-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_nele.py
431 lines (347 loc) · 17 KB
/
train_nele.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
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
# coding=utf-8
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from joblib import Parallel, delayed
import shutil
import scipy.io
import librosa
import os
import time
import numpy as np
import numpy.matlib
import random
import subprocess
import torch
import torch.nn as nn
from audio_util import *
from pystoi.stoi import stoi
from model import Generator_Conv1D_cLN, Discriminator, Discriminator_Quality
from dataloader import *
from tqdm import tqdm
import soundfile as sf
import pdb
random.seed(666)
TargetMetric='siib&haspi&estoi'
output_path='./output'
pt_dir = './chkpt'
GAN_epoch = 500
num_of_sampling = 300 # 300, in each epoch randonly sample only 300 for training
num_of_valid_sample = 480 # number of validation audio
batch_size = 1
fs = 16000 # sampling rate
p_power = (1/6) # power-law compression
inv_p = 6 # inverse of p_power
weight_qua = 0.5 # weight controlling quality item, given in Eq.(7)
creatdir(pt_dir)
creatdir(output_path)
######################### Training data #######################
# You should replace data path to your own
print('Reading path of training data...')
RIR_path = '/home/smg/haoyuli/datasets/Harvard_SI/RIR_DB16k/'
Train_Noise_path = '/home/smg/haoyuli/datasets/Harvard_SI/Train/Noise/'
Train_Clean_path = '/home/smg/haoyuli/datasets/Harvard_SI/Train/Clean/'
Train_Enhan_path = '/home/smg/haoyuli/datasets/Harvard_SI/Train/MultiEnh/' # contains pre-enhanced speech examples y_hat shown in Eqs.(5) and (6). In our paper, we use SSDRC to generate them
Generator_Train_paths = get_filepaths('/home/smg/haoyuli/datasets/Harvard_SI/Train/Clean/')
# Data_shuffle
random.shuffle(Generator_Train_paths)
######################### validation data #########################
# You should replace the addresses to your own
print('Reading path of validation data...')
Test_Noise_path ='/home/smg/haoyuli/datasets/Harvard_SI/Test/Noise/'
Test_Clean_path = '/home/smg/haoyuli/datasets/Harvard_SI/Test/Clean/'
Generator_Test_paths = get_filepaths('/home/smg/haoyuli/datasets/Harvard_SI/Test/Clean/')
Generator_Test_paths = [x for x in Generator_Test_paths if 'AIR_stairway' in x]
assert len(Generator_Test_paths)==720 # Just for double-check data size, change it to fit in you dataset
# Data_shuffle
random.shuffle(Generator_Test_paths)
################################################################
G = Generator_Conv1D_cLN().cuda()
D = Discriminator().cuda()
D_Qua = Discriminator_Quality().cuda()
# Load trained model
# chkpt_path = '/home/smg/haoyuli/SI-Extend/NELE-GAN/trained_model/chkpt_GD.pt'
# save_model = torch.load(chkpt_path)['enhance-model']
# model_dict = G.state_dict()
# state_dict = {k:v for k,v in save_model.items() if k in model_dict.keys()}
# model_dict.update(state_dict)
# G.load_state_dict(model_dict)
# D.load_state_dict(torch.load(chkpt_path)['intel-model'])
# D_Qua.load_state_dict(torch.load(chkpt_path)['quality-model'])
# print('Load Chkpt Finished')
MSELoss = nn.MSELoss().cuda()
optimizer_g = torch.optim.Adam(G.parameters(), lr=5e-4)
optimizer_d = torch.optim.Adam(D.parameters(), lr=2.5e-4)
optimizer_dqua = torch.optim.Adam(D_Qua.parameters(), lr=2.5e-4)
Test_HASPI = []
Test_ESTOI = []
Test_SIIB = []
Test_PESQ = []
Test_VISQOL = []
Previous_Discriminator_training_list = []
shutil.rmtree(output_path)
step_g = 0
step_d = 0
cuda_device = '0'
os.environ["CUDA_VISIBLE_DEVICES"] = cuda_device
# OCCUPY_MEM = occumpy_mem(cuda_device, 0.5) # occupy 50% memory
for gan_epoch in np.arange(1, GAN_epoch+1):
# Prepare directories
creatdir(output_path+"/epoch"+str(gan_epoch))
creatdir(output_path+"/epoch"+str(gan_epoch)+"/"+"Test_epoch"+str(gan_epoch))
creatdir(output_path+'/For_discriminator_training')
creatdir(output_path+'/temp')
# random sample some training data
random.shuffle(Generator_Train_paths)
genloader = create_dataloader(Generator_Train_paths[0:round(1*num_of_sampling)],Train_Noise_path)
if gan_epoch>=2:
print('Generator training (with discriminator fixed)...')
