forked from thunguyenphuoc/HoloGAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_HoloGAN.py
713 lines (603 loc) · 35.5 KB
/
model_HoloGAN.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
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
from __future__ import division
import os
import shutil
import sys
import torch
import numpy as np
import importlib
import tensorflow as tf
from torchvision import transforms
from glob import glob
from tensorflow.contrib import graph_editor as ge
from tqdm import tqdm
DEPLOY = True
if DEPLOY:
from runpy import run_path
from argparse import Namespace
cfg = Namespace(**run_path(sys.argv[1]))
else:
from configs import config as cfg
IMAGE_PATH = cfg.image_path
OUTPUT_DIR = cfg.output_dir
LOGDIR = cfg.log_dir
from tools.ops import *
from tools.utils import get_image, merge, inverse_transform, to_bool
from tools.rotation_utils import *
from tools.model_utils import transform_voxel_to_match_image
from warnings import filterwarnings
filterwarnings('ignore')
# ----------------------------------------------------------------------------
GRID_X, GRID_Y = 8, 8
SAMPLE_VIEW = np.tile(np.concatenate([
generate_random_rotation_translation(
1,
cfg.ele_low, cfg.ele_high,
angle, angle + 1,
cfg.scale_low, cfg.scale_high,
cfg.x_low, cfg.x_high,
cfg.y_low, cfg.y_high,
cfg.z_low, cfg.z_high,
with_translation=False,
with_scale=to_bool(str(cfg.with_translation))
)
for angle in np.linspace(cfg.azi_low, cfg.azi_high, GRID_X)]), [GRID_Y, 1])
SAMPLE_VIEW = SAMPLE_VIEW.reshape((GRID_Y, GRID_X, -1)).transpose((1, 0, 2)).reshape((GRID_X * GRID_Y, -1))
assert SAMPLE_VIEW.shape == (GRID_X * GRID_Y, 6)
ADD_EMBEDDING = True
if ADD_EMBEDDING:
cfg.z_dim += cfg.emb_dim
class HoloGAN(object):
def __init__(self, sess, emb_graph, input_height=108, input_width=108, crop=True,
output_height=64, output_width=64,
gf_dim=64, df_dim=64,
c_dim=3, dataset_name='lsun',
input_fname_pattern='*.webp'):
self.emb_graph = emb_graph
self.sess = sess
self.crop = crop
self.input_height = input_height
self.input_width = input_width
self.output_height = output_height
self.output_width = output_width
self.gf_dim = gf_dim
self.df_dim = df_dim
self.c_dim = c_dim
self.dataset_name = dataset_name
self.input_fname_pattern = input_fname_pattern
self.data = glob.glob(os.path.join(IMAGE_PATH, self.input_fname_pattern))
self.checkpoint_dir = LOGDIR
# checkpoint = torch.load(cfg.emb_checkpoint)
# self.embedder = checkpoint['model'].module
self.emb_transforms = transforms.Compose([
transforms.CenterCrop((self.input_height, self.input_width)),
transforms.Resize((cfg.emb_input_size, cfg.emb_input_size), Image.BICUBIC),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
# with tf_graph.as_default() as g:
# self.emb_input = tf.placeholder(tf.float32, [None, self.c_dim, cfg.emb_input_size, cfg.emb_input_size],
# name='input')
# g_def = g.as_graph_def()
# self.emb_output = tf.import_graph_def(g_def, input_map={"emb_input:0": self.emb_input}, return_elements=['emb_output:0'])
# print(g.get_tensor_by_name('emb_input:0'))
# print(tf_graph.get_tensor_by_name('emb_input:0'))
if ADD_EMBEDDING:
self.emb_input = self.emb_graph.get_tensor_by_name('emb_input:0')
self.emb_output = self.emb_graph.get_tensor_by_name('emb_output:0')
run_config = tf.ConfigProto()
run_config.gpu_options.allow_growth = True
self.sess_emb = tf.Session(config=run_config, graph=self.emb_graph)
global SAMPLE_Z
if ADD_EMBEDDING:
