-
Notifications
You must be signed in to change notification settings - Fork 24
/
main.py
executable file
·569 lines (472 loc) · 32.8 KB
/
main.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
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
import tensorflow as tf
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
import tensorlayer as tl
#tl.logging.set_verbosity(tl.logging.ERROR)
import numpy as np
from random import shuffle
import matplotlib
import datetime
import time
import shutil
import os
from config import config, log_config, get_eval_path
from utils import *
from model import *
batch_size = config.TRAIN.batch_size
batch_size_init = config.TRAIN.batch_size_init
lr_init = config.TRAIN.lr_init
lr_init_init = config.TRAIN.lr_init_init
beta1 = config.TRAIN.beta1
n_epoch = config.TRAIN.n_epoch
n_epoch_init = config.TRAIN.n_epoch_init
lr_decay = config.TRAIN.lr_decay
decay_every = config.TRAIN.decay_every
lambda_adv = config.TRAIN.lambda_adv
lambda_lr_d = config.TRAIN.lambda_lr_d
lambda_binary = config.TRAIN.lambda_binary
lambda_perceptual = config.TRAIN.lambda_perceptual
h = config.TRAIN.height
w = config.TRAIN.width
ni = int(np.ceil(np.sqrt(batch_size)))
def train():
## CREATE DIRECTORIES
mode_dir = config.TRAIN.root_dir + '{}'.format(tl.global_flag['mode'])
ckpt_dir = os.path.join(mode_dir, 'checkpoint')
init_dir = os.path.join(mode_dir, 'init')
log_dir_scalar_init = os.path.join(mode_dir, 'log/scalar_init')
log_dir_image_init = os.path.join(mode_dir, 'log/image_init')
log_dir_scalar = os.path.join(mode_dir, 'log/scalar')
log_dir_image = os.path.join(mode_dir, 'log/image')
sample_dir = os.path.join(mode_dir, 'samples/train')
config_dir = os.path.join(mode_dir, 'config')
if tl.global_flag['delete_log']:
shutil.rmtree(ckpt_dir, ignore_errors = True)
if tl.global_flag['is_pretrain']:
shutil.rmtree(log_dir_scalar_init, ignore_errors = True)
shutil.rmtree(log_dir_image_init, ignore_errors = True)
shutil.rmtree(log_dir_scalar, ignore_errors = True)
shutil.rmtree(log_dir_image, ignore_errors = True)
shutil.rmtree(sample_dir, ignore_errors = True)
shutil.rmtree(config_dir, ignore_errors = True)
tl.files.exists_or_mkdir(ckpt_dir)
tl.files.exists_or_mkdir(init_dir)
tl.files.exists_or_mkdir(log_dir_scalar_init)
tl.files.exists_or_mkdir(log_dir_image_init)
tl.files.exists_or_mkdir(log_dir_scalar)
tl.files.exists_or_mkdir(log_dir_image)
tl.files.exists_or_mkdir(sample_dir)
tl.files.exists_or_mkdir(config_dir)
log_config(config_dir, config)
## DEFINE SESSION
sess = tf.Session(config = tf.ConfigProto(allow_soft_placement = True, log_device_placement = False))
## READ DATASET LIST
# train
train_real_img_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.real_img_path, regx = '.*', printable = False)))
train_real_binary_map_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.real_binary_map_path, regx = '.*', printable = False)))
train_real_img_no_label_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.real_img_no_label_path, regx = '.*', printable = False)))
# test
test_blur_img_list = np.array(sorted(tl.files.load_file_list(path = config.TEST.cuhk_img_path, regx = '.*', printable = False)))
test_gt_list = np.array(sorted(tl.files.load_file_list(path = config.TEST.cuhk_binary_map_path, regx = '.*', printable = False)))
test_blur_imgs = read_all_imgs(test_blur_img_list, path = config.TEST.cuhk_img_path, mode = 'RGB')
test_gt_imgs = read_all_imgs(test_gt_list, path = config.TEST.cuhk_binary_map_path, mode = 'GRAY')
## DEFINE MODEL
# input
with tf.variable_scope('input'):
patches_synthetic = tf.placeholder('float32', [None, h, w, 3], name = 'input_synthetic')