for clean_band, clean_mag, clean_phase, noise_band, noise_mag, noise_phase, target, target_qua, filename in tqdm(genloader):
clean_band = clean_band.cuda()
noise_band = noise_band.cuda()
target = target.cuda()
target_qua = target_qua.cuda()
mask = G(clean_band, noise_band) # outout mask is actually alpha^2 shown in paper, which should be applied to power spectrum
# Do utterance-level energy normalization
clean_power = torch.pow(clean_band.detach(), inv_p)
beta_2 = torch.sum(clean_power) / torch.sum(mask*clean_power)
# Comment parts are frame-level energy normalization Eq.(10) in paper
# beta_2 = torch.sum(clean_power, dim=2) / torch.sum(mask*clean_power, dim=2)
# beta_2 = beta_2.unsqueeze(2)
beta_p = beta_2 ** p_power
enh_band = clean_band * torch.pow(mask, p_power) * beta_p
ref_band = clean_band.detach()
enh_band = enh_band.view(1,1,enh_band.shape[1],enh_band.shape[2]).transpose(2,3).contiguous()
noise_band = noise_band.view(1,1,noise_band.shape[1],noise_band.shape[2]).transpose(2,3).contiguous()
ref_band = ref_band.view(1,1,ref_band.shape[1],ref_band.shape[2]).transpose(2,3).contiguous()
d_inputs = torch.cat((enh_band,noise_band,ref_band),dim=1)
d_inputs_qua = torch.cat((enh_band, ref_band),dim=1)
score = D(d_inputs)
score_qua = D_Qua(d_inputs_qua)
loss = MSELoss(score, target) + weight_qua * MSELoss(score_qua, target_qua)
optimizer_g.zero_grad()
loss.backward()
optimizer_g.step()
step_g += 1
# Evaluate the performance of generator in a validation set.
interval_epoch = 1
if gan_epoch % interval_epoch == 0:
print('Evaluate G by validation data ...')
