-
Notifications
You must be signed in to change notification settings - Fork 8
/
train_libri.py
executable file
·235 lines (185 loc) · 7.17 KB
/
train_libri.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
import argparse
from argparse import ArgumentParser
import logging
import time
import torch
from torch.optim import SGD, Adam
import torch.nn.functional as F
from dev.loaders import LibriSpeech4SpeakerRecognition, LibriSpeechSpeakers
from dev.models import RawAudioCNN, ALR, TDNN
from dev.utils import infinite_iter
from hparams import hp
import pdb, sys, os
import numpy as np
def _is_cuda_available():
return torch.cuda.is_available()
def _get_device():
return torch.device("cuda" if _is_cuda_available() else "cpu")
def noise_augmenter(inputs, labels, epsilon):
""" Data augmentation with additive white uniform noise"""
a = torch.rand([])
noise = torch.rand_like(inputs)
noise = noise.to(inputs.device)
noise = 2 * a * epsilon * noise - a * epsilon
noisy = inputs + noise
inputs = torch.cat([inputs, noisy])
labels = torch.cat([labels, labels])
return inputs, labels
device = _get_device()
def main(args):
logging.basicConfig(filename=args.log, level=logging.DEBUG)
if args.model_ckpt is None:
ckpt = f"model/libri_model_raw_audio_{time.strftime('%Y%m%d%H%M')}.pt"
else:
ckpt = args.model_ckpt
generator_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': args.num_workers
}
# Step 1: load data set
data_resolver = LibriSpeechSpeakers(hp.data_root, hp.data_subset)
train_data = LibriSpeech4SpeakerRecognition(
root=hp.data_root,
url=hp.data_subset,
train_speaker_ratio=hp.train_speaker_ratio,
train_utterance_ratio=hp.train_utterance_ratio,
subset="train",
project_fs=hp.sr,
wav_length=args.wav_length,
)
train_generator = torch.utils.data.DataLoader(train_data, **generator_params)
if args.model_type=='cnn':
model = RawAudioCNN(num_class=data_resolver.get_num_speakers())
elif args.model_type=='tdnn':
model = TDNN(data_resolver.get_num_speakers())
else:
logging.error('Please provide a valid model architecture type!')
sys.exit(-1)
print(model)
if _is_cuda_available():
model.to(device)
logging.info(device)
alr = ALR()
criterion = torch.nn.CrossEntropyLoss()
if args.optimizer=='sgd':
optimizer = torch.optim.SGD(model.parameters(), lr=2e-3, momentum=0.9)
elif args.optimizer=='adam':
print()
print('Using Adam optimizer\n')
optimizer = Adam(model.parameters(), lr=1e-3, betas=(.5, .999))
# Step 3: train
model.train()
batch_idx = 0
loss_epoch = []
acc_epoch = []
for batch_data in infinite_iter(train_generator):
batch_idx += 1
inputs, labels = (x.to(device) for x in batch_data)
model.train()
if args.epsilon > 0:
inputs, labels = noise_augmenter(inputs, labels, args.epsilon)
real_feature = model.encode(inputs)
outputs = model.predict_from_embeddings(real_feature)
class_loss = criterion(outputs, labels)
if args.alr_weight > 0:
input_adv = alr.get_adversarial_perturbations(model, inputs, labels)
adv_feature = model.encode(input_adv)
output_adv = model.predict_from_embeddings(adv_feature)
alr_outputs = alr.get_alp(inputs, input_adv, outputs, output_adv, labels)
alr_loss = alr_outputs[0].mean()
loss = class_loss + args.alr_weight * alr_loss
if args.cr_weight > 0:
cr_loss = F.mse_loss(real_feature, adv_feature)
loss += cr_loss
else:
loss = class_loss
# Model computations
optimizer.zero_grad()
loss.backward()
optimizer.step()
# training accuracy
acc_ = np.mean((torch.argmax(outputs, dim=1) == labels).detach().cpu().numpy())
loss_epoch.append(loss.item())
acc_epoch.append(acc_)
message = f"It [{batch_idx}] train-loss: {loss.item():.4f} \ttrain-acc (batch): {acc_:.4f}"
if args.alr_weight > 0:
message += f"\talr={alr_loss.item():.4e} "
if args.cr_weight > 0:
message += f"\tcr={cr_loss.item():.4e} "
print(message, end="\r")
logging.info(message)
# Checkpointing
if batch_idx % args.save_every == 0:
torch.save(model, ckpt + ".tmp")
torch.save(optimizer, ckpt + ".optimizer.tmp")
print()
# Termination
if batch_idx > args.n_iters:
break
# Termination -- if n_epochs are provided
if args.n_epochs is not None:
done_ = batch_idx//len(train_generator)
if batch_idx%len(train_generator)==0:
msg_ = f"Epoch {done_}: loss = {np.mean(loss_epoch)} acc = {np.mean(acc_epoch)}"
print(msg_)
logging.info(msg_)
loss_epoch=[]
acc_epoch=[]
if done_ >= args.n_epochs:
msg_ = "Finishing training based on n_epochs provided..."
print(msg_)
logging.info(msg_)
break
logging.info("Finished Training")
torch.save(model, ckpt)
def parse_args():
parser = ArgumentParser("Speaker Classification model on LibriSpeech dataset", \
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
"-m", "--model_ckpt", type=str, default=None, help="Model checkpoint")
parser.add_argument(
"-g", "--log", type=str, default="train.log", help="Experiment log")
parser.add_argument(
"-mt", "--model_type", type=str, default='cnn', help="Model type: cnn or tdnn")
parser.add_argument(
"-l", "--wav_length", type=int, default=80000,
help="Max length of waveform in a batch")
parser.add_argument(
"-n", "--n_iters", type=int, default=500000,
help="Number of iterations for training"
)
parser.add_argument(
"-ne", "--n_epochs", type=int, default=None,
help="Number of epochs for training. Optional. Ignored if not provided."
)
parser.add_argument(
"-s", "--save_every", type=int, default=10000, help="Save after this number of gradient updates"
)
parser.add_argument(
"-e", "--epsilon", type=float, default=0,
help="Noise magnitude in data augmentation; set it to 0 to disable augmentation")
parser.add_argument(
"-w", "--alr_weight", type=float, default=0,
help="Weight of the adversarial Lipschitz regularizer"
)
parser.add_argument(
"-c", "--cr_weight", type=float, default=0,
help="Weight of consistency regularizer"
)
parser.add_argument(
"-opt", "--optimizer", type=str, default='adam',
help="Optimizer: sgd, adam")
parser.add_argument(
"-b", "--batch_size", type=int, default=128,
help="Batch size")
parser.add_argument(
"-nw", "--num_workers", type=int, default=8,
help="Number of workers related to pytorch data loader")
args = parser.parse_args()
#clean log file
with open(args.log, "w") as f:
f.write("")
return args
if __name__ == "__main__":
main(parse_args())