forked from MahdiyarMM/MSG-CapsGAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
msg_capsgan.py
546 lines (439 loc) · 19.6 KB
/
msg_capsgan.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
#Image lib
import cv2
# math libraries
import numpy as np
import scipy.misc
# ml libraries
import tensorflow as tf
from keras import layers, models, optimizers
from keras import backend as K
from keras.utils import to_categorical
from keras.datasets import mnist, cifar10
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, Lambda, concatenate, Multiply
from keras.layers import BatchNormalization, Activation, ZeroPadding2D, Add
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.preprocessing.image import ImageDataGenerator
from keras import callbacks
import datetime
# visualization
import skimage
from skimage import data, color, exposure
from skimage.transform import resize
import matplotlib.pyplot as plt
import math
from skimage.measure import compare_ssim as ssim
from skimage.measure import compare_psnr as psnr
from keras.applications import VGG19
# sys and helpers
import sys
import os
import glob
from tqdm import tqdm
print('Modules imported.')
class DataLoader3_all():
def __init__(self, dataset_name, img_res=(128, 128)):
self.dataset_name = dataset_name
self.img_res = img_res
def load_data(self, batch_size=1, is_testing=False):
data_type = "train" if not is_testing else "test"
path = glob.glob('/content/drive/My Drive/CelebA_sample/CelebA/*')
batch_images = np.random.choice(path, size=batch_size)
imgs_hr = []
imgs_lr = []
imgs_32 = []
imgs_64 = []
for img_path in batch_images:
imgl = self.imread(img_path)
img = imgl[20:218-20,:]
h, w = self.img_res
low_h, low_w = int(h / 8), int(w / 8)
# img_hr = scipy.misc.imresize(img, self.img_res)
# img_lr = scipy.misc.imresize(img, (low_h, low_w))
# img_32 = scipy.misc.imresize(img, (32, 32))
# img_64 = scipy.misc.imresize(img, (64, 64))
img_hr = cv2.resize(img, self.img_res)
img_lr = cv2.resize(img, (low_h, low_w))
img_32 = cv2.resize(img, (32, 32))
img_64 = cv2.resize(img, (64, 64))
# If training => do random flip
if not is_testing and np.random.random() < 0.5:
img_hr = np.fliplr(img_hr)
img_lr = np.fliplr(img_lr)
img_32 = np.fliplr(img_32)
img_64 = np.fliplr(img_64)
imgs_hr.append(img_hr)
imgs_lr.append(img_lr)
imgs_32.append(img_32)
imgs_64.append(img_64)
imgs_hr = np.array(imgs_hr) / 127.5 - 1.
imgs_lr = np.array(imgs_lr) / 127.5 - 1.
imgs_32 = np.array(imgs_32) / 127.5 - 1.
imgs_64 = np.array(imgs_64) / 127.5 - 1.
return imgs_hr, imgs_lr , imgs_32, imgs_64
def imread(self, path):
return plt.imread(path).astype(np.float)
def load_test(self,path):
img = self.imread(path)
img_hr = scipy.misc.imresize(img, (256,256))
# Configure data loader
dataset_name = 'img_align_celeba'
hr_height = 128
hr_width = 128
data_loader = DataLoader3_all(dataset_name=dataset_name,
img_res=(hr_height, hr_width))
def all_psnr(imageA, imageB):
psnrs = []
for ii in range(len(imageA)):
psnrs.append(psnr(imageA[ii],imageB[ii]))
return np.mean(psnrs)
def all_ssim(imageA, imageB):
psnrs = []
for ii in range(len(imageA)):
psnrs.append(ssim(imageA[ii],imageB[ii],multichannel =True))
return np.mean(psnrs)
# squash function of capsule layers, borrowed from Xifeng Guo's implementation of Keras CapsNet `https://github.com/XifengGuo/CapsNet-Keras`
def squash(vectors, axis=-1):
"""
The non-linear activation used in Capsule. It drives the length of a large vector to near 1 and small vector to 0
:param vectors: some vectors to be squashed, N-dim tensor
:param axis: the axis to squash
:return: a Tensor with same shape as input vectors
"""
s_squared_norm = K.sum(K.square(vectors), axis, keepdims=True)
scale = s_squared_norm / (1 + s_squared_norm) / K.sqrt(s_squared_norm + K.epsilon())
return scale * vectors
# device check
from tensorflow.python.client import device_lib
print('Devices:', device_lib.list_local_devices())
# GPU check
if not tf.test.gpu_device_name():
print('No GPU found.')