SAMPLE_Z = np.random.uniform(-1., 1., (1, GRID_Y, cfg.z_dim - cfg.emb_dim))
else:
SAMPLE_Z = np.random.uniform(-1., 1., (1, GRID_Y, cfg.z_dim))
# SAMPLE_Z = np.random.uniform(-1., 1., (GRID_X, 1, cfg.z_dim))
if ADD_EMBEDDING:
emb_test = glob.glob(os.path.join(cfg.emb_test, self.input_fname_pattern))
batch_emb = torch.stack([self.emb_transforms(Image.open(file).convert('RGB')) for file in emb_test]).numpy()
batch_emb = self.sess_emb.run(self.emb_output, {self.emb_input: np.tile(batch_emb, (4, 1, 1, 1))})[:8]
batch_emb = np.expand_dims(batch_emb, 0)
SAMPLE_Z = np.concatenate((SAMPLE_Z, batch_emb), axis=-1)
SAMPLE_Z = SAMPLE_Z * np.ones((GRID_X, 1, cfg.z_dim))
SAMPLE_Z = SAMPLE_Z.reshape((-1, cfg.z_dim))
if ADD_EMBEDDING:
self.images = []
for i, image in enumerate([transforms.Compose([
transforms.CenterCrop((self.input_height, self.input_width)),
transforms.Resize((self.output_height, self.output_width)),
])(Image.open(file).convert('RGB')) for file in emb_test]):
# image = inverse_transform(image)
# image = np.clip(255 * image, 0, 255).astype(np.uint8)
# Image.fromarray(inverse_transform(image).astype(np.uint8))\
self.images.append(np.array(image).astype(np.uint8))
# image.save(
# os.path.join(OUTPUT_DIR, "{0}_ASD.png".format(i)))
self.images = np.stack(self.images)
def build(self, build_func_name):
build_func = eval("self." + build_func_name)
build_func()
def build_HoloGAN(self):
self.view_in = tf.placeholder(tf.float32, [None, 6], name='view_in')
self.inputs = tf.placeholder(tf.float32, [None, self.output_height, self.output_width, self.c_dim],
name='real_images')
self.z = tf.placeholder(tf.float32, [None, cfg.z_dim], name='z')
inputs = self.inputs
gen_func = eval("self." + (cfg.generator))
dis_func = eval("self." + (cfg.discriminator))
self.gen_view_func = eval(cfg.view_func)
self.G = gen_func(self.z, self.view_in)
if str.lower(str(cfg.style_disc)) == "true":
print("Style Disc")
self.D, self.D_logits, _, self.d_h1_r, self.d_h2_r, self.d_h3_r, self.d_h4_r = dis_func(
inputs, cont_dim=cfg.z_dim, reuse=False)
self.D_, self.D_logits_, self.Q_c_given_x, self.d_h1_f, self.d_h2_f, self.d_h3_f, self.d_h4_f = dis_func(
self.G, cont_dim=cfg.z_dim, reuse=True)
self.d_h1_loss = cfg.DStyle_lambda * (
tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h1_r, tf.ones_like(self.d_h1_r))) \
+ tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h1_f, tf.zeros_like(self.d_h1_f))))
self.d_h2_loss = cfg.DStyle_lambda * (
tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h2_r, tf.ones_like(self.d_h2_r))) \
+ tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h2_f, tf.zeros_like(self.d_h2_f))))
self.d_h3_loss = cfg.DStyle_lambda * (
tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h3_r, tf.ones_like(self.d_h3_r))) \
+ tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h3_f, tf.zeros_like(self.d_h3_f))))
self.d_h4_loss = cfg.DStyle_lambda * (
tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h4_r, tf.ones_like(self.d_h4_r))) \
+ tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.d_h4_f, tf.zeros_like(self.d_h4_f))))
else:
self.D, self.D_logits, _ = dis_func(inputs, cont_dim=cfg.z_dim, reuse=False)
self.D_, self.D_logits_, self.Q_c_given_x = dis_func(self.G, cont_dim=cfg.z_dim, reuse=True)
self.d_loss_real = tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.D_logits, tf.ones_like(self.D)))
self.d_loss_fake = tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.D_logits_, tf.zeros_like(self.D_)))