labels_synthetic_defocus = tf.placeholder('float32', [None, h, w, 1], name = 'labels_synthetic_defocus')
patches_real = tf.placeholder('float32', [None, h, w, 3], name = 'input_real')
labels_real_binary = tf.placeholder('float32', [None, h, w, 1], name = 'labels_real_binary')
patches_real_no_label = tf.placeholder('float32', [None, h, w, 3], name = 'input_real_no_label')
patches_real_test = tf.placeholder('float32', [None, h, w, 3], name = 'input_real')
labels_real_binary_test = tf.placeholder('float32', [None, h, w, 1], name = 'labels_real_binary')
# model
with tf.variable_scope('main_net') as scope:
with tf.variable_scope('defocus_net') as scope:
with tf.variable_scope('encoder') as scope:
net_vgg, feats_synthetic_down, _, _ = VGG19_down(patches_synthetic, reuse = False, scope = scope)
_, feats_real_down, _, _ = VGG19_down(patches_real, reuse = True, scope = scope)
_, feats_real_no_label_down, _, _ = VGG19_down(patches_real_no_label, reuse = True, scope = scope)
with tf.variable_scope('decoder') as scope:
output_synthetic_defocus, feats_synthetic_up_aux, feats_synthetic_da, _ = UNet_up(patches_synthetic, feats_synthetic_down, is_train = True, reuse = False, scope = scope)
output_real_defocus, _, feats_real_da, _ = UNet_up(patches_real, feats_real_down, is_train = True, reuse = True, scope = scope)
output_real_no_label_defocus, _, feats_real_no_label_da, _ = UNet_up(patches_real_no_label, feats_real_no_label_down, is_train = True, reuse = True, scope = scope)
with tf.variable_scope('binary_net') as scope:
output_real_binary_logits, output_real_binary = Binary_Net(output_real_defocus, is_train = True, reuse = False, scope = scope)
with tf.variable_scope('discriminator') as scope:
with tf.variable_scope('feature') as scope:
d_feature_logits_synthetic, d_feature_synthetic = feature_discriminator(feats_synthetic_da, is_train = True, reuse = False, scope = scope)
d_feature_logits_real, d_feature_real = feature_discriminator(feats_real_da, is_train = True, reuse = True, scope = scope)
d_feature_logits_real_no_label, d_feature_real_no_label = feature_discriminator(feats_real_no_label_da, is_train = True, reuse = True, scope = scope)
# fixed
with tf.variable_scope('perceptual') as scope:
output_synthetic_defocus_3c = tf.concat([output_synthetic_defocus, output_synthetic_defocus, output_synthetic_defocus], axis = 3)
net_vgg_perceptual, _, perceptual_synthetic_out, logits_perceptual_out = VGG19_down(output_synthetic_defocus_3c, reuse = False, scope = scope)
labels_synthetic_defocus_3c = tf.concat([labels_synthetic_defocus, labels_synthetic_defocus, labels_synthetic_defocus], axis = 3)
_, _, perceptual_synthetic_label, logits_perceptual_label = VGG19_down(labels_synthetic_defocus_3c, reuse = True, scope = scope)
# for test
with tf.variable_scope('main_net') as scope:
with tf.variable_scope('defocus_net') as scope:
with tf.variable_scope('encoder') as scope:
_, feats_real_down_test, _, _ = VGG19_down(patches_real_test, reuse = True, scope = scope)
with tf.variable_scope('decoder') as scope:
output_real_defocus_test, _, _, _ = UNet_up(patches_real, feats_real_down_test, is_train = True, reuse = True, scope = scope)
## DEFINE LOSS
with tf.variable_scope('loss'):
with tf.variable_scope('discriminator'):
with tf.variable_scope('feature'):
loss_synthetic_d_feature = tl.cost.sigmoid_cross_entropy(d_feature_logits_synthetic, tf.ones_like(d_feature_logits_synthetic), name = 'synthetic')
loss_real_d_feature = tl.cost.sigmoid_cross_entropy(d_feature_logits_real, tf.zeros_like(d_feature_logits_real), name = 'real')
