/
demo.py
456 lines (340 loc) · 18.2 KB
/
demo.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
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
import argparse
import torch
import torch.nn as nn
from torch.utils import data, model_zoo
import numpy as np
import pickle
from torch.autograd import Variable
import torch.optim as optim
import scipy.misc
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import sys
import os
import os.path as osp
import matplotlib.pyplot as plt
import random
import model
import discriminator
import dataset
SOURCE_DATA_DIR = '/path/to/source/dataset'
SOURCE_DATA_LIST = '/path/to/source/datalist'
TARGET_DATA_DIR = '/path/to/target/dataset'
TARGET_DATA_LIST = '/path/to/target/datalist'
SOURCE_IMAGE_SIZE = '256,256'
TARGET_IMAGE_SIZE = '256,256'
def seg_loss(pred, label, gpu):
label = Variable(label.long()).cuda(gpu)
criterion = torch.nn.CrossEntropyLoss(ignore_index=255).cuda(gpu)
return criterion(pred, label)
def consis_loss(t_pred, t2s_pred, gpu):
loss_t2s_t = t2s_pred * torch.log(F.softmax(t_pred, dim=1))
loss_t_t2s = t_pred * torch.log(F.softmax(t2s_pred, dim=1))
loss = (loss_t2s_t + loss_t_t2s) / 2.0
return loss
def loss_adv(pred, gt):
criterion = nn.BCEWithLogitsLoss()
loss = criterion(pred, gt)
return loss
def main():
cudnn.enabled = True
gpu = args.gpu
model_S = create_seg_model(num_classes=args.num_classes)
model_T = create_seg_model(num_classes=args.num_classes)
model_S.load_state_dict(pretrained_model_path)
model_T.load_state_dict(pretrained_model_path)
model_S.train()
model_T.train()
model_S.cuda(args.gpu)
model_T.cuda(args.gpu)
cudnn.benchmark = True
model_D1_S = create_dis_model(num_classes=args.num_classes)
model_D2_S = create_dis_model(num_classes=args.num_classes)
model_D1_S.train()
model_D2_S.train()
model_D1_S.cuda(args.gpu)
model_D2_S.cuda(args.gpu)
model_D1_T = create_dis_model(num_classes=args.num_classes)
model_D2_T = create_dis_model(num_classes=args.num_classes)
model_D1_T.train()
model_D2_T.train()
model_D1_T.cuda(args.gpu)
model_D2_T.cuda(args.gpu)
netS_T = create_trans_model()
netT_S = create_trans_model()
netS_T.load_state_dict(pretrained_S2T_path)
netT_S.load_state_dict(pretrained_T2S_path)
netS_T.cuda(args.gpu)
netT_S.cuda(args.gpu)
for param in netS_T.parameters():
param.requires_grad = False
for param in netT_S.parameters():
param.requires_grad = False
trainloader = data.DataLoader(
GTA5DataSet(args.data_dir,
args.data_list,
max_iters=args.num_steps * args.iter_size * args.batch_size,
crop_size=input_size,
remap_labels=True,
scale=args.random_scale,
mirror=args.random_mirror,
mean=IMG_MEAN),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True
)
trainloader_iter = enumerate(trainloader)
targetloader = data.DataLoader(cityscapesDataSet(args.data_dir_target, args.data_list_target,
max_iters=args.num_steps * args.iter_size * args.batch_size,
crop_size=input_size_target,
scale=False, mirror=args.random_mirror, mean=IMG_MEAN,
set=args.set),
batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers,
pin_memory=True)
targetloader_iter = enumerate(targetloader)
optimizer_S = optim.SGD(model_S.optim_parameters(args),
lr=args.learning_rate, momentum=args.momentum, weight_decay=args.weight_decay)
optimizer_T = optim.SGD(model_T.optim_parameters(args),
lr=args.learning_rate, momentum=args.momentum, weight_decay=args.weight_decay)
optimizer_S.zero_grad()
optimizer_T.zero_grad()
optimizer_D1 = optim.Adam(model_D1.parameters(), lr=args.learning_rate_D, betas=(0.9, 0.99))
optimizer_D1.zero_grad()
optimizer_D2 = optim.Adam(model_D2.parameters(), lr=args.learning_rate_D, betas=(0.9, 0.99))
optimizer_D2.zero_grad()