self.d_loss = self.d_loss_real + self.d_loss_fake
self.g_loss = tf.reduce_mean(sigmoid_cross_entropy_with_logits(self.D_logits_, tf.ones_like(self.D_)))
if str.lower(str(cfg.style_disc)) == "true":
print("Style disc")
self.d_loss = self.d_loss + self.d_h1_loss + self.d_h2_loss + self.d_h3_loss + self.d_h4_loss
# ====================================================================================================================
# Identity loss
self.q_loss = cfg.lambda_latent * tf.reduce_mean(tf.square(self.Q_c_given_x - self.z))
self.d_loss = self.d_loss + self.q_loss
self.g_loss = self.g_loss + self.q_loss
# ====================================================================================================================
# Emb loss
if ADD_EMBEDDING:
self.emb_reconstructed = tf.image.resize_images(self.G, [cfg.emb_input_size] * 2)
self.emb_reconstructed = tf.transpose(self.emb_reconstructed, [0, 3, 1, 2])
self.emb_reconstructed = ge.graph_replace(self.emb_output, {self.emb_input: self.emb_reconstructed})
self.e_loss = cfg.lambda_emb * tf.reduce_mean(tf.square(self.emb_reconstructed - self.z[:, cfg.z_dim - cfg.emb_dim:]))
self.g_loss = self.g_loss + self.e_loss
self.d_loss_real_sum = scalar_summary("d_loss_real", self.d_loss_real)
self.d_loss_fake_sum = scalar_summary("d_loss_fake", self.d_loss_fake)
self.g_loss_sum = scalar_summary("g_loss", self.g_loss)
self.d_loss_sum = scalar_summary("d_loss", self.d_loss)
t_vars = tf.trainable_variables()
self.d_vars = [var for var in t_vars if 'd_' in var.name]
self.g_vars = [var for var in t_vars if 'g_' in var.name]
self.saver = tf.train_on_loader.Saver()
def train_HoloGAN(self):
self.d_lr_in = tf.placeholder(tf.float32, None, name='d_eta')
self.g_lr_in = tf.placeholder(tf.float32, None, name='d_eta')
d_optim = tf.train_on_loader.AdamOptimizer(cfg.d_eta, beta1=cfg.beta1, beta2=cfg.beta2).minimize(self.d_loss,
var_list=self.d_vars)
g_optim = tf.train_on_loader.AdamOptimizer(cfg.g_eta, beta1=cfg.beta1, beta2=cfg.beta2).minimize(self.g_loss,
var_list=self.g_vars)
tf.global_variables_initializer().run()
self.g_sum = merge_summary([self.d_loss_fake_sum, self.g_loss_sum])
self.d_sum = merge_summary([self.d_loss_real_sum, self.d_loss_sum])
self.writer = SummaryWriter(LOGDIR, self.sess.graph)
# Sample noise Z and view parameters to test during training
# sample_z = self.sampling_Z(cfg.z_dim - cfg.emb_dim, str(cfg.sample_z))
# sample_view = self.gen_view_func(cfg.batch_size,
# cfg.ele_low, cfg.ele_high,
# cfg.azi_low, cfg.azi_high,
# cfg.scale_low, cfg.scale_high,
# cfg.x_low, cfg.x_high,
# cfg.y_low, cfg.y_high,
# cfg.z_low, cfg.z_high,
# with_translation=False,
# with_scale=to_bool(str(cfg.with_translation)))
sample_files = self.data[0:cfg.batch_size]
if cfg.dataset == "cats" or cfg.dataset == "cars":
sample_images = [get_image(sample_file,
input_height=self.input_height,
input_width=self.input_width,
resize_height=self.output_height,
resize_width=self.output_width,
crop=False) for sample_file in sample_files]
else:
sample_images = [get_image(sample_file,
input_height=self.input_height,
input_width=self.input_width,
resize_height=self.output_height,
resize_width=self.output_width,
crop=self.crop) for sample_file in sample_files]
counter = 1
start_time = time.time()
could_load, checkpoint_counter = self.load(self.checkpoint_dir)
if could_load:
counter = checkpoint_counter
print(" [*] Load SUCCESS")
else:
print(" [!] Load failed...")
