-
Notifications
You must be signed in to change notification settings - Fork 33
/
Copy pathtrain_dul_reg.py
executable file
·211 lines (161 loc) · 7.95 KB
/
train_dul_reg.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import os
import sys
import time
import torch
import random
import numpy as np
import torchvision
import torch.nn as nn
import torch.optim as optim
from sklearn import metrics
import torch.nn.functional as F
from torch.utils.data import DataLoader
import model as mlib
import dataset as dlib
from config import reg_args
torch.backends.cudnn.bencmark = True
# os.environ["CUDA_VISIBLE_DEVICES"] = "6,7" # TODO
from IPython import embed
class DulRegTrainer(mlib.Faster1v1):
def __init__(self, args):
mlib.Faster1v1.__init__(self, args)
self.args = args
self.model = dict()
self.data = dict()
self.result = dict()
self.softmax = torch.nn.Softmax(dim=1)
self.use_gpu = args.use_gpu and torch.cuda.is_available()
def _report_settings(self):
''' Report the settings '''
str = '-' * 16
print('%sEnvironment Versions%s' % (str, str))
print("- Python : {}".format(sys.version.strip().split('|')[0]))
print("- PyTorch : {}".format(torch.__version__))
print("- TorchVison: {}".format(torchvision.__version__))
print("- USE_GPU : {}".format(self.use_gpu))
print("- IS_DEBUG : {}".format(self.args.is_debug))
print('-' * 52)
def _model_loader(self):
self.model['backbone'] = mlib.dulres_zoo(self.args.backbone, \
drop_ratio=self.args.drop_ratio, \
use_se=self.args.use_se,\
used_as='backbone') # ResBlock
self.model['reg_head'] = mlib.RegHead(drop_ratio=self.args.drop_ratio)
self.model['criterion'] = mlib.RegLoss(feat_dim=self.args.in_feats, classnum=self.args.classnum)
self.model['optimizer'] = torch.optim.SGD(
[{'params': self.model['reg_head'].parameters()}],
lr=self.args.base_lr,
weight_decay=self.args.weight_decay,
momentum=0.9,
nesterov=True)
self.model['scheduler'] = torch.optim.lr_scheduler.MultiStepLR(
self.model['optimizer'], milestones=self.args.lr_adjust, gamma=self.args.gamma)
if self.use_gpu:
self.model['backbone'] = self.model['backbone'].cuda()
self.model['criterion'] = self.model['criterion'].cuda()
if self.use_gpu and len(self.args.gpu_ids) > 1:
self.model['backbone'] = torch.nn.DataParallel(self.model['backbone'], device_ids=self.args.gpu_ids)
print('Parallel mode was going ...')
elif self.use_gpu:
print('Single-gpu mode was going ...')
else:
print('CPU mode was going ...')
if len(self.args.resume) > 2:
checkpoint = torch.load(self.args.resume, map_location=lambda storage, loc: storage)
self.args.start_epoch = checkpoint['epoch']
self.model['backbone'].load_state_dict(checkpoint['backbone'])
self.model['criterion'].fetch_center_from_fc_layer(checkpoint['fc_layer'])
self.model['backbone'].eval() # backbone was fixed
print('Resuming the train process at %3d epoches ...' % self.args.start_epoch)
print('Model loading was finished ...')
@staticmethod
def collate_fn_1v1(batch):
imgs, pairs_info = [], []
for unit in batch:
pairs_info.append([unit['name1'], unit['name2'], unit['label']])
imgs.append(torch.cat((unit['face1'], unit['face2']), dim=0))
return (torch.stack(imgs, dim=0), np.array(pairs_info))
def _data_loader(self):
self.data['train'] = DataLoader(
dlib.DataBase(self.args),
batch_size=self.args.batch_size, \
shuffle=True,
num_workers=self.args.workers)
self.data['lfw'] = DataLoader(
dlib.VerifyBase(self.args, benchmark = 'lfw'),
batch_size=self.args.batch_size // 2, \
num_workers=self.args.workers,
drop_last=False,
collate_fn=self.collate_fn_1v1)
print('Data loading was finished ...')
def _train_one_epoch(self, epoch = 0):
self.model['backbone'].train()
self.model['fc_layer'].train()
loss_recorder, batch_acc = [], []
for idx, (imgs, gtys, _) in enumerate(self.data['train']):
imgs.requires_grad = False
gty.requires_grad = False
if self.use_gpu:
imgs = imgs.cuda()
gtys = gtys.cuda()
outfeat, _, _ = self.model['backbone'](imgs) # CORE
mu, logvar = self.model['reg_head'](outfeat) # CORE
loss = self.model['criterion'](mu, logvar, gty) # CORE
self.model['optimizer'].zero_grad()
loss.backward()
self.model['optimizer'].step()
loss_recorder.append(loss.item())
if (idx + 1) % self.args.print_freq == 0:
print('epoch : %2d|%2d, iter : %4d|%4d, loss : %.4f' % \
(epoch, self.args.end_epoch, idx+1, len(self.data['train']), np.mean(loss_recorder)))
train_loss = np.mean(loss_recorder)
print('train_loss : %.4f' % train_loss)
return train_loss
def _save_weights(self, testinfo = {}):
''' save the weights during the process of training '''
if not os.path.exists(self.args.save_to):
os.mkdir(self.args.save_to)
freq_flag = self.result['epoch'] % self.args.save_freq == 0
sota_flag = self.result['sota_acc'] < testinfo['test_acc']
save_name = '%s/epoch_%02d_lfw_opt_thresh_%.4f_racc_%.4f.pth' % \
(self.args.save_to, self.result['epoch'], testinfo['opt_thresh'], testinfo['test_acc'])
if sota_flag:
save_name = '%s/sota.pth' % self.args.save_to
self.result['opt_thresh'] = testinfo['opt_thresh']
self.result['sota_acc'] = testinfo['test_acc']
print('%s Yahoo, SOTA model was updated %s' % ('*'*16, '*'*16))
if sota_flag or freq_flag:
torch.save({
'epoch' : self.result['epoch'],
'backbone': self.model['backbone'].state_dict(),
'fc_layer': self.model['fc_layer'].state_dict(),
'thresh' : testinfo['opt_thresh'],
'sota_acc': testinfo['test_acc']}, save_name)
if sota_flag and freq_flag:
normal_name = '%s/epoch_%02d-lfw_opt_thresh_%.4f-racc_%.4f.pth' % \
(self.args.save_to, self.result['epoch'], testinfo['opt_thresh'], testinfo['test_acc'])
shutil.copy(save_name, normal_name)
def _dul_training(self):
self.result['opt_thresh'] = -1.0
self.result['sota_acc'] = 0
for epoch in range(self.args.start_epoch, self.args.end_epoch + 1):
start_time = time.time()
self.result['epoch'] = epoch
train_loss = self._train_one_epoch(epoch)
self.model['scheduler'].step()
eval_info = self._evaluate_one_epoch(loader='lfw')
end_time = time.time()
print('Single epoch cost time : %.2f mins' % ((end_time - start_time)/60))
self._save_weights(eval_info)
if self.args.is_debug:
break
def train_runner(self):
self._report_settings()
self._model_loader()
self._data_loader()
self._dul_training()
if __name__ == "__main__":
dul_reg = DulRegTrainer(reg_args())
dul_reg.train_runner()