optimizer.optimizer.load_state_dict(checkpoint['optimizer'])
self.model.load_state_dict(checkpoint['model'])#module.
def _train_epoch(self, epoch):
total_loss = 0
self.train_dataset.dataset.sample()
self.train_dataloader = tqdm(self.train_dataset, ncols=70)#tqdm(self.train_dataloader, ncols=75)
'''
for param_group in self.optimizer.optimizer.param_groups:
lr = param_group['lr']
print(lr)
'''
for step, data in enumerate(self.train_dataloader, 1):
mixture,_,_,_,_,target = data
mixture = mixture.to(torch.float32).to(self.device) # [B,N,M]
target = target.to(torch.float32).to(self.device) # [B, N]
with torch.no_grad():
esti_tagt = self.model_pret(mixture) # [B,N,M]
batch = esti_tagt[1]
pitch_est = self.model_pitch(esti_tagt[0][-1])
scale_mix0 = mixture / torch.max(torch.abs(mixture), 1)[0][:,None]
scale_mix1 = esti_tagt[0][0] / torch.max(torch.abs(esti_tagt[0][0]), 1)[0][:,None]
scale_mix2 = esti_tagt[0][1] / torch.max(torch.abs(esti_tagt[0][1]), 1)[0][:,None]
scale_mix3 = esti_tagt[0][2] / torch.max(torch.abs(esti_tagt[0][2]), 1)[0][:,None]
pitch_input = torch.argmax(pitch_est, dim=-1)
filtered_mix0 = self.model_comb(scale_mix0, pitch_input)[:,:,:,0]
filtered_spec0 = torch.fft.rfft(filtered_mix0,dim=2)
filtered_spec0 = torch.stack([filtered_spec0.real,filtered_spec0.imag], dim=1) # [B, 2, T, F]
filtered_mix1 = self.model_comb(scale_mix1, pitch_input)[:,:,:,0]
filtered_spec1 = torch.fft.rfft(filtered_mix1,dim=2)
filtered_spec1 = torch.stack([filtered_spec1.real,filtered_spec1.imag], dim=1)
filtered_mix2 = self.model_comb(scale_mix2, pitch_input)[:,:,:,0]
filtered_spec2 = torch.fft.rfft(filtered_mix2,dim=2)
filtered_spec2 = torch.stack([filtered_spec2.real,filtered_spec2.imag], dim=1)
filtered_mix3 = self.model_comb(scale_mix3, pitch_input)[:,:,:,0]
filtered_spec3 = torch.fft.rfft(filtered_mix3,dim=2)
filtered_spec3 = torch.stack([filtered_spec3.real,filtered_spec3.imag], dim=1)
inp = [mixture, filtered_spec0, filtered_spec1, filtered_spec2, filtered_spec3, batch]
esti_tagt = self.model(inp)
#print(target.shape,esti_tagt[0][0].shape)
loss = self.loss_func(esti_tagt[0][0],target)#+ self.loss_func(esti_nois, noise)
loss = torch.mean(loss)#bs!=1时
total_loss += loss.item()
self.train_dataloader.desc = 'train[{}/{}]'.format(
epoch, self.epochs + self.start_epoch-1)
self.train_dataloader.postfix = 'ls{:.2f}'.format(total_loss / step)
self.optimizer.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clip_grad_norm_value)
self.optimizer.step()
#for name, parms in self.model.named_parameters():
# if 'fc.bias' in name:print(print('-->grad_value:',parms.grad))
# 等待所有进程计算完毕
if self.device != torch.device("cpu"):
torch.cuda.synchronize(self.device)
for param_group in self.optimizer.optimizer.param_groups:
lr = param_group['lr']
if self.rank == 1:
self.writer.add_scalars('train_loss', {'train_loss': total_loss / step,'lr':lr}, epoch)
@torch.no_grad()
def _validation_epoch(self, epoch):
total_loss = 0
total_stoi_score = 0
self.validation_dataloader = tqdm(self.validation_dataloader, ncols=70)
for step, data in enumerate(self.validation_dataloader, 1):
mix,_,_,_,_,tgt,name = data
mix = mix.to(torch.float32).to(self.device)
tgt = tgt.to(torch.float32).to(self.device)
esti_tagt = self.model_pret(mix) # [B,N,M]
batch = esti_tagt[1]
pitch_est = self.model_pitch(esti_tagt[0][-1])
scale_mix0 = mix / torch.max(torch.abs(mix), 1)[0][:,None]