optimizer_D1_T = optim.Adam(model_D1_T.parameters(), lr=args.learning_rate_D, betas=(0.9, 0.99))
optimizer_D1_T.zero_grad()
optimizer_D2_T = optim.Adam(model_D2_T.parameters(), lr=args.learning_rate_D, betas=(0.9, 0.99))
optimizer_D2_T.zero_grad()
bce_loss = torch.nn.BCEWithLogitsLoss()
interp = nn.Upsample(size=(input_size[1], input_size[0]), mode='bilinear', align_corners=True)
interp_target = nn.Upsample(size=(input_size_target[1], input_size_target[0]), mode='bilinear', align_corners=True)
source_label = 0
target_label = 1
for i_iter in range(args.num_steps):
loss_seg_value1 = 0
loss_consist_1_value = 0
loss_seg_value1_T = 0
loss_adv_target_value1 = 0
loss_adv_target_value1_T = 0
loss_D_value1 = 0
loss_D_value1_T = 0
loss_seg_value2 = 0
loss_consist_2_value = 0
loss_seg_value2_T = 0
loss_adv_target_value2 = 0
loss_adv_target_value2_T = 0
loss_D_value2 = 0
loss_D_value2_T = 0
# print(i_iter)
optimizer_S.zero_grad()
optimizer_T.zero_grad()
adjust_learning_rate(optimizer_S, i_iter)
adjust_learning_rate(optimizer_T, i_iter)
optimizer_D1.zero_grad()
optimizer_D2.zero_grad()
adjust_learning_rate_D(optimizer_D1, i_iter)
adjust_learning_rate_D(optimizer_D2, i_iter)
optimizer_D1_T.zero_grad()
optimizer_D2_T.zero_grad()
adjust_learning_rate_D(optimizer_D1_T, i_iter)
adjust_learning_rate_D(optimizer_D2_T, i_iter)
for sub_i in range(args.iter_size):
for param in model_D1.parameters():
param.requires_grad = False
for param in model_D2.parameters():
param.requires_grad = False
for param in model_D1_T.parameters():
param.requires_grad = False
for param in model_D2_T.parameters():
param.requires_grad = False
# train with source
_, batch = trainloader_iter.__next__()
images_raw, labels, _ = batch
images = Variable(images_raw).cuda(args.gpu) # Xa
GTA_to_Real_images = Variable(netG_A(images).detach()).cuda(args.gpu) #Xac
### Segmentation modle trained with Xa and Xac (those two have same label) ###
pred1, pred2 = model_S(images)
pred1_T, pred2_T = model_T(GTA_to_Real_images)
pred1 = interp(pred1)
pred2 = interp(pred2)
pred1_T = interp(pred1_T)
pred2_T = interp(pred2_T)
# Source label loss #
loss_seg1 = loss_calc(pred1, labels, args.gpu)
loss_seg2 = loss_calc(pred2, labels, args.gpu)
loss_seg1_T = loss_calc(pred1_T, labels, args.gpu)
loss_seg2_T = loss_calc(pred2_T, labels, args.gpu)
loss_S = loss_seg2 + args.lambda_seg * loss_seg1
loss_T = loss_seg2_T + args.lambda_seg * loss_seg1_T
# model_S backward with Xa image and label
loss = loss_S / args.iter_size
loss.backward(retain_graph=True)
# model T backward with Xac image and label
loss = loss_T / args.iter_size
loss.backward(retain_graph=True)
# itemizae loss
loss_seg_value1 += loss_seg1.item() / args.iter_size
loss_seg_value2 += loss_seg2.item() / args.iter_size
loss_seg_value1_T += loss_seg1_T.item() / args.iter_size
loss_seg_value2_T += loss_seg2_T.item() / args.iter_size
# train with target
_, batch = targetloader_iter.__next__()
images_raw, _, _ = batch
images = Variable(images_raw).cuda(args.gpu) # Xc
Real_to_GTA_images = Variable(netG_B(images).detach()).cuda(args.gpu) # Xca
### Segmentation modle trained with Xc and Xca (those two don't have a label) ###
pred_target1, pred_target2 = model_T(images)
pred_target1_S, pred_target2_S = model_S(Real_to_GTA_images)
pred_target1 = interp_target(pred_target1)
pred_target2 = interp_target(pred_target2)
pred_target1_S = interp_target(pred_target1_S)
pred_target2_S = interp_target(pred_target2_S)
# consistency Loss #
pred_target1_prob = F.softmax(pred_target1, dim=1)
pred_target2_prob = F.softmax(pred_target2, dim=1)
pred_target1_S_prob = F.softmax(pred_target1_S, dim=1)