self.data = glob.glob(os.path.join(IMAGE_PATH, self.input_fname_pattern))
d_lr = cfg.d_eta
g_lr = cfg.g_eta
for epoch in range(cfg.max_epochs):
d_lr = d_lr if epoch < cfg.epoch_step else d_lr * (cfg.max_epochs - epoch) / (
cfg.max_epochs - cfg.epoch_step)
g_lr = g_lr if epoch < cfg.epoch_step else g_lr * (cfg.max_epochs - epoch) / (
cfg.max_epochs - cfg.epoch_step)
random.shuffle(self.data)
batch_idxs = min(len(self.data), cfg.train_size) // cfg.batch_size
for idx in tqdm(range(0, batch_idxs)):
batch_files = self.data[idx * cfg.batch_size:(idx + 1) * cfg.batch_size]
if cfg.dataset == "cats" or cfg.dataset == "cars":
batch_images = [get_image(batch_file,
input_height=self.input_height,
input_width=self.input_width,
resize_height=self.output_height,
resize_width=self.output_width,
crop=False) for batch_file in batch_files]
else:
batch_images = [get_image(batch_file,
input_height=self.input_height,
input_width=self.input_width,
resize_height=self.output_height,
resize_width=self.output_width,
crop=self.crop) for batch_file in batch_files]
if ADD_EMBEDDING:
batch_z = self.sampling_Z(cfg.z_dim - cfg.emb_dim, str(cfg.sample_z))
else:
batch_z = self.sampling_Z(cfg.z_dim, str(cfg.sample_z))
# batch_z = self.sampling_Z(cfg.z_dim, str(cfg.sample_z))
# batch_emb = torch.stack([self.emb_transforms(Image.open(file).convert('RGB')) for file in batch_files])
# batch_emb = self.embedder(batch_emb.to(torch.float32).cuda()).cpu().detach().numpy()
if ADD_EMBEDDING:
batch_emb = [self.emb_transforms(Image.open(file).convert('RGB')) for file in batch_files]
if len(batch_emb) != cfg.batch_size:
batch_emb = batch_emb * cfg.batch_size
batch_emb = batch_emb[:cfg.batch_size]
batch_emb = torch.stack(batch_emb).numpy()
batch_emb = self.sess_emb.run(self.emb_output, {self.emb_input: batch_emb})
batch_z = np.hstack((batch_z, batch_emb))
batch_view = self.gen_view_func(cfg.batch_size,
cfg.ele_low, cfg.ele_high,
cfg.azi_low, cfg.azi_high,
cfg.scale_low, cfg.scale_high,
cfg.x_low, cfg.x_high,
cfg.y_low, cfg.y_high,
cfg.z_low, cfg.z_high,
with_translation=False,
with_scale=to_bool(str(cfg.with_translation)))
# control_weight = [
# op for op in tf.get_default_graph().get_operations()
# if op.name == 'layer3.20.conv1.weight']
# control_weight = control_weight[0].values()[0].eval(session=self.sess)
# print(np.mean(control_weight), np.std(control_weight))
feed = {self.inputs: batch_images,
self.z: batch_z,
self.view_in: batch_view,
self.d_lr_in: d_lr,
self.g_lr_in: g_lr}
# Update D network
_, summary_str = self.sess.run([d_optim, self.d_sum], feed_dict=feed)
self.writer.add_summary(summary_str, counter)
# Update G network
_, summary_str = self.sess.run([g_optim, self.g_sum], feed_dict=feed)
self.writer.add_summary(summary_str, counter)
# Run g_optim twice
_, summary_str = self.sess.run([g_optim, self.g_sum], feed_dict=feed)
self.writer.add_summary(summary_str, counter)
errD_fake = self.d_loss_fake.eval(feed)
errD_real = self.d_loss_real.eval(feed)
errG = self.g_loss.eval(feed)
errQ = self.q_loss.eval(feed)
if ADD_EMBEDDING:
errE = self.e_loss.eval(feed)
else:
errE = 0
counter += 1
print("Epoch: [%2d] [%4d/%4d] time: %4.4f, d_loss: %.8f, g_loss: %.8f, q_loss: %.8f, e_loss: %.8f" \
% (epoch, idx, batch_idxs,
time.time() - start_time, errD_fake + errD_real, errG, errQ, errE))
if np.mod(counter, cfg.refresh_rate) == 1:
assert 'log' in str(LOGDIR)
shutil.rmtree(LOGDIR, ignore_errors=True)
shutil.copyfile(sys.argv[1], os.path.join(LOGDIR, 'config.py'))
self.save(LOGDIR, counter)
# Use global vectors
feed_eval = {self.inputs: sample_images,
self.z: SAMPLE_Z,
self.view_in: SAMPLE_VIEW,
self.d_lr_in: d_lr,
self.g_lr_in: g_lr}
samples, d_loss, g_loss = self.sess.run(
[self.G, self.d_loss, self.g_loss],
feed_dict=feed_eval)
ren_img = inverse_transform(samples)
ren_img = np.clip(255 * ren_img, 0, 255).astype(np.uint8)
grid_x = GRID_X
if ADD_EMBEDDING:
ren_img = np.concatenate([self.images, ren_img], axis=0)
grid_x += 1
# try:
Image.fromarray(merge(ren_img, [grid_x, GRID_Y]).astype(np.uint8)).save(
os.path.join(OUTPUT_DIR, "{0}_GAN.png".format(counter)))
print("[Sample] d_loss: %.8f, g_loss: %.8f" % (d_loss, g_loss))
# except:
# Image.fromarray(ren_img[0]).save(
# os.path.join(OUTPUT_DIR, "{0}_GAN.png".format(counter)))
# print("[Sample] d_loss: %.8f, g_loss: %.8f" % (d_loss, g_loss))
def sample_HoloGAN(self):
could_load, checkpoint_counter = self.load(self.checkpoint_dir)
if could_load:
counter = checkpoint_counter
print(" [*] Load SUCCESS")
else:
print(" [!] Load failed...")