Test_enhanced_Name = []
utterance = 0
G.eval()
with torch.no_grad():
for i, path in enumerate(Generator_Test_paths[0:num_of_valid_sample]):
S = path.split('/')
wave_name = S[-1]
clean_wav,sr = librosa.load(path, sr=None)
assert sr==16000
noise_wav,sr = librosa.load(Test_Noise_path+wave_name, sr=None)
assert sr==16000
clean_band, clean_mag, clean_phase = Sp_and_phase_Speech(clean_wav, power=p_power, Normalization=True)
noise_band, noise_mag, noise_phase = Sp_and_phase_Noise(noise_wav, power=p_power, Normalization=True)
clean_in = clean_band.reshape(1,clean_band.shape[0],-1)
clean_in = torch.from_numpy(clean_in).cuda()
noise_in = noise_band.reshape(1,noise_band.shape[0],-1)
noise_in = torch.from_numpy(noise_in).cuda()
mask = G(clean_in, noise_in)
clean_power = torch.pow(clean_in, inv_p)
beta_2 = torch.sum(clean_power) / torch.sum(mask*clean_power)
# beta_2 = torch.sum(clean_power, dim=2) / torch.sum(mask*clean_power, dim=2)
# beta_2 = beta_2.unsqueeze(2)
mask = mask * beta_2 # normed alpha2
mask = mask.detach().cpu().squeeze(0).numpy()
enh_wav = SP_to_wav(mask, clean_mag, clean_phase)
if utterance<20: # Only seperatly save the firt 20 utterance for listening comparision
enhanced_name=output_path+"/epoch"+str(gan_epoch)+"/"+"Test_epoch"+str(gan_epoch)+"/"+ wave_name[0:-4]+"@"+str(gan_epoch)+wave_name[-4:]
else:
enhanced_name=output_path+"/temp"+"/"+ wave_name[0:-4]+"@"+str(gan_epoch)+wave_name[-4:]
sf.write(enhanced_name, enh_wav, fs,'PCM_16')
utterance+=1
Test_enhanced_Name.append(enhanced_name)
#print(i)
G.train()
# Calculate True HASPI
test_HASPI = read_batch_HASPI(Test_Clean_path, Test_Noise_path, Test_enhanced_Name, norm=False)
Test_HASPI.append(np.mean(test_HASPI))
# Calculate True ESTOI
test_ESTOI = read_batch_STOI(Test_Clean_path, Test_Noise_path, Test_enhanced_Name, norm=False)
Test_ESTOI.append(np.mean(test_ESTOI))
# Calculate True SIIB
test_SIIB = read_batch_SIIB(Test_Clean_path, Test_Noise_path, Test_enhanced_Name, norm=False)
Test_SIIB.append(np.mean(test_SIIB))
# Calculate True PESQ
test_PESQ = read_batch_PESQ(Test_Clean_path, Test_enhanced_Name, norm=False)
Test_PESQ.append(np.mean(test_PESQ))
# Calculate True VISQOL
test_VISQOL = read_batch_VISQOL(Test_Clean_path, Test_enhanced_Name, norm=False)
Test_VISQOL.append(np.mean(test_VISQOL))
with open('./log.txt','a') as f:
f.write('SIIB is %.3f, HASPI is %.3f, ESTOI is %.3f, PESQ is %.3f, VISQOL is %.3f, EPOCH:%d \n'%(np.mean(test_SIIB), np.mean(test_HASPI), np.mean(test_ESTOI), 0, 0, gan_epoch))
# Plot learning curves
plt.figure(1)
plt.plot(range(1,gan_epoch+1,interval_epoch),Test_HASPI,'b',label='ValidHASPI')
plt.xlim([1,gan_epoch])
plt.xlabel('GAN_epoch')
plt.ylabel('HASPI')
plt.grid(True)
plt.show()
plt.savefig('Test_HASPI.png', dpi=150)
plt.figure(2)
plt.plot(range(1,gan_epoch+1,interval_epoch),Test_SIIB,'r',label='ValidSIIB')
plt.xlim([1,gan_epoch])
plt.xlabel('GAN_epoch')
plt.ylabel('SIIB')
plt.grid(True)
plt.show()
plt.savefig('Test_SIIB.png', dpi=150)
plt.figure(3)
plt.plot(range(1,gan_epoch+1,interval_epoch),Test_ESTOI,'b',label='ValidESTOI')
plt.xlim([1,gan_epoch])
plt.xlabel('GAN_epoch')
plt.ylabel('ESTOI')
plt.grid(True)
plt.show()
plt.savefig('Test_ESTOI.png', dpi=150)
plt.figure(3)
plt.plot(range(1,gan_epoch+1,interval_epoch),Test_PESQ,'b',label='ValidPESQ')
plt.xlim([1,gan_epoch])
plt.xlabel('GAN_epoch')
plt.ylabel('PESQ')
plt.grid(True)
plt.show()
plt.savefig('Test_PESQ.png', dpi=150)
plt.figure(3)
plt.plot(range(1,gan_epoch+1,interval_epoch),Test_VISQOL,'b',label='ValidVISQOL')
plt.xlim([1,gan_epoch])
plt.xlabel('GAN_epoch')
plt.ylabel('VISQOL')
plt.grid(True)
plt.show()
plt.savefig('Test_VISQOL.png', dpi=150)
# save the current enhancement model
save_path = os.path.join(pt_dir, 'chkpt_%d.pt' % gan_epoch)
torch.save({
'enhance-model': G.state_dict(),
'intel-model': D.state_dict(),
}, save_path)
print('Sample training data for discriminator training...')