pred_target2_S_prob = F.softmax(pred_target2_S, dim=1)
pred_target1_problog = F.log_softmax(pred_target1, dim=1)
pred_target2_problog = F.log_softmax(pred_target2, dim=1)
pred_target1_S_problog = F.log_softmax(pred_target1_S, dim=1)
pred_target2_S_problog = F.log_softmax(pred_target2_S, dim=1)
loss_consist_1 = KL_criterion(pred_target1_problog, pred_target1_S_prob) + KL_criterion(pred_target1_S_problog, pred_target1_prob)
loss_consist_2 = KL_criterion(pred_target2_problog, pred_target2_S_prob) + KL_criterion(pred_target2_S_problog, pred_target2_prob)
loss = (loss_consist_1 + loss_consist_2) / args.iter_size
loss.backward(retain_graph=True)
# itemize Loss
loss_consist_1_value += loss_consist_1.item() / args.iter_size
loss_consist_2_value += loss_consist_2.item() / args.iter_size
### Feature learning for confusing discriminator ###
# pred_target1(2)_S : model_S(Xca) --> needs to be source like by D_out_S
D_out1_S = model_D1(F.softmax(pred_target1_S, dim=1))
D_out2_S = model_D2(F.softmax(pred_target2_S, dim=1))
# pred1(2)_T : model_T(Xac) --> needs to be target like by D_out_T
D_out1_T = model_D1_T(F.softmax(pred1_T, dim=1))
D_out2_T = model_D2_T(F.softmax(pred2_T, dim=1))
# Loss for confusing D
loss_adv_target1 = bce_loss(D_out1_S,
Variable(torch.FloatTensor(D_out1_S.data.size()).fill_(source_label)).cuda(
args.gpu))
loss_adv_target2 = bce_loss(D_out2_S,
Variable(torch.FloatTensor(D_out2_S.data.size()).fill_(source_label)).cuda(
args.gpu))
loss_adv_target1_T = bce_loss(D_out1_T,
Variable(torch.FloatTensor(D_out1_T.data.size()).fill_(target_label)).cuda(
args.gpu))
loss_adv_target2_T = bce_loss(D_out2_T,
Variable(torch.FloatTensor(D_out2_T.data.size()).fill_(target_label)).cuda(
args.gpu))
loss_S = args.lambda_adv_target1 * loss_adv_target1 + args.lambda_adv_target2 * loss_adv_target2
loss_T = args.lambda_adv_target1 * loss_adv_target1_T + args.lambda_adv_target2 * loss_adv_target2_T
# Backward for D_out_S
loss = loss_S / args.iter_size
loss.backward()
# Backward for D_out_T
loss = loss_T / args.iter_size
loss.backward()
# Itemize the Loss
loss_adv_target_value1 += loss_adv_target1.item()
loss_adv_target_value2 += loss_adv_target2.item()
loss_adv_target_value1_T += loss_adv_target1_T.item()
loss_adv_target_value2_T += loss_adv_target2_T.item()
for param in model_D1.parameters():
param.requires_grad = True
for param in model_D2.parameters():
param.requires_grad = True
for param in model_D1_T.parameters():
param.requires_grad = True
for param in model_D2_T.parameters():
param.requires_grad = True
pred1 = pred1.detach() # GTA
pred2 = pred2.detach()
pred_target1 = pred_target1.detach() # Real
pred_target2 = pred_target2.detach()
D_out1_S = model_D1(F.softmax(pred1))
D_out2_S = model_D2(F.softmax(pred2))
D_out1_T = model_D1_T(F.softmax(pred_target1))
D_out2_T = model_D2_T(F.softmax(pred_target2))
loss_D1 = bce_loss(D_out1_S,
Variable(torch.FloatTensor(D_out1_S.data.size()).fill_(source_label)).cuda(args.gpu))
loss_D2 = bce_loss(D_out2_S,
Variable(torch.FloatTensor(D_out2_S.data.size()).fill_(source_label)).cuda(args.gpu))
loss_D1_T = bce_loss(D_out1_T,
Variable(torch.FloatTensor(D_out1_T.data.size()).fill_(target_label)).cuda(args.gpu))
loss_D2_T = bce_loss(D_out2_T,
Variable(torch.FloatTensor(D_out2_T.data.size()).fill_(target_label)).cuda(args.gpu))
loss_D1 = loss_D1 / args.iter_size / 2
loss_D2 = loss_D2 / args.iter_size / 2
loss_D1_T = loss_D1_T / args.iter_size / 2
loss_D2_T = loss_D2_T / args.iter_size / 2
loss_D1.backward()
loss_D2.backward()
loss_D1_T.backward()
loss_D2_T.backward()
# Itemize the loss
loss_D_value1 += loss_D1.item()
loss_D_value2 += loss_D2.item()