loss_real_no_label_d_feature = tl.cost.sigmoid_cross_entropy(d_feature_logits_real_no_label, tf.zeros_like(d_feature_logits_real_no_label), name = 'real')
loss_d_feature = tf.identity((2 * loss_synthetic_d_feature + loss_real_d_feature + loss_real_no_label_d_feature) / 2. * lambda_adv, name = 'total')
loss_d = tf.identity(loss_d_feature, name = 'total')
with tf.variable_scope('generator'):
with tf.variable_scope('feature'):
loss_real_g_feature = tl.cost.sigmoid_cross_entropy(d_feature_logits_real, tf.ones_like(d_feature_logits_real), name = 'real')
loss_real_no_label_g_feature = tl.cost.sigmoid_cross_entropy(d_feature_logits_real_no_label, tf.ones_like(d_feature_logits_real_no_label), name = 'real')
loss_g_feature = tf.identity((loss_real_g_feature + loss_real_no_label_g_feature) / 2., name = 'total')
loss_g = tf.identity(loss_g_feature * lambda_adv, name = 'total')
with tf.variable_scope('defocus'):
loss_defocus = tl.cost.mean_squared_error(output_synthetic_defocus, labels_synthetic_defocus, is_mean = True, name = 'synthetic')
with tf.variable_scope('auxilary'):
labels_layer = InputLayer(labels_synthetic_defocus)
loss_aux_1 = tl.cost.mean_squared_error(feats_synthetic_up_aux[0],
DownSampling2dLayer(labels_layer, (int(h / 16), int(w / 16)), is_scale = False, method = 1, align_corners=True).outputs, is_mean = True, name = 'aux1')
loss_aux_2 = tl.cost.mean_squared_error(feats_synthetic_up_aux[1],
DownSampling2dLayer(labels_layer, (int(h / 8), int(w / 8)), is_scale = False, method = 1, align_corners=True).outputs, is_mean = True, name = 'aux2')
loss_aux_3 = tl.cost.mean_squared_error(feats_synthetic_up_aux[2],
DownSampling2dLayer(labels_layer, (int(h / 4), int(w / 4)), is_scale = False, method = 1, align_corners=True).outputs, is_mean = True, name = 'aux3')
loss_aux_4 = tl.cost.mean_squared_error(feats_synthetic_up_aux[3],
DownSampling2dLayer(labels_layer, (int(h / 2), int(w / 2)), is_scale = False, method = 1, align_corners=True).outputs, is_mean = True, name = 'aux4')
loss_aux_5 = tl.cost.mean_squared_error(feats_synthetic_up_aux[4], labels_synthetic_defocus, is_mean = True, name = 'aux5')
loss_aux = tf.identity(loss_aux_1 + loss_aux_2 + loss_aux_3 + loss_aux_4 + loss_aux_5, name = 'total')
with tf.variable_scope('perceptual'):
loss_synthetic_perceptual = tl.cost.mean_squared_error(perceptual_synthetic_out, perceptual_synthetic_label, is_mean = True, name = 'synthetic')
loss_perceptual = tf.identity(loss_synthetic_perceptual * lambda_perceptual, name = 'total')
with tf.variable_scope('binary'):
loss_real_binary = tl.cost.sigmoid_cross_entropy(output_real_binary_logits, labels_real_binary, name = 'real')
loss_binary = tf.identity(loss_real_binary * lambda_binary, name = 'total')
loss_main = tf.identity(loss_defocus + loss_binary + loss_perceptual + loss_aux + loss_g, name = 'total')
loss_init = tf.identity(loss_defocus, name = 'loss_init')
## DEFINE OPTIMIZER
# variables to save / train
d_vars = tl.layers.get_variables_with_name('discriminator', True, False)
main_vars = tl.layers.get_variables_with_name('main_net', True, False)
save_vars = tl.layers.get_variables_with_name('main_net', False, False) + tl.layers.get_variables_with_name('discriminator', False, False)
init_vars = tl.layers.get_variables_with_name('defocus_net', True, False)
save_init_vars = tl.layers.get_variables_with_name('defocus_net', False, False)
# define optimizer
with tf.variable_scope('Optimizer'):
learning_rate = tf.Variable(lr_init, trainable = False)
learning_rate_init = tf.Variable(lr_init_init, trainable = False)
optim_d = tf.train.AdamOptimizer(learning_rate * lambda_lr_d, beta1 = beta1).minimize(loss_d, var_list = d_vars)