D_paths = Generator_Train_paths[0:num_of_sampling]
Enhanced_name = []
G.eval()
# Generate samples for discriminator training
with torch.no_grad():
for path in D_paths:
S = path.split('/')
wave_name = S[-1]
clean_wav, sr = librosa.load(path, sr=fs)
assert sr==16000
noise_wav, _ = librosa.load(Train_Noise_path+wave_name, sr=fs)
clean_band, clean_mag, clean_phase = Sp_and_phase_Speech(clean_wav, power=p_power, Normalization=True)
noise_band, noise_mag, noise_phase = Sp_and_phase_Noise(noise_wav, power=p_power, Normalization=True)
clean_in = clean_band.reshape(1,clean_band.shape[0],-1)
clean_in = torch.from_numpy(clean_in).cuda()
noise_in = noise_band.reshape(1,noise_band.shape[0],-1)
noise_in = torch.from_numpy(noise_in).cuda()
# Energy normalization
mask = G(clean_in, noise_in)
clean_power = torch.pow(clean_in, inv_p)
beta_2 = torch.sum(clean_power) / torch.sum(mask*clean_power)
# beta_2 = torch.sum(clean_power, dim=2) / torch.sum(mask*clean_power, dim=2)
# beta_2 = beta_2.unsqueeze(2)
mask = mask * beta_2 # normed alpha2
mask = mask.detach().cpu().squeeze(0).numpy()
enh_wav = SP_to_wav(mask, clean_mag, clean_phase)
enhanced_name=output_path+"/For_discriminator_training/"+ wave_name[0:-4]+"@"+str(gan_epoch)+wave_name[-4:]
sf.write(enhanced_name, enh_wav, fs,'PCM_16')
Enhanced_name.append(enhanced_name)
G.train()
if TargetMetric=='siib&haspi&estoi':
# Calculate True SIIB score
train_SIIB = read_batch_SIIB(Train_Clean_path, Train_Noise_path, Enhanced_name)
train_HASPI = read_batch_HASPI(Train_Clean_path, Train_Noise_path, Enhanced_name)
train_ESTOI = read_batch_STOI(Train_Clean_path, Train_Noise_path, Enhanced_name)
train_PESQ = read_batch_PESQ(Train_Clean_path, Enhanced_name)
train_VISQOL = read_batch_VISQOL(Train_Clean_path, Enhanced_name)
train_SIIB = List_concat_5scores(train_SIIB, train_HASPI, train_ESTOI, train_PESQ, train_VISQOL) # SIIB, HASPI, ESTOI, PESQ, VISQOL
current_sampling_list=List_concat(train_SIIB, Enhanced_name) # This list is used to train discriminator.