loss_D_value1_T += loss_D1_T.item()
loss_D_value2_T += loss_D2_T.item()
# train with target
pred1_T = pred1_T.detach()
pred2_T = pred2_T.detach()
pred_target1_S = pred_target1_S.detach()
pred_target2_S = pred_target2_S.detach()
D_out1_S = model_D1(F.softmax(pred_target1_S))
D_out2_S = model_D2(F.softmax(pred_target2_S))
D_out1_T = model_D1_T(F.softmax(pred1_T))
D_out2_T = model_D2_T(F.softmax(pred2_T))
loss_D1 = bce_loss(D_out1_S,
Variable(torch.FloatTensor(D_out1_S.data.size()).fill_(target_label)).cuda(args.gpu))
loss_D2 = bce_loss(D_out2_S,
Variable(torch.FloatTensor(D_out2_S.data.size()).fill_(target_label)).cuda(args.gpu))
loss_D1_T = bce_loss(D_out1_T,
Variable(torch.FloatTensor(D_out1_T.data.size()).fill_(source_label)).cuda(args.gpu))
loss_D2_T = bce_loss(D_out2_T,
Variable(torch.FloatTensor(D_out2_T.data.size()).fill_(source_label)).cuda(args.gpu))
loss_D1 = loss_D1 / args.iter_size / 2
loss_D2 = loss_D2 / args.iter_size / 2
loss_D1_T = loss_D1_T / args.iter_size / 2
loss_D2_T = loss_D2_T / args.iter_size / 2
# loss_D_second = loss_D1 + loss_D2 + loss_D1_T + loss_D2_T
loss_D1.backward()
loss_D2.backward()
loss_D1_T.backward()
loss_D2_T.backward()
loss_D_value1 += loss_D1.item()
loss_D_value2 += loss_D2.item()
loss_D_value1_T += loss_D1_T.item()
loss_D_value2_T += loss_D2_T.item()
optimizer_S.step()
optimizer_T.step()
optimizer_D1.step()
optimizer_D2.step()
optimizer_D1_T.step()
optimizer_D2_T.step()
print('exp = {}'.format(args.snapshot_dir))
print(
'iter = {0:8d}/{1:8d}, loss_seg1 = {2:.3f} loss_seg2 = {3:.3f} loss_seg_value1_T = {4:.3f} loss_seg_value2_T = {5:.3f} loss_adv1 = {5:.3f}, loss_adv2 = {6:.3f} loss_adv_target_value1_T = {7:.3f} loss_adv_target_value1_T = {8:.3f} loss_D1 = {9:.3f} loss_D2 = {10:.3f} loss_D1_T = {11:.3f} loss_D2_T = {12:.3f}'.format(
i_iter, args.num_steps, loss_seg_value1, loss_seg_value2, loss_seg_value1_T, loss_seg_value2_T, loss_adv_target_value1, loss_adv_target_value2, loss_adv_target_value1_T, loss_adv_target_value2_T, loss_D_value1, loss_D_value2, loss_D_value1_T, loss_D_value2_T))
Loss_list_Tensorboard = [loss_seg_value1, loss_seg_value2, loss_seg_value1_T, loss_seg_value2_T, loss_adv_target_value1, loss_adv_target_value2, loss_adv_target_value1_T, loss_adv_target_value2_T, loss_D_value1, loss_D_value2, loss_D_value1_T, loss_D_value2_T]
# write_loss(writer, Loss_list_Tensorboard, i_iter)
if i_iter >= args.num_steps_stop - 1:
print('save model ...')
torch.save(model_S.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + 'S.pth'))
torch.save(model_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + 'T.pth'))
torch.save(model_D1.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + '_D1.pth'))
torch.save(model_D2.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + '_D2.pth'))
torch.save(model_D1_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + '_D1_T.pth'))
torch.save(model_D2_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(args.num_steps_stop) + '_D2_T.pth'))
break
if i_iter % args.save_pred_every == 0 and i_iter != 0:
print('taking snapshot ...')
torch.save(model_S.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + 'S.pth'))
torch.save(model_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + 'T.pth'))
torch.save(model_D1.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + '_D1.pth'))
torch.save(model_D2.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + '_D2.pth'))
torch.save(model_D1_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + '_D1_T.pth'))
torch.save(model_D2_T.state_dict(), osp.join(args.snapshot_dir, 'GTA5_' + str(i_iter) + '_D2_T.pth'))
if __name__ == '__main__':
main()