forked from yistLin/dvector
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
executable file
·229 lines (191 loc) · 8.24 KB
/
train.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Train d-vector."""
# original codebase: https://github.com/yistLin/dvector
# modified and re-distributed by Zifeng Zhao @ Peking University
# 2022.03
import json
from argparse import ArgumentParser
from collections import deque
from datetime import datetime
from itertools import count
from multiprocessing import cpu_count
from pathlib import Path
import torch
from torch.optim import SGD
from torch.optim.lr_scheduler import StepLR
from torch.utils.data import random_split
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
import data
import modules
import os
import equal_error_rate
import numpy as np
def train(
data_dir,
save_dir,
n_speakers,
n_utterances,
seg_len,
save_every,
valid_every,
decay_every,
batch_per_valid,
n_workers,
comment,
gpu,
max_epoch,
test_dir,
test_txt,
):
"""Train a d-vector network."""
# setup job name
start_time = datetime.now().strftime("%Y-%m-%d_%H:%M:%S")
job_name = f"{start_time}_{comment}" if comment is not None else start_time
# setup checkpoint and log dirs
print(f'# 实验保存路径 {save_dir}')
checkpoints_path = Path(save_dir) / "checkpoints"
checkpoints_path.mkdir(parents=True, exist_ok=True)
writer = SummaryWriter(Path(save_dir) / "logs")
# create data loader, iterator
print(f'# 读取json {Path(data_dir, "metadata.json")}')
with open(Path(data_dir, "metadata.json"), "r") as f:
metadata = json.load(f)
print(f'# 加载data.GE2EDataset...')
dataset = data.GE2EDataset(data_dir, metadata["speakers"], n_utterances, seg_len)
print(f'# 数据集划分 spk_train/spk_valid={len(dataset) - n_speakers}/{n_speakers}...')
trainset, validset = random_split(dataset, [len(dataset) - n_speakers, n_speakers])
# 无限重复采样
train_loader = data.InfiniteDataLoader(
trainset,
batch_size=n_speakers,
num_workers=n_workers,
collate_fn=data.collate_batch,
drop_last=True,
)
valid_loader = data.InfiniteDataLoader(
validset,
batch_size=n_speakers,
num_workers=n_workers,
collate_fn=data.collate_batch,
drop_last=True,
)
train_iter = data.infinite_iterator(train_loader)
valid_iter = data.infinite_iterator(valid_loader)
# display training infos
assert len(trainset) >= n_speakers
assert len(validset) >= n_speakers
print(f"[INFO] Use {len(trainset)} speakers for training.")
print(f"[INFO] Use {len(validset)} speakers for validation.")
print(f'# 使用GPU: {gpu}')
os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu)
# build network and training tools
print(f'# 加载模型')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dvector = modules.AttentivePooledLSTMDvector(
dim_input=metadata["n_mels"],
seg_len=seg_len,
).to(device)
dvector = torch.jit.script(dvector)
criterion = modules.GE2ELoss().to(device)
optimizer = SGD(list(dvector.parameters()) + list(criterion.parameters()), lr=0.01)
scheduler = StepLR(optimizer, step_size=decay_every, gamma=0.95)
print('')
print('### START TRAINING ###')
batch_size = n_speakers * n_utterances
voxceleb1_train_utt = 68224
steps_per_epoch = voxceleb1_train_utt // batch_size
# record training infos
pbar = tqdm(total=steps_per_epoch, ncols=0, desc="Train")
running_train_loss, running_grad_norm = deque(maxlen=steps_per_epoch), deque(maxlen=steps_per_epoch)
running_valid_loss = deque(maxlen=batch_per_valid)
# start training
for epoch in range(max_epoch):
tqdm.write('')
tqdm.write(f'# epoch {epoch} - train...')
