-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain_adversary.py
318 lines (239 loc) · 13.4 KB
/
train_adversary.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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
@author: liuyaqi
"""
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import discriminator as dis
import detector as det
import utils
import dmac_vgg_skip as dmac_vgg
import time
def print_network(net):
num_params = 0
for param in net.parameters():
num_params += param.numel()
print(net)
print('Total number of parameters: %d' % num_params)
class SoftmaxMask(nn.Module):
def __init__(self):
super(SoftmaxMask,self).__init__()
self.softmax = nn.Softmax2d()
def forward(self,x):
x = self.softmax(x)
x = torch.chunk(x,2,dim=1)
return x[0],x[1]
def loc_loss_calc(out, label, gpu):
"""
This function returns cross entropy loss for semantic segmentation
"""
# out shape batch_size x channels x h x w
# label shape batch_size x h x w
# print 'loss computation begin!'
# label = torch.from_numpy(label).long()
m = nn.LogSoftmax().cuda(gpu)
criterion = nn.NLLLoss2d().cuda(gpu)
out = m(out)
return criterion(out,label)
def snapshot(model, prefix, epoch, iter):
print 'taking snapshot ...'
torch.save(model.state_dict(), prefix + str(epoch) + '_' + str(iter) + '.pth')
low_prob = 0.1
up_prob = 0.9
def generate_adap_gt(gt, mask, gpu):
s0,s1,s2,s3 = gt.size()
gt1 = gt
gt0 = torch.ones((s0,s1,s2,s3)).cuda(gpu)-gt1
gt = torch.clamp(gt,min=low_prob,max=up_prob)
gt = gt + nn.ReLU()(torch.mul(mask-gt,gt1))
gt = gt - nn.ReLU()(torch.mul(gt-mask,gt0))
return gt
def norm_img(image):
img_temp = image
img_temp[:,0,:,:] = image[:,0,:,:] + 104.008
img_temp[:,1,:,:] = image[:,1,:,:] + 116.669
img_temp[:,2,:,:] = image[:,2,:,:] + 122.675
img_temp = torch.div(img_temp,255.00)
img_temp = torch.div(img_temp - 0.5, 0.5)
return img_temp
class AdversaryLearning(object):
def __init__(self, args):
self.epoch = args.epoch
self.pair_list = args.pair_list
self.epoch_len = args.epoch_len
self.batch_size = args.batch_size
self.gpu = args.gpu
self.loc_update_stride = args.loc_update_stride
self.snapshot_stride = args.snapshot_stride
self.lambda_loc = args.lambda_loc
self.lambda_det = args.lambda_det
self.lambda_dis = args.lambda_dis
self.start_epoch_idx = args.start_epoch_idx
self.start_iter_idx = args.start_iter_idx
self.snapshot_prefix_loc = args.snapshot_prefix_loc
self.snapshot_prefix_dis = args.snapshot_prefix_dis
self.snapshot_prefix_det = args.snapshot_prefix_det
self.data_path = args.data_path
self.input_scale = args.input_scale
self.loc = dmac_vgg.DMAC_VGG(args.nolabel, self.gpu, self.input_scale)
self.dis = dis.Discriminator(8)
self.det = det.Detector(8)
self.adapt_gt_flag = args.adapt_gt_flag
self.norm_im_flag = args.norm_im_flag
if args.loc_pretrained:
loc_saved_state_dict = torch.load(args.loc_pretrain_model)
self.loc.load_state_dict(loc_saved_state_dict)
if args.dis_pretrained:
dis_saved_state_dict = torch.load(args.dis_pretrain_model)
self.dis.load_state_dict(dis_saved_state_dict)
if args.det_pretrained:
det_saved_state_dict = torch.load(args.det_pretrain_model)
self.det.load_state_dict(det_saved_state_dict)
self.soft_max = SoftmaxMask()
self.CE_Loss = nn.CrossEntropyLoss().cuda(self.gpu)
self.loss_type = args.loss_type
self.BCElog_Loss = nn.BCEWithLogitsLoss().cuda(self.gpu)
self.loss_type = args.loss_type
self.loc.cuda(self.gpu)
self.dis.cuda(self.gpu)
self.det.cuda(self.gpu)
self.soft_max.cuda(self.gpu)
self.loc_optimizer = optim.Adam(self.loc.parameters(),lr=args.lr_loc,betas=(args.beta1,args.beta2))
self.dis_optimizer = optim.Adam(filter(lambda p: p.requires_grad, self.dis.parameters()),lr=args.lr_dis,betas=(args.beta1,args.beta2))
self.det_optimizer = optim.Adam(self.det.parameters(),lr=args.lr_det,betas=(args.beta1,args.beta2))
print('---------- Networks architecture -------------')
print_network(self.loc)
print_network(self.dis)
print_network(self.det)
print('-----------------------------------------------')
def train(self):
self.train_hist = {}
self.train_hist['per_epoch_time'] = []
self.train_hist['total_time'] = []
self.loc.train()
self.dis.train()
self.det.train()
print('training start!!')