scale_mix1 = esti_tagt[0][0] / torch.max(torch.abs(esti_tagt[0][0]), 1)[0][:,None]
scale_mix2 = esti_tagt[0][1] / torch.max(torch.abs(esti_tagt[0][1]), 1)[0][:,None]
scale_mix3 = esti_tagt[0][2] / torch.max(torch.abs(esti_tagt[0][2]), 1)[0][:,None]
pitch_input = torch.argmax(pitch_est, dim=-1)
filtered_mix0 = self.model_comb(scale_mix0, pitch_input)[:,:,:,0]
filtered_spec0 = torch.fft.rfft(filtered_mix0,dim=2)
filtered_spec0 = torch.stack([filtered_spec0.real,filtered_spec0.imag], dim=1) # [B, 2, T, F]
filtered_mix1 = self.model_comb(scale_mix1, pitch_input)[:,:,:,0]
filtered_spec1 = torch.fft.rfft(filtered_mix1,dim=2)
filtered_spec1 = torch.stack([filtered_spec1.real,filtered_spec1.imag], dim=1)
filtered_mix2 = self.model_comb(scale_mix2, pitch_input)[:,:,:,0]
filtered_spec2 = torch.fft.rfft(filtered_mix2,dim=2)
filtered_spec2 = torch.stack([filtered_spec2.real,filtered_spec2.imag], dim=1)
filtered_mix3 = self.model_comb(scale_mix3, pitch_input)[:,:,:,0]
filtered_spec3 = torch.fft.rfft(filtered_mix3,dim=2)
filtered_spec3 = torch.stack([filtered_spec3.real,filtered_spec3.imag], dim=1)
inp = [mix, filtered_spec0, filtered_spec1, filtered_spec2, filtered_spec3, batch]
# FT-JNF
esti_tagt = self.model(inp) # [B, F, T, 2]
#print(tgt.shape,esti_tagt[0][0].shape)
loss = self.loss_func(esti_tagt[0][0],tgt)# + self.loss_func(esti_nois, noise)
total_loss += loss.item()
enhanced = esti_tagt[0][0].squeeze().cpu().numpy()
clean = tgt.squeeze().cpu().numpy()
enh_len = enhanced.shape[-1]
###
stoi_score = stoi(enhanced, clean[0:enh_len], 16000, extended=True)
total_stoi_score += stoi_score
if step<3:
sf.write(os.path.join(self.sample_path,
'{}_enhanced_epoch{}_estoi={:.3f}.wav'.format(name[0],epoch, stoi_score)),
enhanced, 16000)
sf.write(os.path.join(self.sample_path,
'{}_clean.wav'.format(name[0])),
clean, 16000)
# enhanced = enhanced / enhanced.max() * 0.5
self.validation_dataloader.desc = 'val[{}/{}]'.format(
epoch, self.epochs + self.start_epoch-1)
self.validation_dataloader.postfix = 'ls{:.2f},est{:.2f}'.format(
total_loss / step, total_stoi_score / step)
# 等待所有进程计算完毕
if self.device != torch.device("cpu"):
torch.cuda.synchronize(self.device)
if self.rank == 1:
self.writer.add_scalars(
'val_loss', {'val_loss': total_loss / step,
'estoi': total_stoi_score / step}, epoch)
return total_loss / step, total_stoi_score / step
torch.cuda.empty_cache()
def train(self):
if self.rank == 1:
timestamp_txt = os.path.join(self.exp_path, 'timestamp.txt')
mode = 'a' if os.path.exists(timestamp_txt) else 'w'
with open(timestamp_txt, mode) as f:
f.write('[{}] start for {} epochs\n'.format(
datetime.now().strftime("%Y-%m-%d-%H:%M"), self.epochs))
if self.resume:
self._resume_checkpoint()
for epoch in range(self.start_epoch, self.epochs + self.start_epoch):
self.train_sampler.set_epoch(epoch)
self._set_train_mode()
self._train_epoch(epoch)
self._set_eval_mode()
valid_loss, score = self._validation_epoch(epoch)
#self.scheduler.step(valid_loss)
if (self.rank == 1) and (epoch % self.save_checkpoint_interval == 0):
self._save_checkpoint(epoch, score)
if self.rank == 1:
torch.save(self.state_dict_best,
os.path.join(self.checkpoint_path,
'best_model_{}.tar'.format(str(self.state_dict_best['epoch']).zfill(4))))
print('------------Training for {} epochs has done!------------'.format(self.epochs))
with open(timestamp_txt, 'a') as f:
f.write('[{}] end\n'.format(datetime.now().strftime("%Y-%m-%d-%H:%M")))