else:
print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
def build_vgg(hr_shape):
"""
Builds a pre-trained VGG19 model that outputs image features extracted at the
third block of the model
"""
vgg = VGG19(weights="imagenet")
# Set outputs to outputs of last conv. layer in block 3
# See architecture at: https://github.com/keras-team/keras/blob/master/keras/applications/vgg19.py
vgg.outputs = [vgg.layers[9].output]
img = Input(hr_shape)
# Extract image features
img_features = vgg(img)
return Model(img, img_features)
optimizer = Adam(0.0003, 0.5)
vgg = build_vgg((128,128,3))
vgg.trainable = False
vgg.compile(loss='mse',
optimizer=optimizer,
metrics=['accuracy'])
vgg64 = build_vgg((64,64,3))
vgg64.trainable = False
vgg64.compile(loss='mse',
optimizer=optimizer,
metrics=['accuracy'])
vgg32 = build_vgg((32,32,3))
vgg32.trainable = False
vgg32.compile(loss='mse',
optimizer=optimizer,
metrics=['accuracy'])
# discriminator structure
def build_discriminator():
# depending on dataset we define input shape for our network
img = Input(shape=(128, 128, 3))
img64 = Input(shape=(64, 64, 3))
img32 = Input(shape=(32, 32, 3))
x1 = Conv2D(filters=128, kernel_size=9, strides=2, padding='same', name='conv0')(img)
img64f = Conv2D(filters=128, kernel_size=9, strides=1, padding='same', name='convimg0')(img64)
x2 = concatenate([x1,img64f],axis=-1)
x3 = Conv2D(filters=128, kernel_size=9, strides=2, padding='same', name='conv1')(x2)
img32f = Conv2D(filters=128, kernel_size=9, strides=1, padding='same', name='convimg1')(img32)
x4 = concatenate([x3,img32f],axis=-1)
# first typical convlayer outputs a 20x20x256 matrix
x5 = Conv2D(filters=256, kernel_size=9, strides=1, padding='valid', name='conv2')(x4)
x6 = LeakyReLU()(x5)
# original 'Dynamic Routing Between Capsules' paper does not include the batch norm layer after the first conv group
x = BatchNormalization(momentum=0.8)(x6)
#
# primarycaps coming first
#
# filters 256 (n_vectors=8 * channels=32)
x = Conv2D(filters=8 * 32, kernel_size=9, strides=2, padding='valid', name='primarycap_conv2')(x)
#Flatten to add dense layer
x = Flatten()(x)
# reshape into the 8D vector for all 32 feature maps combined
# (primary capsule has collections of activations which denote orientation of the digit
# while intensity of the vector which denotes the presence of the digit)
x = Reshape(target_shape=[-1, 8], name='primarycap_reshape')(x)
# the purpose is to output a number between 0 and 1 for each capsule where the length of the input decides the amount
x = Lambda(squash, name='primarycap_squash')(x)
x = BatchNormalization(momentum=0.8)(x)
#
# digitcaps are here
#
x = Flatten()(x)
uhat = Dense(160, kernel_initializer='he_normal', bias_initializer='zeros', name='uhat_digitcaps')(x)
# c = coupling coefficient (softmax over the bias weights, log prior) | "the coupling coefficients between capsule (i) and all the capsules in the layer above sum to 1"
# we treat the coupling coefficiant as a softmax over bias weights from the previous dense layer
c = Activation('softmax', name='softmax_digitcaps1')(uhat) # softmax will make sure that each weight c_ij is a non-negative number and their sum equals to one
# s_j (output of the current capsule level) = uhat * c
c = Dense(160)(c) # compute s_j
x = Multiply()([uhat, c])
"""
NOTE: Squashing the capsule outputs creates severe blurry artifacts, thus we replace it with Leaky ReLu.
