-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
271 lines (232 loc) · 10.9 KB
/
trainer.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
import torch.nn as nn
import torch, time, torchaudio
import pandas as pd
import os, sys
from tqdm import tqdm
import numpy as np
from utils.util import get_filepaths, check_folder, cal_score, get_feature, progress_bar
from utils.eval_composite import eval_composite
# from model.res_LSTM import res_LSTM
class Trainer:
def __init__(self, model, epochs, epoch, best_loss, optimizer,
criterion, device, loader,Test_path, writer, model_path, score_path, args):
# self.step = 0
self.epoch = epoch
self.epochs = epochs
self.best_loss = best_loss
self.model = model.to(device)
self.optimizer = optimizer
# fea = {
# 'BLSTM' :'log1p',
# 'DPTNet' :'wav'
# }
fea = {
'odconv' :'log1p',
'DPTNet' :'wav',
'BLSTM':'log1p',
'odconv_revise' :'log1p',
'odconv2d' : 'log1p',
'res_LSTM' : 'log1p',
'CBAM' : 'log1p',
'sConformer' : 'log1p',
'new_conformer': 'log1p',
'new_conformer_1': 'log1p',
'new_conformer_2': 'log1p',
'new_conformer_3': 'log1p',
'new_conformer_4': 'log1p',
'new_conformer_5': 'log1p',
'new_conformer_6': 'log1p',
'new_conformer_7': 'log1p',
'new_conformer_7re': 'log1p',
'new_conformer_7res': 'log1p',
'new_conformer_7res2conv': 'log1p',
'new_conformer_7res_conv': 'log1p',
'new_conformer_7res_layer6': 'log1p',
'new_conformer_7res_layer6_2ssl': 'log1p',
'new_conformer_7res_layer6_2sslw': 'log1p',
'new_conformer_7res_feature': 'log1p',
'new_conformer_7res_fastcnn': 'log1p',
'new_conformer_7se': 'log1p',
'new_conformer_8': 'log1p',
'new_conformer_9': 'log1p',
'new_conformer_7res2': 'log1p',
'new_conformer_7res3': 'log1p',
'new_conformer_7res4': 'log1p',
'new_conformer_7res5': 'log1p',
'new_conformer_10': 'log1p',
'new_conformer_11': 'log1p',
'new_conformer_12': 'log1p',
'new_conformer_13': 'log1p',
'test_1conv': 'log1p',
'FAodconv': 'log1p',
'FAodconv01': 'log1p',
'conformer_res_7_conv': 'log1p',
}
self.fea = fea[args.model]
self.device = device
self.loader = loader
self.criterion = criterion
self.Test_path = Test_path
self.train_loss = 0
self.val_loss = 0
self.writer = writer
self.model_path = model_path
self.score_path = score_path
self.args = args
self.transform = get_feature()
if not args.finetune_SSL and args.feature!='raw':
self.model.model_SSL.eval()
for name,param in self.model.model_SSL.named_parameters():
param.requires_grad = False
if args.finetune_SSL=='PF':
for name,param in self.model.model_SSL.feature_extractor.named_parameters():
param.requires_grad = False
def save_checkpoint(self,):
save_model = self.model.module if isinstance(self.model, nn.DataParallel) else self.model
state_dict = {
'epoch': self.epoch,
'model': save_model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'best_loss': self.best_loss
}
check_folder(self.model_path_)
torch.save(state_dict, self.model_path_)
def get_fea(self,wav,ftype='log1p'):
if ftype=='wav':
return wav
elif ftype=='complex':
return self.transform(wav,ftype=ftype)[0]
else:
return self.transform(wav,ftype=ftype)[0][0]
def _train_step(self, nwav,cwav):
device = self.device
nwav,cwav = nwav.to(device),cwav.to(device)
# print("cwav",cwav.shape)
cdata = self.get_fea(cwav, ftype=self.fea)
# print("cdata",cdata.shape)
pred = self.model(nwav)
loss = self.criterion(pred,cdata)
self.train_loss += loss.item()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# if USE_GRAD_NORM:
# nn.utils.clip_grad_norm_(self.model['discriminator'].parameters(), DISCRIMINATOR_GRAD_NORM)
# self.optimizer['discriminator'].step()
def _train_epoch(self):
self.train_loss = 0
self.model.train()
if not self.args.finetune_SSL and self.args.feature!='raw':
self.model.model_SSL.eval()
step = 0
t_start = time.time()
for nwav,cwav in self.loader['train']:
#(batchsize,N,T,2)
self._train_step(nwav,cwav)
step += 1
progress_bar(self.epoch,self.epochs,step,self.train_step,time.time()-t_start,loss=self.train_loss,mode='train')
self.train_loss /= len(self.loader['train'])
print(f'train_loss:{self.train_loss}')
# @torch.no_grad()
def _val_step(self, nwav,cwav):
device = self.device
nwav,cwav = nwav.to(device),cwav.to(device)
cdata = self.get_fea(cwav, ftype=self.fea)
pred = self.model(nwav)
loss = self.criterion(pred,cdata)
self.val_loss += loss.item()
