-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_enhancement_model.py
194 lines (147 loc) · 6.9 KB
/
train_enhancement_model.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
import argparse
import yaml
import os
import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from src.data_utils import EnhancementDataset
from src.loss_utils import snr_loss
from models import EnhancementNet
def _parse_args():
parser = argparse.ArgumentParser(description='enh')
parser.add_argument('--batch-size', type=int, default=6,
help='input batch size for training')
parser.add_argument('--epochs', type=int, default=100,
help='number of epochs to train')
parser.add_argument('--cuda', action='store_true', default=True,
help='enables CUDA training')
parser.add_argument('--lr', type=float, default=1e-3,
help='learning rate')
parser.add_argument('--config', default='./configs/config_enhancement.yaml', type=str,
help='model config')
parser.add_argument('--seed', type=int, default=2023028,
help='random seed')
parser.add_argument('--training-file-path', default='./', type=str,
help='training file path')
parser.add_argument('--validation-file-path', default='./', type=str,
help='validation file path')
parser.add_argument('--checkpoint-path', type=str, default='./',
help='path to save the model')
args, _ = parser.parse_known_args()
return args
def _load_config(path):
with open(path, 'r') as ymlfile:
config = yaml.safe_load(ymlfile)
return config
def _save_checkpoint(filepath, obj):
print("Saving checkpoint to {}".format(filepath))
torch.save(obj, filepath)
print("Complete.")
def train(train_loader, validation_loader, model, optimizer, scheduler, summary_writer, args):
for epoch in range(1, args.epochs + 1):
model.train()
train_loss = 0.
for batch_idx, data in enumerate(train_loader):
batch_s1 = Variable(data[0]).contiguous()
batch_s2 = Variable(data[1]).contiguous()
batch_est_s1 = Variable(data[2]).contiguous()
batch_est_s2 = Variable(data[3]).contiguous()
batch_noise = Variable(data[4]).contiguous()
batch_mix = batch_s1 + batch_s2 + batch_noise
if args.cuda:
batch_mix = batch_mix.cuda()
batch_s1 = batch_s1.cuda()
batch_s2 = batch_s2.cuda()
batch_est_s1 = batch_est_s1.cuda()
batch_est_s2 = batch_est_s2.cuda()
batch_mix = torch.cat([batch_mix, batch_mix], dim=0)
batch_est = torch.cat([batch_est_s1, batch_est_s2], dim=0)
batch_clean = torch.cat([batch_s1, batch_s2], dim=0)
batch_output = model(batch_mix, batch_est)
optimizer.zero_grad()
loss = torch.mean(snr_loss(batch_clean.view(batch_clean.size(0)*2, batch_clean.size(2)),
batch_output.view(batch_clean.size(0)*2, batch_clean.size(2)))
)
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 5.)
train_loss += loss.data.item()
optimizer.step()
train_loss /= (batch_idx+1)
summary_writer.add_scalar("training/train_loss", train_loss, epoch)
checkpoint_path = "{}/epoch_{:03d}".format(args.checkpoint_path, epoch)
_save_checkpoint(
checkpoint_path,
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict(),
})
model.eval()
validation_loss = 0.
for batch_idx, data in enumerate(validation_loader):
batch_s1 = Variable(data[0]).contiguous()
batch_s2 = Variable(data[1]).contiguous()
batch_est_s1 = Variable(data[2]).contiguous()
batch_est_s2 = Variable(data[3]).contiguous()
batch_noise = Variable(data[4]).contiguous()
batch_mix = batch_s1 + batch_s2 + batch_noise
if args.cuda:
batch_mix = batch_mix.cuda()
batch_s1 = batch_s1.cuda()
batch_s2 = batch_s2.cuda()
batch_est_s1 = batch_est_s1.cuda()
batch_est_s2 = batch_est_s2.cuda()
batch_mix = torch.cat([batch_mix, batch_mix], dim=0)
batch_est = torch.cat([batch_est_s1, batch_est_s2], dim=0)
batch_clean = torch.cat([batch_s1, batch_s2], dim=0)
with torch.no_grad():
batch_output = model(batch_mix, batch_est)
loss = torch.mean(snr_loss(batch_clean.view(batch_clean.size(0)*2, batch_clean.size(2)),
batch_output.view(batch_clean.size(0)*2, batch_clean.size(2)))
)
validation_loss += loss.data.item()
validation_loss /= (batch_idx+1)
summary_writer.add_scalar("training/val_loss", validation_loss, epoch)
if epoch % 2 == 0:
scheduler.step()
def main(args):
args.cuda = args.cuda and torch.cuda.is_available()
if args.cuda:
torch.cuda.manual_seed(args.seed)
kwargs = {'num_workers': 4, 'pin_memory': True}
else:
kwargs = {}
random.seed(args.seed)
torch.manual_seed(args.seed)
train_loader = DataLoader(EnhancementDataset(args.training_file_path),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
validation_loader = DataLoader(EnhancementDataset(args.validation_file_path),
batch_size=args.batch_size,
shuffle=False,
**kwargs)
config = _load_config(args.config)
model = EnhancementNet(
enc_dim = config['enc_dim'],
feature_dim = config['feature_dim'],
hidden_dim = config['hidden_dim'],
enc_win = config['enc_win'],
enc_stride = config['enc_stride'],
num_block = config['num_block'],
num_layer = config['num_layer'],
kernel_size = config['kernel_size'],
num_spk = config['num_spk'],
)
if args.cuda:
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.98)
sw = SummaryWriter(os.path.join(args.checkpoint_path, 'logs'))
train(train_loader, validation_loader, model, optimizer, scheduler, sw, args)
if __name__ == '__main__':
main(_parse_args())