"""
s_j = LeakyReLU()(x)
#
# we will repeat the routing part 2 more times (num_routing=3) to unfold the loop
#
c = Activation('softmax', name='softmax_digitcaps2')(s_j) # softmax will make sure that each weight c_ij is a non-negative number and their sum equals to one
c = Dense(160)(c) # compute s_j
x = Multiply()([uhat, c])
s_j = LeakyReLU()(x)
c = Activation('softmax', name='softmax_digitcaps3')(s_j) # softmax will make sure that each weight c_ij is a non-negative number and their sum equals to one
c = Dense(160)(c) # compute s_j
x = Multiply()([uhat, c])
s_j = LeakyReLU()(x)
pred = Dense(25*25, activation='sigmoid')(s_j)
patch = Reshape(target_shape=[-1, 25], name='last_out_reshape')(pred)
return Model([img,img64,img32], patch)
# build and compile the discriminator
discriminator = build_discriminator()
print('DISCRIMINATOR:')
discriminator.summary()
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
#discriminator.load_weights('images_MSG/wdis6400.h5')
discriminator.compile(loss='binary_crossentropy', optimizer=sgd)
# generator structure
def build_generator():
from keras.applications import VGG19
vgg = VGG19(weights="imagenet")
vgg.outputs = [vgg.layers[9].output]
vgg.trainable =False
def deconv2d(layer_input,num=256):
"""Layers used during upsampling"""
u = UpSampling2D(size=2)(layer_input)
u = Conv2D(num, kernel_size=3, strides=1, padding='same')(u)
u = Activation('relu')(u)
return u
def residual_block(layer_input, filters):
"""Residual block described in paper"""
d = Conv2D(filters, kernel_size=3, strides=1, padding='same')(layer_input)
d = Activation('relu')(d)
d = BatchNormalization(momentum=0.8)(d)
d = Conv2D(filters, kernel_size=3, strides=1, padding='same')(d)
d = BatchNormalization(momentum=0.8)(d)
d = Add()([d, layer_input])
return d
img_lr_in = Input(shape=(16,16,3))
img_bilinear = Lambda(lambda image: tf.image.resize_images(image,(128, 128),method = tf.image.ResizeMethod.BICUBIC,align_corners = True, preserve_aspect_ratio = True))(img_lr_in)
img_bilinear_64 = Lambda(lambda image: tf.image.resize_images(image,(64, 64),method = tf.image.ResizeMethod.BICUBIC,align_corners = True, preserve_aspect_ratio = True))(img_lr_in)
img_bilinear_32 = Lambda(lambda image: tf.image.resize_images(image,(32, 32),method = tf.image.ResizeMethod.BICUBIC,align_corners = True, preserve_aspect_ratio = True))(img_lr_in)
vgg_features = vgg(img_bilinear)
vgg_features_cnn = Conv2D(64, kernel_size=3, strides=1, padding='same',name = "vgg_features_cnn")(vgg_features)
pd = Conv2D(64, kernel_size=3, strides=1, padding='same')(img_lr_in)
x = residual_block(pd,64)
x = deconv2d(x,num=64)
xconcat = concatenate([x,vgg_features_cnn],axis=-1,name='concatenation')
pd = Conv2D(64, kernel_size=3, strides=1, padding='same')(xconcat)
x = residual_block(pd,64)
x_out_32 = Conv2D(3, kernel_size=3, strides=1, padding='same')(x)
x_out_32 = Activation('tanh')(x_out_32)
x_32_add = Add()([x_out_32, img_bilinear_32])
x_32 = Activation('tanh')(x_32_add)
x = deconv2d(x)
pd = Conv2D(64, kernel_size=3, strides=1, padding='same')(x)
x = residual_block(pd,64)
x_out_64 = Conv2D(3, kernel_size=3, strides=1, padding='same')(x)