pbar.reset()
for step in count(start=1):
if step > steps_per_epoch:
break
batch = next(train_iter).to(device)
embds = dvector(batch).view(n_speakers, n_utterances, -1)
loss = criterion(embds)
optimizer.zero_grad()
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(
list(dvector.parameters()) + list(criterion.parameters()),
max_norm=3,
norm_type=2.0,
)
dvector.embedding.weight.grad *= 0.5
dvector.embedding.bias.grad *= 0.5
criterion.w.grad *= 0.01
criterion.b.grad *= 0.01
optimizer.step()
running_train_loss.append(loss.item())
running_grad_norm.append(grad_norm.item())
avg_train_loss = sum(running_train_loss) / len(running_train_loss)
avg_grad_norm = sum(running_grad_norm) / len(running_grad_norm)
pbar.update(1)
pbar.set_postfix(loss=avg_train_loss, grad_norm=avg_grad_norm)
#if step % valid_every == 0:
if True:
# 交叉验证
for _ in range(batch_per_valid):
batch = next(valid_iter).to(device)
with torch.no_grad():
embd = dvector(batch).view(n_speakers, n_utterances, -1)
loss = criterion(embd)
running_valid_loss.append(loss.item())
avg_valid_loss = sum(running_valid_loss) / len(running_valid_loss)
#tqdm.write(f"Valid: epoch={epoch}, error={avg_valid_loss:.1f}")
writer.add_scalar("train/loss", avg_train_loss, epoch)
writer.add_scalar("valid/error", avg_valid_loss, epoch)
writer.add_scalar("learning_rate", optimizer.state_dict()['param_groups'][0]['lr'], epoch)
tqdm.write(f'# loss={avg_train_loss}, error={avg_valid_loss}, lr='+str(optimizer.state_dict()['param_groups'][0]['lr']))
#if step % save_every == 0:
if epoch % 20 == 0:
ckpt_path = checkpoints_path / f"dvector-epoch{epoch}.pt"
dvector.cpu()
dvector.save(str(ckpt_path))
dvector.to(device)
eer, thr = equal_error_rate.get_eer(test_dir=test_dir,
test_txt=test_txt,
wav2mel_path=os.path.join(data_dir, "wav2mel.pt"),
checkpoint_path=str(ckpt_path),
)
tqdm.write(f'# testEER={np.round(eer*100,2)}%, threshold={thr}')
writer.add_scalar("test/EER", eer, epoch)
scheduler.step()
ckpt_path = checkpoints_path / f"dvector-epoch{epoch}.pt"
dvector.cpu()
dvector.save(str(ckpt_path))
dvector.to(device)
print('')
print('### TRAINING COMPLETED ###')
from IPython import embed
embed()
print('\ndone.')
if __name__ == "__main__":
PARSER = ArgumentParser()
PARSER.add_argument("--data_dir", type=str) # 数据集(Mel谱)路径
PARSER.add_argument("--save_dir", type=str) # 实验保存路径
# batch_size = n_speakers * n_utterances = 64 * 4 = 256
PARSER.add_argument("-n", "--n_speakers", type=int, default=64) # 一个batch中speaker数目
PARSER.add_argument("-m", "--n_utterances", type=int, default=4) # 轮到的说话人每人每次随机抽取n_utterance条进入batch 语料不足n_utterance条的说话人将被滤除
PARSER.add_argument("--seg_len", type=int, default=160) # 长度不足seg_len(帧)的音频将被滤除
PARSER.add_argument("--save_every", type=int, default=10000)
PARSER.add_argument("--valid_every", type=int, default=1000)
PARSER.add_argument("--decay_every", type=int, default=10)
PARSER.add_argument("--batch_per_valid", type=int, default=10)
PARSER.add_argument("--n_workers", type=int, default=cpu_count()) # 线程
PARSER.add_argument("--comment", type=str)
# 新增
PARSER.add_argument("--gpu", type=str, default="0") # 所用GPU
PARSER.add_argument("--max_epoch", type=int, default=300) # epoch
PARSER.add_argument("--test_dir", type=str) # 测试集路径
PARSER.add_argument("--test_txt", type=str) # 测试列表路径
print('')
print('# 解析parser运行参数')
train(**vars(PARSER.parse_args()))