start_time = time.time()
for epoch in range(self.start_epoch_idx, self.epoch):
epoch_start_time = time.time()
data_gen = utils.chunker(self.pair_list, self.batch_size)
if epoch == self.start_epoch_idx:
start_iter_idx = self.start_iter_idx
else:
start_iter_idx = 0
for iter in range(start_iter_idx, self.epoch_len):
if iter == self.epoch_len // self.batch_size:
break
# read images
chunk = data_gen.next()
images1, images2, labels, gt1, gt2 = utils.get_data_from_chunk(self.data_path,chunk,self.input_scale)
images1 = images1.cuda(self.gpu)
images2 = images2.cuda(self.gpu)
if self.norm_im_flag:
images1_d = norm_img(images1)
images2_d = norm_img(images2)
else:
images1_d = images1
images2_d = images2
# gt masks variable
gt1_ = torch.squeeze(gt1,dim=1).long()
gt2_ = torch.squeeze(gt2,dim=1).long()
gt1_ = gt1_.cuda(self.gpu)
gt2_ = gt2_.cuda(self.gpu)
gt1 = gt1.cuda(self.gpu)
gt2 = gt2.cuda(self.gpu)
# dis labels variable
dis_label_gt, dis_label_ge = torch.ones((self.batch_size,1)).cuda(self.gpu), torch.zeros(self.batch_size, 1).cuda(self.gpu)
det_label = labels.cuda(self.gpu)
# localization
output1, output2 = self.loc(images1,images2)
#localization update
if (iter+1) % self.loc_update_stride == 0:
mask1_0,mask1_1 = self.soft_max(output1)
mask2_0,mask2_1 = self.soft_max(output2)
self.loc_optimizer.zero_grad()
dis_label_1_ge,dis_label_2_ge = self.dis(images1_d,images2_d,mask1_0,mask2_0,mask1_1,mask2_1)
det_label_ge = self.det(images1_d,images2_d,mask1_1,mask2_1)
#localization net update
loc_loss_1 = loc_loss_calc(output1,gt1_,self.gpu)
loc_loss_2 = loc_loss_calc(output2,gt2_,self.gpu)
if self.loss_type == 'BCE':
dis_loss_1_ge_ = self.BCElog_Loss(dis_label_1_ge,dis_label_gt)
dis_loss_2_ge_ = self.BCElog_Loss(dis_label_2_ge,dis_label_gt)
else:
dis_loss_1_ge_ = -dis_label_1_ge.mean()
dis_loss_2_ge_ = -dis_label_2_ge.mean()
det_loss_ge_ = self.CE_Loss(det_label_ge,det_label)
loc_loss = self.lambda_loc * (loc_loss_1 + loc_loss_2) + self.lambda_dis * (dis_loss_1_ge_ + dis_loss_2_ge_) + self.lambda_det * det_loss_ge_
loc_loss.backward()
self.loc_optimizer.step()
if (iter+1) % 10 == 0:
print '********************************************************************************'
print 'iter = ',iter, ' epoch = ', epoch, 'completed, loc_loss = ', loc_loss.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, loc_loss_1 = ', loc_loss_1.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, loc_loss_2 = ', loc_loss_2.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, det_loss_ge = ', det_loss_ge_.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_1_ge = ', dis_loss_1_ge_.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_2_ge = ', dis_loss_2_ge_.data.cpu().numpy()
output1 = output1.detach()
output2 = output2.detach()
mask1_0,mask1_1 = self.soft_max(output1)
mask2_0,mask2_1 = self.soft_max(output2)
if self.adapt_gt_flag:
gt1 = generate_adap_gt(gt1,mask1_1,self.gpu)