x_out_64 = Activation('tanh')(x_out_64)
x_64_add = Add()([x_out_64, img_bilinear_64])
x_64 = Activation('tanh')(x_64_add)
x = deconv2d(x)
x = Conv2D(3, kernel_size=3, strides=1, padding='same')(x)
x_out_pre = Activation('tanh')(x)
x_out_add = Add()([x_out_pre, img_bilinear])
x_out = Activation('tanh')(x_out_add)
return Model([img_lr_in],[x_out,x_64,x_32])
# build and compile the generator
generator = build_generator()
print('GENERATOR:')
generator.summary()
#generator.load_weights('images_MSG/wgen6400.h5')
generator.compile(loss='binary_crossentropy', optimizer=Adam(0.0002, 0.5))
model_json = generator.to_json()
with open("model.json", "w") as json_file:
json_file.write(model_json)
model_json = generator.to_json()
with open("model.json", "w") as json_file:
json_file.write(model_json)
# feeding noise to generator
z = Input(shape=(16,16,3))
imgf = generator(z)
fake_features = vgg(imgf[0])
fake_features_1 = vgg64(imgf[1])
fake_features_2 = vgg32(imgf[2])
# for the combined model we will only train the generator
discriminator.trainable = False
# try to discriminate generated images
valid = discriminator(imgf)
# the combined model (stacked generator and discriminator) takes
# noise as input => generates images => determines validity
combined = Model(z, [valid,fake_features,fake_features_1,fake_features_2])
print('COMBINED:')
combined.summary()
combined.compile(loss=['binary_crossentropy', 'mse', 'mse', 'mse'],
loss_weights=[3e-3, 0.001,0.001,1],
optimizer=Adam(0.0002, 0.5))
# loss values for further plotting
D_L_REAL = []
D_L_FAKE = []
D_L = []
D_ACC = []
G_L = []
def train(dataset_title, epochs, batch_size=32, save_interval=50):
start_time = datetime.datetime.now()
half_batch = int(batch_size / 2)
for epoch in range(0,50000):
'''
#Progressive Weight Loss Adjustment
if epoch == 600:
combined.compile(loss=['binary_crossentropy', 'mse'],
loss_weights=[1e-2, 1],
optimizer=Adam(0.0002, 0.5))
# ---------------------
if epoch == 1500:
combined.compile(loss=['binary_crossentropy', 'mse'],
loss_weights=[1e-1, 1],
optimizer=Adam(0.0002, 0.5))
if epoch == 2500:
combined.compile(loss=['binary_crossentropy', 'mse'],
loss_weights=[1e-2, 1],
optimizer=Adam(0.0002, 0.5))
if epoch == 3500:
combined.compile(loss=['binary_crossentropy', 'mse'],
loss_weights=[1e-1, 1],
optimizer=Adam(0.0002, 0.5))
'''
# Train Discriminator
# ---------------------
# select a random half batch of images
imgs_hr, imgs_lr, imgs_32, imgs_64 = data_loader.load_data(half_batch)
# generate a half batch of new images
gen_imgs = generator.predict(imgs_lr)
# train the discriminator by feeding both real and fake (generated) images one by one
d_loss_real = discriminator.train_on_batch([imgs_hr,imgs_64,imgs_32], np.random.uniform(low=0.8, high=1.2, size=(half_batch, 25,25))) # 0.9 for label smoothing
d_loss_fake = discriminator.train_on_batch(gen_imgs, np.random.uniform(low=0.0, high=0.3, size=(half_batch, 25,25)))
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# ---------------------
# Train Generator
# ---------------------
imgs_hr, imgs_lr, imgs_32, imgs_64 = data_loader.load_data(batch_size)