return
SAMPLE_DIR = os.path.join(OUTPUT_DIR, "samples")
if not os.path.exists(SAMPLE_DIR):
os.makedirs(SAMPLE_DIR)
# sample_z = self.sampling_Z(cfg.z_dim, str(cfg.sample_z))
if cfg.rotate_azimuth:
low = cfg.azi_low
high = cfg.azi_high
step = 10
elif cfg.rotate_elevation:
low = cfg.ele_low
high = cfg.ele_high
step = 5
else:
low = 0
high = 10
step = 1
for i in range(low, high, step):
if cfg.rotate_azimuth:
sample_view = np.tile(
np.array([i * math.pi / 180.0, 0 * math.pi / 180.0, 1.0, 0, 0, 0]), (cfg.batch_size, 1))
elif cfg.rotate_elevation:
sample_view = np.tile(
np.array([270 * math.pi / 180.0, (90 - i) * math.pi / 180.0, 1.0, 0, 0, 0]), (cfg.batch_size, 1))
else:
sample_view = self.gen_view_func(cfg.batch_size,
cfg.ele_low, cfg.ele_high,
cfg.azi_low, cfg.azi_high,
cfg.scale_low, cfg.scale_high,
cfg.x_low, cfg.x_high,
cfg.y_low, cfg.y_high,
cfg.z_low, cfg.z_high,
with_translation=False,
with_scale=to_bool(str(cfg.with_translation)))
# Use global vectors
feed_eval = {self.z: SAMPLE_Z,
self.view_in: SAMPLE_VIEW}
samples = self.sess.run(self.G, feed_dict=feed_eval)
ren_img = inverse_transform(samples)
ren_img = np.clip(255 * ren_img, 0, 255).astype(np.uint8)
grid_x = GRID_X
if ADD_EMBEDDING:
ren_img = np.concatenate([self.images, ren_img], axis=0)
grid_x += 1
Image.fromarray(merge(ren_img, [grid_x, GRID_Y]).astype(np.uint8)).save(
os.path.join(SAMPLE_DIR, "{0}_samples_{1}.png".format(counter, i)))
# =======================================================================================================================
def sampling_Z(self, z_dim, type="uniform"):
if str.lower(type) == "uniform":
np.random.seed()
return np.random.uniform(-1., 1., (cfg.batch_size, z_dim))
else:
return np.random.normal(0, 1, (cfg.batch_size, z_dim))
def linear_classifier(self, features, scope="lin_class", stddev=0.02, reuse=False):
with tf.variable_scope(scope) as sc:
w = tf.get_variable('w', [features.get_shape()[-1], 1],
initializer=tf.random_normal_initializer(stddev=stddev))
b = tf.get_variable('biases', 1, initializer=tf.constant_initializer(0.0))
logits = tf.matmul(features, w) + b
return tf.nn.sigmoid(logits), logits
def z_mapping_function(self, z, output_channel, scope='z_mapping', act="relu", stddev=0.02):
with tf.variable_scope(scope) as sc:
w = tf.get_variable('w', [z.get_shape()[-1], output_channel * 2],
initializer=tf.random_normal_initializer(stddev=stddev))
b = tf.get_variable('biases', output_channel * 2, initializer=tf.constant_initializer(0.0))
if act == "relu":
out = tf.nn.relu(tf.matmul(z, w) + b)
else:
out = lrelu(tf.matmul(z, w) + b)
return out[:, :output_channel], out[:, output_channel:]
# =======================================================================================================================
def discriminator_IN(self, image, cont_dim, reuse=False):
if str(cfg.add_D_noise) == "true":
image = image + tf.random_normal(tf.shape(image), stddev=0.02)
with tf.variable_scope("discriminator") as scope:
if reuse:
scope.reuse_variables()
h0 = lrelu(conv2d(image, self.df_dim, name='d_h0_conv'))
h1 = lrelu(instance_norm(conv2d_specNorm(h0, self.df_dim * 2, name='d_h1_conv'), 'd_in1'))
