-
Notifications
You must be signed in to change notification settings - Fork 12
/
Network_Training.py
157 lines (130 loc) · 6.09 KB
/
Network_Training.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
# -*- coding: utf-8 -*-
"""
Created on Fri Apr 11 13:32:17 2022
@author: Zhongshu.Hou & Qinwen.Hu
network training
"""
import os
import torch
import torch.optim as optim
# from torch.utils.tensorboard import SummaryWriter
import numpy as np
from torch.optim.lr_scheduler import StepLR
from tqdm import tqdm
import soundfile as sf
from Dataloader import Dataset, collate_fn
from Modules import DPModel
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
torch.set_default_tensor_type(torch.FloatTensor)
from signal_processing import iSTFT_module
WINDOW = torch.sqrt(torch.hann_window(1200,device=device) + 1e-8)
#------------------------warm up strategy------------------------
class NoamOpt:
"Optim wrapper that implements rate."
def __init__(self, model_size, factor, warmup, optimizer):
self.optimizer = optimizer
self._step = 0
self.warmup = warmup
self.factor = factor
self.model_size = model_size
self._rate = 0
def step(self):
"Update parameters and rate"
self._step += 1
rate = self.rate()
for p in self.optimizer.param_groups:
p['lr'] = rate
self._rate = rate
self.optimizer.step()
def rate(self, step = None):
"Implement `lrate` above"
if step is None:
step = self._step
return self.factor * \
(self.model_size ** (-0.5) *
min(step ** (-0.5), step * self.warmup ** (-1.5)))
#------------------------start training------------------------
def train(end_epoch = 100):
'''Loss functions'''
def Loss(y_pred, y_true):
snr = torch.div(torch.mean(torch.square(y_pred - y_true), dim=1, keepdim=True),(torch.mean(torch.square(y_true), dim=1, keepdim=True) + 1e-7))
snr_loss = 10 * torch.log10(snr + 1e-7)
pred_stft = torch.stft(y_pred,1200,600,win_length=1200,window=WINDOW,center=True)
true_stft = torch.stft(y_true,1200,600,win_length=1200,window=WINDOW,center=True)
pred_stft_real, pred_stft_imag = pred_stft[:,:,:,0], pred_stft[:,:,:,1]
true_stft_real, true_stft_imag = true_stft[:,:,:,0], true_stft[:,:,:,1]
pred_mag = torch.sqrt(pred_stft_real**2 + pred_stft_imag**2 + 1e-12)
true_mag = torch.sqrt(true_stft_real**2 + true_stft_imag**2 + 1e-12)
pred_real_c = pred_stft_real / (pred_mag**(2/3))
pred_imag_c = pred_stft_imag / (pred_mag**(2/3))
true_real_c = true_stft_real / (true_mag**(2/3))
true_imag_c = true_stft_imag / (true_mag**(2/3))
real_loss = torch.mean((pred_real_c - true_real_c)**2)
imag_loss = torch.mean((pred_imag_c - true_imag_c)**2)
mag_loss = torch.mean((pred_mag**(1/3)-true_mag**(1/3))**2)
return real_loss + imag_loss + mag_loss, snr_loss
'''model'''
model = DPModel(model_type='DPARN', device=device)
model.init_load()
model = model.to(device)
''' train from checkpoints'''
# checkpoint_DPARN = torch.load('',map_location=device)
# model.load_state_dict(checkpoint_DPARN['state_dict'])
'''optimizer & lr_scheduler'''
optimizer = NoamOpt(model_size=model.process_model.numUnits, factor=1., warmup=40000,
optimizer=torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))
'''load train data, set the configuration according to your device'''
dataset = Dataset(length_in_seconds=8, random_start_point=True, train=True)
data_loader = torch.utils.data.DataLoader(dataset, batch_size=4, shuffle=True, num_workers=8, drop_last=True)
os.makedirs('./checkpoints_dparn/', exist_ok=True)
'''start train'''
for epoch in range(end_epoch):
train_loss = []
asnr_loss = []
model.train()
dataset.train = True
dataset.random_start_point = True
'''train'''
print('epoch %s--training' %(epoch))
for i, data in enumerate(tqdm(data_loader)):
noisy, clean = data
noisy = noisy.to(device)
clean = clean.to(device)
optimizer.optimizer.zero_grad()
noisy_stft = torch.stft(noisy,1200,600,win_length=1200,window=WINDOW,center=True)
enh_stft = model(noisy_stft)
enh_s = iSTFT_module(n_fft=1200, hop_length=600, win_length=1200,window=WINDOW,center = True,length = noisy.shape[-1])(enh_stft)
stft_loss, snr_loss = Loss(enh_s, clean)
stft_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=3)
optimizer.step()
train_loss.append(stft_loss.cpu().detach().numpy())
train_loss = np.mean(train_loss)
'''eval'''
valid_loss = []
model.eval()
print('epoch %s--validating' %(epoch))
dataset.train = False
dataset.random_start_point = False
with torch.no_grad():
for i, data in enumerate(tqdm(data_loader)):
noisy, clean = data
noisy = noisy.to(device)
clean = clean.to(device)
noisy_stft = torch.stft(noisy,1200,600,win_length=1200,window=WINDOW,center=True)
enh_stft = model(noisy_stft)
enh_s = iSTFT_module(n_fft=1200, hop_length=600, win_length=1200,window=WINDOW,center = True,length = noisy.shape[-1])(enh_stft)
stft_loss, snr_loss = Loss(enh_s, clean)
valid_loss.append(stft_loss.cpu().detach().numpy())
asnr_loss.append(snr_loss.cpu().detach().numpy())
valid_loss = np.mean(valid_loss)
asnr_loss = np.mean(asnr_loss)
print('train loss: %s, valid loss %s, snr loss: %s' %(train_loss, valid_loss, asnr_loss))
print('current step:{}, current lr:{}'.format(optimizer._step, optimizer._rate))
torch.save(
{'epoch': epoch,
'state_dict': model.state_dict(),
'optimizer': optimizer.optimizer.state_dict()},
'./checkpoints_dparn/model_epoch_{}_train_{:.5f}_valid_{:.5f}_snr_{:.5f}.pth'.format(epoch, train_loss, valid_loss, asnr_loss))
if __name__ == '__main__':
train(end_epoch=200)