optim_main = tf.train.AdamOptimizer(learning_rate, beta1 = beta1).minimize(loss_main, var_list = main_vars)
optim_init = tf.train.AdamOptimizer(learning_rate_init, beta1 = beta1).minimize(loss_init, var_list = init_vars)
## DEFINE SUMMARY
# writer
writer_scalar = tf.summary.FileWriter(log_dir_scalar, sess.graph, flush_secs=30, filename_suffix = '.loss_log')
writer_image = tf.summary.FileWriter(log_dir_image, sess.graph, flush_secs=30, filename_suffix = '.image_log')
if tl.global_flag['is_pretrain']:
writer_scalar_init = tf.summary.FileWriter(log_dir_scalar_init, sess.graph, flush_secs=30, filename_suffix = '.loss_log_init')
writer_image_init = tf.summary.FileWriter(log_dir_image_init, sess.graph, flush_secs=30, filename_suffix = '.image_log_init')
# for pretrain
loss_sum_list_init = []
with tf.variable_scope('loss_init'):
loss_sum_list_init.append(tf.summary.scalar('1_total_loss_init', loss_init))
loss_sum_list_init.append(tf.summary.scalar('2_defocus_loss_init', loss_defocus))
loss_sum_init = tf.summary.merge(loss_sum_list_init)
image_sum_list_init = []
image_sum_list_init.append(tf.summary.image('1_synthetic_input_init', patches_synthetic))
image_sum_list_init.append(tf.summary.image('2_synthetic_defocus_out_init', fix_image_tf(output_synthetic_defocus, 1)))
image_sum_list_init.append(tf.summary.image('3_synthetic_defocus_gt_init', fix_image_tf(labels_synthetic_defocus, 1)))
image_sum_init = tf.summary.merge(image_sum_list_init)
# for train
loss_sum_g_list = []
with tf.variable_scope('loss_generator'):
loss_sum_g_list.append(tf.summary.scalar('1_total', loss_main))
loss_sum_g_list.append(tf.summary.scalar('2_g', loss_g))
loss_sum_g_list.append(tf.summary.scalar('3_defocus', loss_defocus))
loss_sum_g_list.append(tf.summary.scalar('4_perceptual', loss_perceptual))
loss_sum_g_list.append(tf.summary.scalar('5_auxilary', loss_aux))
loss_sum_g_list.append(tf.summary.scalar('6_binary', loss_binary))
loss_sum_g = tf.summary.merge(loss_sum_g_list)
loss_sum_d_list = []
with tf.variable_scope('loss_discriminator'):
loss_sum_d_list.append(tf.summary.scalar('1_d', loss_d_feature))
loss_sum_d = tf.summary.merge(loss_sum_d_list)
image_sum_list = []
image_sum_list.append(tf.summary.image('1_synthetic_input', patches_synthetic))
image_sum_list.append(tf.summary.image('2_synthetic_defocus_out', fix_image_tf(output_synthetic_defocus, 1)))
image_sum_list.append(tf.summary.image('3_synthetic_defocus_gt', fix_image_tf(labels_synthetic_defocus, 1)))
image_sum_list.append(tf.summary.image('4_real_input', patches_real))
image_sum_list.append(tf.summary.image('5_real_defocus_out', fix_image_tf(output_real_defocus, 1)))
image_sum_list.append(tf.summary.image('6_real_binary_out', fix_image_tf(output_real_binary, 1)))
image_sum_list.append(tf.summary.image('7_real_binary_gt', fix_image_tf(labels_real_binary, 1)))
image_sum_list.append(tf.summary.image('8_real_binary_gt_no_label', patches_real_no_label))
image_sum_list.append(tf.summary.image('9_real_defocus_out_no_label', fix_image_tf(output_real_no_label_defocus, 1)))
image_sum = tf.summary.merge(image_sum_list)
## INITIALIZE SESSION
sess.run(tf.global_variables_initializer())
## LOAD VGG
vgg19_npy_path = "pretrained/vgg19.npy"
if not os.path.isfile(vgg19_npy_path):
print("Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg")
exit()
npz = np.load(vgg19_npy_path, encoding='latin1', allow_pickle=True).item()
params = []
for val in sorted( npz.items() ):
if val[0] == 'fc6':
break;
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
tl.files.assign_params(sess, params, net_vgg_perceptual)