h2 = lrelu(instance_norm(conv2d_specNorm(h1, self.df_dim * 4, name='d_h2_conv'), 'd_in2'))
h3 = lrelu(instance_norm(conv2d_specNorm(h2, self.df_dim * 8, name='d_h3_conv'), 'd_in3'))
# Returning logits to determine whether the images are real or fake
h4 = linear(slim.flatten(h3), 1, 'd_h4_lin')
# Recognition network for latent variables has an additional layer
encoder = lrelu((linear(slim.flatten(h3), 128, 'd_latent')))
cont_vars = linear(encoder, cont_dim, "d_latent_prediction")
return tf.nn.sigmoid(h4), h4, tf.nn.tanh(cont_vars)
def discriminator_IN_style_res128(self, image, cont_dim, reuse=False):
batch_size = tf.shape(image)[0]
if str(cfg.add_D_noise) == "true":
image = image + tf.random_normal(tf.shape(image), stddev=0.02)
with tf.variable_scope("discriminator") as scope:
if reuse:
scope.reuse_variables()
h0 = lrelu(conv2d(image, self.df_dim, name='d_h0_conv'))
h1 = conv2d_specNorm(h0, self.df_dim * 2, name='d_h1_conv')
h1, h1_mean, h1_var = instance_norm(h1, 'd_in1', True)
h1_mean = tf.reshape(h1_mean, (batch_size, self.df_dim * 2))
h1_var = tf.reshape(h1_var, (batch_size, self.df_dim * 2))
d_h1_style = tf.concat([h1_mean, h1_var], 0)
d_h1, d_h1_logits = self.linear_classifier(d_h1_style, "d_h1_class")
h1 = lrelu(h1)
h2 = conv2d_specNorm(h1, self.df_dim * 4, name='d_h2_conv')
h2, h2_mean, h2_var = instance_norm(h2, 'd_in2', True)
h2_mean = tf.reshape(h2_mean, (batch_size, self.df_dim * 4))
h2_var = tf.reshape(h2_var, (batch_size, self.df_dim * 4))
d_h2_style = tf.concat([h2_mean, h2_var], 0)
d_h2, d_h2_logits = self.linear_classifier(d_h2_style, "d_h2_class")
h2 = lrelu(h2)
h3 = conv2d_specNorm(h2, self.df_dim * 8, name='d_h3_conv')
h3, h3_mean, h3_var = instance_norm(h3, 'd_in3', True)
h3_mean = tf.reshape(h3_mean, (batch_size, self.df_dim * 8))
h3_var = tf.reshape(h3_var, (batch_size, self.df_dim * 8))
d_h3_style = tf.concat([h3_mean, h3_var], 0)
d_h3, d_h3_logits = self.linear_classifier(d_h3_style, "d_h3_class")
h3 = lrelu(h3)
h4 = conv2d_specNorm(h3, self.df_dim * 16, name='d_h4_conv')
h4, h4_mean, h4_var = instance_norm(h4, 'd_in4', True)
h4_mean = tf.reshape(h4_mean, (batch_size, self.df_dim * 16))
h4_var = tf.reshape(h4_var, (batch_size, self.df_dim * 16))
d_h4_style = tf.concat([h4_mean, h4_var], 0)
d_h4, d_h4_logits = self.linear_classifier(d_h4_style, "d_h4_class")
h4 = lrelu(h4)
# Returning logits to determine whether the images are real or fake
h5 = linear(slim.flatten(h4), 1, 'd_h5_lin')
# Recognition network for latent variables has an additional layer
encoder = lrelu((linear(slim.flatten(h4), 128, 'd_latent')))
cont_vars = linear(encoder, cont_dim, "d_latent_prediction")
return tf.nn.sigmoid(h5), h5, tf.nn.tanh(cont_vars), d_h1_logits, d_h2_logits, d_h3_logits, d_h4_logits
def generator_AdaIN(self, z, view_in, reuse=False):
batch_size = tf.shape(z)[0]
s_h, s_w, s_d = 64, 64, 64
# s_h, s_w, s_d = 112, 112, 112
s_h2, s_w2, s_d2 = conv_out_size_same(s_h, 2), conv_out_size_same(s_w, 2), conv_out_size_same(s_d, 2)
s_h4, s_w4, s_d4 = conv_out_size_same(s_h2, 2), conv_out_size_same(s_w2, 2), conv_out_size_same(s_d2, 2)