# DRC_Enhanced_name = [Train_Enhan_path+'Train_'+S.split('/')[-1].split('_')[-1].split('@')[0]+'.wav' for S in Enhanced_name]
DRC_Enhanced_name = [Train_Enhan_path+S.split('/')[-1].split('@')[0]+'.wav' for S in Enhanced_name]
#pdb.set_trace()
train_SIIB_DRC = read_batch_SIIB_DRC(Train_Clean_path, Train_Noise_path, DRC_Enhanced_name)
train_HASPI_DRC = read_batch_HASPI_DRC(Train_Clean_path, Train_Noise_path, DRC_Enhanced_name)
train_ESTOI_DRC = read_batch_STOI_DRC(Train_Clean_path, Train_Noise_path, DRC_Enhanced_name)
train_PESQ_DRC = read_batch_PESQ_DRC(Train_Clean_path, DRC_Enhanced_name)
train_VISQOL_DRC = read_batch_VISQOL_DRC(Train_Clean_path, DRC_Enhanced_name)
train_SIIB_DRC = List_concat_5scores(train_SIIB_DRC, train_HASPI_DRC, train_ESTOI_DRC, train_PESQ_DRC, train_VISQOL_DRC) # SIIB, HASPI, ESTOI
Co_DRC_list = List_concat(train_SIIB_DRC, DRC_Enhanced_name)
print("Discriminator training...")
# Training for current list
Current_Discriminator_training_list = current_sampling_list+Co_DRC_list
random.shuffle(Current_Discriminator_training_list)
disloader = create_dataloader(Current_Discriminator_training_list, Train_Noise_path, Train_Clean_path, loader='D')
for x, x_qua, target, target_qua in tqdm(disloader):
x = x.cuda()
x_qua = x_qua.cuda()
target = target.cuda()
target_qua = target_qua.cuda()
score = D(x)
score_qua = D_Qua(x_qua)
loss = MSELoss(score, target)
optimizer_d.zero_grad()
loss.backward()
optimizer_d.step()
loss_qua = MSELoss(score_qua, target_qua)
optimizer_dqua.zero_grad()
loss_qua.backward()
optimizer_dqua.step()
step_d += 1
#if step_d % 1000 ==0:
# print('Step %d: Loss in D training is %.3f'%(step_d,loss.item()))
## Training for current list + Previous list (like replay buffer in RL, optional)
random.shuffle(Previous_Discriminator_training_list)
Total_Discriminator_training_list=Previous_Discriminator_training_list[0:len(Previous_Discriminator_training_list)//30]+Current_Discriminator_training_list # Discriminator_Train_list is the list used for pretraining.
random.shuffle(Total_Discriminator_training_list)
disloader_past = create_dataloader(Total_Discriminator_training_list, Train_Noise_path, Train_Clean_path, loader='D')
for x, x_qua, target, target_qua in tqdm(disloader_past):
x = x.cuda()
x_qua = x_qua.cuda()
target = target.cuda()
target_qua = target_qua.cuda()
score = D(x)
score_qua = D_Qua(x_qua)
loss = MSELoss(score, target)
optimizer_d.zero_grad()
loss.backward()
optimizer_d.step()
loss_qua = MSELoss(score_qua, target_qua)
optimizer_dqua.zero_grad()
loss_qua.backward()
optimizer_dqua.step()
step_d += 1
#if step_d % 1000 ==0:
# print('Step %d: Loss in D training is %.3f'%(step_d,loss.item()))
# Update the history list
Previous_Discriminator_training_list=Previous_Discriminator_training_list+Current_Discriminator_training_list
# Training current list again
for x, x_qua, target, target_qua in tqdm(disloader):
x = x.cuda()
x_qua = x_qua.cuda()
target = target.cuda()
target_qua = target_qua.cuda()
score = D(x)
score_qua = D_Qua(x_qua)
loss = MSELoss(score, target)
optimizer_d.zero_grad()
loss.backward()
optimizer_d.step()
loss_qua = MSELoss(score_qua, target_qua)
optimizer_dqua.zero_grad()
loss_qua.backward()
optimizer_dqua.step()
step_d += 1
#if step_d % 1000 ==0:
# print('Step %d: Loss in D training is %.3f'%(step_d,loss.item()))
shutil.rmtree(output_path+'/temp')
print('Epoch %d Finished' % gan_epoch)
print('Finished')