init_name = init_dir + '/{}_init.npz'.format(tl.global_flag['mode'])
if os.path.isfile(init_name):
tl.files.load_and_assign_npz_dict(name = init_name, sess = sess)
if tl.global_flag['is_pretrain']:
print('*****************************************')
print(' PRE-TRAINING START')
print('*****************************************')
global_step = 0
for epoch in range(0, n_epoch_init):
total_loss_init, n_iter = 0, 0
# reload image list
train_synthetic_img_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.synthetic_img_path, regx = '.*', printable = False)))
train_defocus_map_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.defocus_map_path, regx = '.*', printable = False)))
# shuffle datasets
shuffle_index = np.arange(len(train_synthetic_img_list))
np.random.shuffle(shuffle_index)
train_synthetic_img_list = train_synthetic_img_list[shuffle_index]
train_defocus_map_list = train_defocus_map_list[shuffle_index]
epoch_time = time.time()
#for idx in range(0, len(train_synthetic_img_list), batch_size_init):
for idx in range(0, 10):
step_time = time.time()
## READ DATA
# read synthetic data
b_idx = (idx + np.arange(batch_size_init)) % len(train_synthetic_img_list)
synthetic_images_blur = read_all_imgs(train_synthetic_img_list[b_idx], path = config.TRAIN.synthetic_img_path, mode = 'RGB')
synthetic_defocus_maps = read_all_imgs(train_defocus_map_list[b_idx], path = config.TRAIN.defocus_map_path, mode = 'DEPTH')
synthetic_images_blur, synthetic_defocus_maps = \
crop_pair_with_different_shape_images(synthetic_images_blur, synthetic_defocus_maps, [h, w], is_gaussian_noise = tl.global_flag['is_noise'])
err_init, lr, _ = \
sess.run([loss_init, learning_rate_init, optim_init], {patches_synthetic: synthetic_images_blur, labels_synthetic_defocus: synthetic_defocus_maps})
print('[%s] Ep [%2d/%2d] %4d/%4d time: %4.2fs, err_init: %1.2e, lr: %1.2e' % \
(tl.global_flag['mode'], epoch, n_epoch_init, n_iter, len(train_synthetic_img_list)/batch_size_init, time.time() - step_time, err_init, lr))
if global_step % config.TRAIN.write_log_every == 0:
summary_loss_init, summary_image_init = sess.run([loss_sum_init, image_sum_init], {patches_synthetic: synthetic_images_blur, labels_synthetic_defocus: synthetic_defocus_maps})
writer_scalar_init.add_summary(summary_loss_init, global_step)
writer_image_init.add_summary(summary_image_init, global_step)
total_loss_init += err_init
n_iter += 1
global_step += 1
if epoch % config.TRAIN.refresh_image_log_every and epoch != n_epoch_init == 0:
writer_image_init.close()
remove_file_end_with(log_dir_image_init, '*.image_log')
writer_image_init.reopen()
if epoch % 2 or epoch == n_epoch_init - 1:
tl.files.save_npz_dict(save_init_vars, name = init_dir + '/{}_init.npz'.format(tl.global_flag['mode']), sess = sess)
writer_image_init.close()
writer_scalar_init.close()
## START TRAINING
print('*****************************************')
print(' TRAINING START')
print('*****************************************')
global_step = 0
for epoch in range(0, n_epoch + 1):
total_loss, n_iter = 0, 0
# reload synthetic datasets
train_synthetic_img_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.synthetic_img_path, regx = '.*', printable = False)))
train_defocus_map_list = np.array(sorted(tl.files.load_file_list(path = config.TRAIN.defocus_map_path, regx = '.*', printable = False)))
# shuffle datasets
shuffle_index = np.arange(len(train_synthetic_img_list))
np.random.shuffle(shuffle_index)
train_synthetic_img_list = train_synthetic_img_list[shuffle_index]
train_defocus_map_list = train_defocus_map_list[shuffle_index]
shuffle_index = np.arange(len(train_real_img_list))