# def _val_epoch(self):
# self.val_loss = 0
# self.model.eval()
# step = 0
# t_start = time.time()
# for nwav,cwav in self.loader['val']:
# self._val_step(nwav,cwav)
# step += 1
# progress_bar(self.epoch,self.epochs,step,self.val_step,time.time()-t_start,loss=self.val_loss,mode='test')
# self.val_loss /= len(self.loader['val'])
# print(f'val_loss:{self.val_loss}')
# if self.best_loss > self.val_loss:
# print(f"Save model to '{self.model_path}'")
# self.save_checkpoint()
# self.best_loss = self.val_loss
def _val_epoch(self):
self.val_loss = 0
self.model.eval()
step = 0
t_start = time.time()
for nwav,cwav in self.loader['val']:
self._val_step(nwav,cwav)
step += 1
progress_bar(self.epoch,self.epochs,step,self.val_step,time.time()-t_start,loss=self.val_loss,mode='test')
self.val_loss /= len(self.loader['val'])
print(f'val_loss:{self.val_loss}')
if self.best_loss > self.val_loss and self.epoch <= 15:
self.model_path_ = self.model_path+'_pre15.pth.tar'
print(f"Save model to '{self.model_path_ }'")
self.save_checkpoint()
self.best_loss = self.val_loss
if self.best_loss > self.val_loss and self.epoch > 15:
self.model_path_ = self.model_path+'_'+str(self.epoch)+'.pth.tar'
print(f"Save model to '{self.model_path_ }'")
self.save_checkpoint()
self.best_loss = self.val_loss
def write_score(self,test_file,c_path):
args = self.args
wavname = test_file.split('/')[-1]
c_file = os.path.join(c_path,wavname)
n_data,sr = torchaudio.load(test_file)
c_data,sr = torchaudio.load(c_file)
enhanced = self.model(n_data.to(self.device),output_wav=True)
out_path = f'/home/zt/ssl/base_demo_BSSE-SE/exp/Enhanced/{self.model.__class__.__name__}_{args.model}_{args.ssl_model}_{args.target}_epochs{args.epochs}' \
f'_{args.optim}_{args.loss_fn}_batch{args.batch_size}_'\
f'lr{args.lr}_{args.feature}_{args.size}_'\
f'WS{args.weighted_sum}_FT{args.finetune_SSL}/{wavname}'
check_folder(out_path)
enhanced = enhanced.cpu()
torchaudio.save(out_path,enhanced,sr)
s_pesq, s_stoi = cal_score(c_data.squeeze().detach().numpy(),enhanced.squeeze().detach().numpy())
date = eval_composite(c_data.squeeze().detach().numpy(),enhanced.squeeze().detach().numpy())
with open(self.score_path_, 'a') as f:
f.write(f'{wavname},{s_pesq},{s_stoi},{date["csig"]},{date["cbak"]},{date["covl"]},{date["pesq"]},{date["stoi"]},{date["ssnr"]}\n')
def train(self):
args = self.args
model_name = self.model.module.__class__.__name__ if isinstance(self.model, nn.DataParallel) else self.model.__class__.__name__
# figname = f'{self.args.task}/{model_name}_{args.model}_{args.ssl_model}_{args.target}_{args.feature}_{args.size}_WS{args.weighted_sum}_FT{args.finetune_SSL}'
figname = f'{self.args.task}/{model_name}_{args.model}_{args.ssl_model}_{args.target}_{args.lr}_{args.loss_fn}_{args.batch_size}_{args.feature}_{args.size}_WS{args.weighted_sum}_FT{args.finetune_SSL}'
self.train_step = len(self.loader['train'])
self.val_step = len(self.loader['val'])
while self.epoch < self.epochs:
self._train_epoch()
self._val_epoch()
self.writer.add_scalars(f'{figname}', {'train': self.train_loss},self.epoch)
self.writer.add_scalars(f'{figname}', {'val': self.val_loss},self.epoch)
self.epoch += 1
def test(self,model_path_test):
# load model
#后加的
self.score_path_ = self.score_path
self.score_path_=self.score_path_+'_'+model_path_test.split('_')[-1].split('.')[0]+'.csv'
self.model.eval()
checkpoint = torch.load(model_path_test)
self.model.load_state_dict(checkpoint['model'])
noisy_paths = get_filepaths(self.Test_path['noisy'])
check_folder(self.score_path_)
if os.path.exists(self.score_path_):
os.remove(self.score_path_)
with open(self.score_path_, 'a') as f:
f.write('Filename,PESQ,STOI,MY_CSIG,MY_CBAK,MY_COVL,MY_PESQ,MY_STOI,MY_SSNR\n')
for noisy_path in tqdm(noisy_paths):
self.write_score(noisy_path,self.Test_path['clean'])
data = pd.read_csv(self.score_path_)
pesq_mean = data['PESQ'].to_numpy().astype('float').mean()
stoi_mean = data['STOI'].to_numpy().astype('float').mean()
my_csig_mean = data['MY_CSIG'].to_numpy().astype('float').mean()
my_cbak_mean = data['MY_CBAK'].to_numpy().astype('float').mean()
my_covl_mean = data['MY_COVL'].to_numpy().astype('float').mean()
my_pesq_mean = data['MY_PESQ'].to_numpy().astype('float').mean()
my_stoi_mean = data['MY_STOI'].to_numpy().astype('float').mean()
my_ssnr_mean = data['MY_SSNR'].to_numpy().astype('float').mean()
with open(self.score_path_, 'a') as f:
f.write(','.join(('Average',str(pesq_mean),str(stoi_mean),str(my_csig_mean),str(my_cbak_mean),
str(my_covl_mean),str(my_pesq_mean),str(my_stoi_mean),str(my_ssnr_mean),))+'\n')