s_h8, s_w8, s_d8 = conv_out_size_same(s_h4, 2), conv_out_size_same(s_w4, 2), conv_out_size_same(s_d4, 2)
s_h16, s_w16, s_d16 = conv_out_size_same(s_h8, 2), conv_out_size_same(s_w8, 2), conv_out_size_same(s_d8, 2)
with tf.variable_scope("generator") as scope:
if reuse:
scope.reuse_variables()
# A learnt constant "template"
with tf.variable_scope('g_w_constant'):
w = tf.get_variable('w', [s_h16, s_w16, s_d16, self.gf_dim * 8],
initializer=tf.random_normal_initializer(stddev=0.02))
w_tile = tf.tile(tf.expand_dims(w, 0), (batch_size, 1, 1, 1, 1))
s0, b0 = self.z_mapping_function(z, self.gf_dim * 8, 'g_z0')
h0 = AdaIn(w_tile, s0, b0)
h0 = tf.nn.relu(h0)
h1 = deconv3d(h0, [batch_size, s_h8, s_w8, s_d8, self.gf_dim * 2], k_h=3, k_d=3, k_w=3, name='g_h1')
s1, b1 = self.z_mapping_function(z, self.gf_dim * 2, 'g_z1')
h1 = AdaIn(h1, s1, b1)
h1 = tf.nn.relu(h1)
h2 = deconv3d(h1, [batch_size, s_h4, s_w4, s_d4, self.gf_dim * 1], k_h=3, k_d=3, k_w=3, name='g_h2')
s2, b2 = self.z_mapping_function(z, self.gf_dim * 1, 'g_z2')
h2 = AdaIn(h2, s2, b2)
h2 = tf.nn.relu(h2)
# =============================================================================================================
h2_rotated = tf_3D_transform(h2, view_in, 16, 16)
h2_rotated = transform_voxel_to_match_image(h2_rotated)
# =============================================================================================================
# Collapsing depth dimension
h2_2d = tf.reshape(h2_rotated, [batch_size, s_h4, s_w4, 16 * self.gf_dim])
# 1X1 convolution
h3 = deconv2d(h2_2d, [batch_size, s_h4, s_w4, self.gf_dim * 16], k_h=1, k_w=1, d_h=1, d_w=1, name='g_h3')
h3 = tf.nn.relu(h3)
# =============================================================================================================
h4 = deconv2d(h3, [batch_size, s_h2, s_w2, self.gf_dim * 4], k_h=4, k_w=4, name='g_h4')
s4, b4 = self.z_mapping_function(z, self.gf_dim * 4, 'g_z4')
h4 = AdaIn(h4, s4, b4)
h4 = tf.nn.relu(h4)
h5 = deconv2d(h4, [batch_size, s_h, s_w, self.gf_dim], k_h=4, k_w=4, name='g_h5')
s5, b5 = self.z_mapping_function(z, self.gf_dim, 'g_z5')
h5 = AdaIn(h5, s5, b5)
h5 = tf.nn.relu(h5)
h6 = deconv2d(h5, [batch_size, s_h, s_w, self.c_dim], k_h=4, k_w=4, d_h=1, d_w=1, name='g_h6')
output = tf.nn.tanh(h6, name="output")
return output
def generator_AdaIN_res128(self, z, view_in, reuse=False):
batch_size = tf.shape(z)[0]
s_h, s_w, s_d = 64, 64, 64
s_h2, s_w2, s_d2 = conv_out_size_same(s_h, 2), conv_out_size_same(s_w, 2), conv_out_size_same(s_d, 2)
s_h4, s_w4, s_d4 = conv_out_size_same(s_h2, 2), conv_out_size_same(s_w2, 2), conv_out_size_same(s_d2, 2)
s_h8, s_w8, s_d8 = conv_out_size_same(s_h4, 2), conv_out_size_same(s_w4, 2), conv_out_size_same(s_d4, 2)
s_h16, s_w16, s_d16 = conv_out_size_same(s_h8, 2), conv_out_size_same(s_w8, 2), conv_out_size_same(s_d8, 2)
with tf.variable_scope("generator") as scope:
if reuse:
scope.reuse_variables()
# A learnt constant "template"
with tf.variable_scope('g_w_constant'):
w = tf.get_variable('w', [s_h16, s_w16, s_d16, self.gf_dim * 8],
initializer=tf.random_normal_initializer(stddev=0.02))
w_tile = tf.tile(tf.expand_dims(w, 0),
(batch_size, 1, 1, 1, 1)) # Repeat the learnt constant features to make a batch