np.random.shuffle(shuffle_index)
train_real_img_list = train_real_img_list[shuffle_index]
train_real_binary_map_list = train_real_binary_map_list[shuffle_index]
shuffle_index = np.arange(len(train_real_img_no_label_list))
np.random.shuffle(shuffle_index)
train_real_img_no_label_list = train_real_img_no_label_list[shuffle_index]
# update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay ** (epoch // decay_every)
sess.run(tf.assign(learning_rate, lr_init * new_lr_decay))
elif epoch == 0:
sess.run(tf.assign(learning_rate, lr_init))
epoch_time = time.time()
for idx in range(0, len(train_synthetic_img_list), batch_size):
step_time = time.time()
## READ DATA
# read synthetic data
b_idx = (idx + np.arange(batch_size)) % len(train_synthetic_img_list)
synthetic_images_blur = read_all_imgs(train_synthetic_img_list[b_idx], path = config.TRAIN.synthetic_img_path, mode = 'RGB')
synthetic_defocus_maps = read_all_imgs(train_defocus_map_list[b_idx], path = config.TRAIN.defocus_map_path, mode = 'DEPTH')
synthetic_images_blur, synthetic_defocus_maps = crop_pair_with_different_shape_images(synthetic_images_blur, synthetic_defocus_maps, [h, w], is_gaussian_noise = tl.global_flag['is_noise'])
# read real data #
b_idx = (idx % len(train_real_img_list) + np.arange(batch_size)) % len(train_real_img_list)
real_images_blur = read_all_imgs(train_real_img_list[b_idx], path = config.TRAIN.real_img_path, mode = 'RGB')
real_binary_maps = read_all_imgs(train_real_binary_map_list[b_idx], path = config.TRAIN.real_binary_map_path, mode = 'GRAY')
real_images_blur, real_binary_maps = crop_pair_with_different_shape_images(real_images_blur, real_binary_maps, [h, w])
real_images_no_label_blur = read_all_imgs(train_real_img_no_label_list[b_idx], path = config.TRAIN.real_img_no_label_path, mode = 'RGB')
real_images_no_label_blur = random_crop(real_images_no_label_blur, [h, w])
## RUN NETWORK
#discriminator
feed_dict = {patches_synthetic: synthetic_images_blur, patches_real: real_images_blur, labels_synthetic_defocus: synthetic_defocus_maps, patches_real_no_label: real_images_no_label_blur}
_ = sess.run(optim_d, feed_dict)
#generator
feed_dict = {patches_synthetic: synthetic_images_blur, labels_synthetic_defocus: synthetic_defocus_maps, patches_real: real_images_blur, labels_real_binary: real_binary_maps, patches_real_no_label: real_images_no_label_blur}
_ = sess.run(optim_main, feed_dict)
#log
err_main, err_g, err_d, d_synthetic, d_real, lr = \
sess.run([loss_main, loss_g, loss_d, d_feature_synthetic, d_feature_real, learning_rate], feed_dict)
d_acc = get_disc_accuracy([d_synthetic, d_real], [0, 1])
g_acc = get_disc_accuracy([d_synthetic, d_real], [1, 0])
print('[%s] Ep [%2d/%2d] %4d/%4d time: %4.2fs, err[main: %1.2e, g(acc): %1.2e(%1.2f), d(acc): %1.2e(%1.2f)], lr: %1.2e' % \
(tl.global_flag['mode'], epoch, n_epoch, n_iter, len(train_synthetic_img_list)/batch_size, time.time() - step_time, err_main, err_g, g_acc, err_d, d_acc, lr))
## SAVE LOGS
# save loss & image log
if global_step % config.TRAIN.write_log_every == 0:
summary_loss_g, summary_loss_d, summary_image = sess.run([loss_sum_g, loss_sum_d, image_sum], {patches_synthetic: synthetic_images_blur, labels_synthetic_defocus: synthetic_defocus_maps, patches_real: real_images_blur, labels_real_binary: real_binary_maps, patches_real_no_label: real_images_no_label_blur})
writer_scalar.add_summary(summary_loss_d, global_step)
writer_scalar.add_summary(summary_loss_g, global_step)
writer_image.add_summary(summary_image, global_step)