gt2 = generate_adap_gt(gt2,mask2_1,self.gpu)
gt1_0 = torch.ones((self.batch_size,1,self.input_scale/8,self.input_scale/8)).cuda(self.gpu)-gt1
gt2_0 = torch.ones((self.batch_size,1,self.input_scale/8,self.input_scale/8)).cuda(self.gpu)-gt2
#discrimination
# generated masks
dis_label_1_ge,dis_label_2_ge = self.dis(images1_d,images2_d,mask1_0,mask2_0,mask1_1,mask2_1)
det_label_ge = self.det(images1_d,images2_d,mask1_1,mask2_1)
# gt masks
dis_label_1_gt,dis_label_2_gt = self.dis(images1_d,images2_d,gt1_0,gt2_0,gt1,gt2)
det_label_gt = self.det(images1_d,images2_d,gt1,gt2)
self.dis_optimizer.zero_grad()
self.det_optimizer.zero_grad()
#discriminator update
if self.loss_type == 'BCE':
dis_loss_1_ge = self.BCElog_Loss(dis_label_1_ge,dis_label_ge)
dis_loss_2_ge = self.BCElog_Loss(dis_label_2_ge,dis_label_ge)
dis_loss_1_gt = self.BCElog_Loss(dis_label_1_gt,dis_label_gt)
dis_loss_2_gt = self.BCElog_Loss(dis_label_2_gt,dis_label_gt)
dis_loss = dis_loss_1_ge + dis_loss_2_ge + dis_loss_1_gt + dis_loss_2_gt
elif self.loss_type == 'HIG':
dis_loss_1_gt = nn.ReLU()(1.0-dis_label_1_gt).mean()
dis_loss_1_ge = nn.ReLU()(1.0+dis_label_1_ge).mean()
dis_loss_1 = dis_loss_1_gt + dis_loss_1_ge
dis_loss_2_gt = nn.ReLU()(1.0-dis_label_2_gt).mean()
dis_loss_2_ge = nn.ReLU()(1.0+dis_label_2_ge).mean()
dis_loss_2 = dis_loss_2_gt + dis_loss_2_ge
dis_loss = dis_loss_1 + dis_loss_2
det_loss_ge = self.CE_Loss(det_label_ge,det_label)
det_loss_gt = self.CE_Loss(det_label_gt,det_label)
dis_loss.backward()
self.dis_optimizer.step()
det_loss = det_loss_ge + det_loss_gt
det_loss.backward()
self.det_optimizer.step()
if (iter+1) % 10 == 0:
print '********************************************************************************'
print 'iter = ',iter, ' epoch = ', epoch, 'completed, dis_loss = ', dis_loss.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_1_ge = ', dis_loss_1_ge.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_2_ge = ', dis_loss_2_ge.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, det_loss_ge = ', det_loss_ge.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_1_gt = ', dis_loss_1_gt.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, dis_loss_2_gt = ', dis_loss_2_gt.data.cpu().numpy()
print 'iter = ',iter, ' epoch = ', epoch,'completed, det_loss_gt = ', det_loss_gt.data.cpu().numpy()
if (iter + 1) % self.snapshot_stride == 0:
snapshot(self.loc, self.snapshot_prefix_loc, epoch, iter)
snapshot(self.dis, self.snapshot_prefix_dis, epoch, iter)
snapshot(self.det, self.snapshot_prefix_det, epoch, iter)
self.train_hist['per_epoch_time'].append(time.time() - epoch_start_time)
self.train_hist['total_time'].append(time.time() - start_time)
print("Avg one epoch time: %.2f, total %d epochs time: %.2f" % (np.mean(self.train_hist['per_epoch_time']),
self.epoch, self.train_hist['total_time'][0]))
print("Training finish!... save training results")