s0, b0 = self.z_mapping_function(z, self.gf_dim * 8, 'g_z0')
h0 = AdaIn(w_tile, s0, b0)
h0 = lrelu(h0)
h1 = deconv3d(h0, [batch_size, s_h8, s_w8, s_d8, self.gf_dim * 4], k_h=3, k_w=3, k_d=3, name='g_h1')
s1, b1 = self.z_mapping_function(z, self.gf_dim * 4, 'g_z1')
h1 = AdaIn(h1, s1, b1)
h1 = lrelu(h1)
h2 = deconv3d(h1, [batch_size, s_h4, s_w4, s_d4, self.gf_dim * 2], k_h=3, k_w=3, k_d=3, name='g_h2')
s2, b2 = self.z_mapping_function(z, self.gf_dim * 2, 'g_z2')
h2 = AdaIn(h2, s2, b2)
h2 = lrelu(h2)
# =============================================================================================================
h2_rotated = tf_3D_transform(h2, view_in, 16, 16)
h2_rotated = transform_voxel_to_match_image(h2_rotated)
h2_proj1 = deconv3d(h2_rotated, [batch_size, s_h4, s_w4, s_d4, self.gf_dim * 1], k_h=3, k_w=3, k_d=3, d_h=1,
d_w=1, d_d=1, name='g_h2_proj1')
h2_proj1 = lrelu(h2_proj1)
h2_proj2 = deconv3d(h2_proj1, [batch_size, s_h4, s_w4, s_d4, self.gf_dim], k_h=3, k_w=3, k_d=3, d_h=1,
d_w=1, d_d=1, name='g_h2_proj2')
h2_proj2 = lrelu(h2_proj2)
# =============================================================================================================
# Collapsing depth dimension
h2_2d = tf.reshape(h2_proj2, [batch_size, s_h4, s_w4, s_d4 * self.gf_dim])
# 1X1 convolution
h3 = deconv2d(h2_2d, [batch_size, s_h4, s_w4, self.gf_dim * 16 // 2], k_h=1, k_w=1, d_h=1, d_w=1,
name='g_h3')
h3 = lrelu(h3)
# =============================================================================================================
h4 = deconv2d(h3, [batch_size, s_h2, s_w2, self.gf_dim * 4], k_h=4, k_w=4, name='g_h4')
s4, b4 = self.z_mapping_function(z, self.gf_dim * 4, 'g_z4')
h4 = AdaIn(h4, s4, b4)
h4 = lrelu(h4)
h5 = deconv2d(h4, [batch_size, s_h, s_w, self.gf_dim], k_h=4, k_w=4, name='g_h5')
s5, b5 = self.z_mapping_function(z, self.gf_dim, 'g_z5')
h5 = AdaIn(h5, s5, b5)
h5 = lrelu(h5)
h6 = deconv2d(h5, [batch_size, s_h * 2, s_w * 2, self.gf_dim // 2], k_h=4, k_w=4, name='g_h6')
s6, b6 = self.z_mapping_function(z, self.gf_dim // 2, 'g_z6')
h6 = AdaIn(h6, s6, b6)
h6 = lrelu(h6)
h7 = deconv2d(h6, [batch_size, s_h * 2, s_w * 2, self.c_dim], k_h=4, k_w=4, d_h=1, d_w=1, name='g_h7')
output = tf.nn.tanh(h7, name="output")
return output
# =======================================================================================================================
@property
def model_dir(self):
return "{}_{}_{}".format(
self.dataset_name,
self.output_height, self.output_width)
def save(self, checkpoint_dir, step):
model_name = "HoloGAN.model"
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
self.saver.save(self.sess,
os.path.join(checkpoint_dir, model_name),
global_step=step)
def load(self, checkpoint_dir):
import re
print(" [*] Reading checkpoints...")
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
ckpt = tf.train_on_loader.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
counter = int(next(re.finditer("(\d+)(?!.*\d)", ckpt_name)).group(0))
print(" [*] Success to read {}".format(ckpt_name))
return True, counter
else:
print(" [*] Failed to find a checkpoint")
return False, 0