# save samples
# if global_step != 0 and global_step % config.TRAIN.write_sample_every == 0:
# synthetic_defocus_out, real_defocus_out, real_binary_out = sess.run([output_synthetic_defocus, output_real_defocus, output_real_binary], {patches_synthetic: synthetic_images_blur, patches_real: real_images_blur, labels_real_binary: real_binary_maps })
# save_images(synthetic_images_blur, [ni, ni], sample_dir + '/{}_{}_1_synthetic_input.png'.format(epoch, global_step))
# save_images(norm_image(synthetic_defocus_out), [ni, ni], sample_dir + '/{}_{}_2_synthetic_defocus_out_norm.png'.format(epoch, global_step))
# save_images(norm_image(synthetic_defocus_maps), [ni, ni], sample_dir + '/{}_{}_3_synthetic_defocus_gt.png'.format(epoch, global_step))
# save_images(real_images_blur, [ni, ni], sample_dir + '/{}_{}_4_real_input.png'.format(epoch, global_step))
# save_images(norm_image(real_defocus_out), [ni, ni], sample_dir + '/{}_{}_5_real_defocus_out_norm.png'.format(epoch, global_step))
# save_images(real_binary_out, [ni, ni], sample_dir + '/{}_{}_6_real_binary_out.png'.format(epoch, global_step))
# save_images(real_binary_maps, [ni, ni], sample_dir + '/{}_{}_7_real_binary_gt.png'.format(epoch, global_step))
total_loss += err_main
n_iter += 1
global_step += 1
print('[TRAIN] Epoch: [%2d/%2d] time: %4.4fs, total_err: %1.2e' % (epoch, n_epoch, time.time() - epoch_time, total_loss/n_iter))
# reset image log
if epoch % config.TRAIN.refresh_image_log_every == 0:
writer_image.close()
remove_file_end_with(log_dir_image, '*.image_log')
writer_image.reopen()
if epoch % config.TRAIN.write_ckpt_every == 0:
#remove_file_end_with(ckpt_dir, '*.npz')
tl.files.save_npz_dict(save_vars, name = ckpt_dir + '/{}_{}.npz'.format(tl.global_flag['mode'], epoch), sess = sess)
def evaluate():
date = datetime.datetime.now().strftime('%Y_%m_%d_%H%M')
# directories
mode_dir = os.path.join(config.TRAIN.root_dir, tl.global_flag['mode'])
ckpt_dir = os.path.join(mode_dir, 'checkpoint')
sample_dir = os.path.join(mode_dir, 'results/{}/{}'.format(tl.global_flag['test_set'], date))
# input
input_path, gt_path = get_eval_path(tl.global_flag['test_set'], config)
test_blur_img_list = np.array(sorted(tl.files.load_file_list(path = input_path, regx = '.*', printable = False)))
test_blur_imgs = read_all_imgs(test_blur_img_list, path = input_path, mode = 'RGB')
# gt
if gt_path is not None:
test_gt_list = np.array(sorted(tl.files.load_file_list(path = gt_path, regx = '.*', printable = False)))
mode = 'NPY' if 'RTF' in tl.global_flag['test_set'] else 'GRAY'
test_gt_imgs = read_all_imgs(test_gt_list, path = gt_path, mode = mode)
# define session
sess = tf.Session(config = tf.ConfigProto(allow_soft_placement = True, log_device_placement = False))
# define model
with tf.variable_scope('input'):
patches_blurred = tf.placeholder('float32', [1, None, None, 3], name = 'input_patches')
with tf.variable_scope('main_net') as scope:
with tf.variable_scope('defocus_net') as scope:
with tf.variable_scope('encoder') as scope:
feats_down = VGG19_down(patches_blurred, reuse = False, scope = scope, is_test = True)
with tf.variable_scope('decoder') as scope:
output_defocus, feats_up, _, refine_lists = UNet_up(patches_blurred, feats_down, is_train = False, reuse = False, scope = scope)
# init vars
sess.run(tf.global_variables_initializer())
# load checkpoint
ckpt_path = os.path.join(ckpt_dir, '{}.npz'.format(tl.global_flag['mode']))
if os.path.isfile(ckpt_path) is False:
print('{} does not exit'.format(ckpt_path))
exit()
tl.files.load_and_assign_npz_dict(name = ckpt_path, sess = sess)
print('================')
print('Evaluation Start')
print('================')
print('Results will be saved in: {}\n'.format(sample_dir))
avg_time = 0.
MSE_total = 0
MAE_total = 0
for i in np.arange(len(test_blur_imgs)):
test_blur_img = refine_image(test_blur_imgs[i])
if gt_path is not None:
test_gt_img = refine_image(test_gt_imgs[i])
test_gt_img = np.squeeze(test_gt_img)
# run network
feed_dict = {patches_blurred: np.expand_dims(test_blur_img, axis = 0)}
tic = time.time()
defocus_map, feats_down_out, feats_up_out, refine_lists_out = sess.run([output_defocus, feats_down, feats_up, refine_lists], feed_dict)
toc = time.time()
defocus_map = np.squeeze(defocus_map)
defocus_map_norm = defocus_map - defocus_map.min()
defocus_map_norm = defocus_map_norm / defocus_map_norm.max()
##################
sigma_map = ((defocus_map * 15) - 1) / 2
sigma_map[np.where(sigma_map < 0)] = 0
# when you read, multipy the image by 7 to get sigma
sigma_map = sigma_map / 7.
##################
# quantitative
if gt_path is not None:
if 'RTF' in tl.global_flag['test_set']:
h, w = sigma_map.shape
h_gt, w_gt = test_gt_img.shape
in_temp = sigma_map[:min(h, h_gt), :min(w, w_gt)] * 7.
in_temp[np.where(in_temp>3.275)] = 3.275
in_temp = in_temp / 3.275
gt_temp = test_gt_img[:min(h, h_gt), :min(w, w_gt)]
elif 'SYNDOF' in tl.global_flag['test_set']:
in_temp = defocus_map
gt_temp = test_gt_img
MSE_total = MSE_total + np.mean((in_temp - gt_temp)**2)
MAE_total = MAE_total + np.mean(np.abs(in_temp - gt_temp))
# qualitative
tl.files.exists_or_mkdir(sample_dir, verbose = False)
tl.files.exists_or_mkdir(sample_dir + '/image')
tl.files.exists_or_mkdir(sample_dir + '/defocus_map')
tl.files.exists_or_mkdir(sample_dir + '/defocus_map_min_max_norm')
tl.files.exists_or_mkdir(sample_dir + '/sigma_map_7_norm')
scipy.misc.toimage(test_blur_img, cmin = 0., cmax = 1.).save(sample_dir + '/image/{0:04d}.png'.format(i))
scipy.misc.toimage(defocus_map, cmin = 0., cmax = 1.).save(sample_dir + '/defocus_map/{0:04d}.png'.format(i))
scipy.misc.toimage(defocus_map_norm, cmin = 0., cmax = 1.).save(sample_dir + '/defocus_map_min_max_norm/{0:04d}.png'.format(i))
scipy.misc.toimage(sigma_map, cmin = 0., cmax = 1.).save(sample_dir + '/sigma_map_7_norm/{0:04d}.png'.format(i))
if gt_path is not None:
tl.files.exists_or_mkdir(sample_dir + '/gt')
scipy.misc.toimage(np.squeeze(1 - refine_image(test_gt_imgs[i])), cmin = 0., cmax = 1.).save(sample_dir + '/gt/{0:04d}.png'.format(i))
avg_time = avg_time + (toc - tic)
print('[{}/{}] {} [{:.3f}s]\n'.format(i+1, len(test_blur_imgs), test_blur_img_list[i], toc - tic))
avg_time = avg_time / len(test_blur_imgs)
print('averge time: {:.3f}s'.format(avg_time))
if gt_path is not None:
print('MSE: ', MSE_total / len(test_blur_imgs), ' MAE: ', MAE_total / len(test_blur_imgs))
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type = str, default = 'DMENet', help = 'model name')
parser.add_argument('--is_train', action = 'store_true', default = False, help = 'whether to train or not')
parser.add_argument('--is_pretrain', action = 'store_true', default = False, help = 'whether to pretrain or not')
parser.add_argument('--is_noise', action = 'store_true', default = False, help = 'whether to add noise to synthetic images')
parser.add_argument('--delete_log', action = 'store_true', default = False, help = 'whether to delete log or not')
parser.add_argument('--test_set', type = str , default = 'CUHK', help = 'test_set CUHK|SYNDOF|RTF0|RTF1|RTF1_6|random')
args = parser.parse_args()
tl.global_flag['mode'] = args.mode
tl.global_flag['is_train'] = args.is_train
tl.global_flag['is_pretrain'] = args.is_pretrain
tl.global_flag['is_noise'] = args.is_noise
tl.global_flag['delete_log'] = args.delete_log
if tl.global_flag['is_train']:
train()
else:
tl.global_flag['test_set'] = args.test_set
evaluate()