# the generator wants the discriminator to label the generated samples
# as valid (ones)
image_features = vgg.predict(imgs_hr)
image_features_1 = vgg64.predict(imgs_64)
image_features_2 = vgg32.predict(imgs_32)
# train the generator
g_loss = combined.train_on_batch(imgs_lr, [np.ones((batch_size,25,25)),image_features,image_features_1,image_features_2])
elapsed_time = datetime.datetime.now() - start_time
# Plot the progress
print ("%d time: %s ,[D loss: %f, acc.: %.2f%%] [G loss: %f]" % (epoch,elapsed_time, d_loss, 100*d_loss, g_loss[0]))
D_L_REAL.append(d_loss_real)
D_L_FAKE.append(d_loss_fake)
D_L.append(d_loss)
D_ACC.append(d_loss)
G_L.append(g_loss)
# if at save interval => save generated image samples
if epoch % save_interval == 0:
sample_images(epoch)
generator.save_weights("images_MSG/wgen%d.h5" % (epoch))
discriminator.save_weights("images_MSG/wdis%d.h5" % (epoch))
def sample_images(epoch):
os.makedirs('images_MSG/%s' % dataset_name, exist_ok=True)
r, c = 2, 5
imgs_hr, imgs_lr,_,_ = data_loader.load_data(batch_size=2, is_testing=True)
fake_hr = generator.predict(imgs_lr)
imgs_hr_inp = imgs_hr
imgs_lr_inp = imgs_lr
# Rescale images 0 - 1
imgs_lr = 0.5 * imgs_lr + 0.5
fake_hr[0] = 0.5 * fake_hr[0] + 0.5
fake_hr[1] = 0.5 * fake_hr[1] + 0.5
fake_hr[2] = 0.5 * fake_hr[2] + 0.5
imgs_hr = 0.5 * imgs_hr + 0.5
# Save generated images and the high resolution originals
titles = ['Original','128x128']
fig, axs = plt.subplots(r, c)
cnt = 0
for row in range(r):
for col, image in enumerate([imgs_hr,fake_hr[0]]):
axs[row, col+3].imshow(imgs_lr[row])
axs[row, col+2].imshow(fake_hr[2][row])
axs[row, col+1].imshow(fake_hr[1][row])
axs[row, col].imshow(image[row])
axs[row, col+1].set_title('64x64')
axs[row, col+2].set_title('32x32')
axs[row, col+3].set_title('input')
axs[row, col].set_title(titles[col])
axs[row, col].axis('off')
axs[row, col+1].axis('off')
axs[row, col+2].axis('off')
axs[row, col+3].axis('off')
cnt += 1
txt = ("psnr = %f - ssim = %f" % (all_psnr(fake_hr[0],imgs_hr), all_ssim(fake_hr[0],imgs_hr)))
fig.text(.5, .05, txt, ha='center')
fig.savefig("images_MSG/%s/%d.png" % (dataset_name, epoch))
plt.close()
# Save low resolution images for comparison
for i in range(r):
fig = plt.figure()
plt.imshow(imgs_lr[i])
fig.savefig('images_MSG/%s/%d_lowres%d.png' % (dataset_name, epoch, i))
plt.close()
print(all_psnr(fake_hr[0],imgs_hr))
print(all_ssim(fake_hr[0],imgs_hr))
if epoch<5000 and epoch>1000:
combined.compile(loss=['binary_crossentropy', 'mse', 'mse', 'mse'],
loss_weights=[3e-3, 0.001,(epoch - 999)/4000,1],
optimizer=Adam(0.0002, 0.5))
if epoch<10000 and epoch>5000:
combined.compile(loss=['binary_crossentropy', 'mse', 'mse', 'mse'],
loss_weights=[3e-3, (epoch - 4999)/5000,1,1],
optimizer=Adam(0.0002, 0.5))
if epoch<15000 and epoch>10000:
combined.compile(loss=['binary_crossentropy', 'mse', 'mse', 'mse'],
loss_weights=[3e-3 + 2e-3*(epoch - 9999)/5000, 1,1,1],
optimizer=Adam(0.0002, 0.5))
return imgs_lr_inp,imgs_hr_inp
history = train('cifar10', epochs=30000, batch_size=32, save_interval=50)
#generator.save('mnist_model.h5')
#generator.save('cifar10_model.h5')
plt.plot(D_L)
plt.title('Discriminator results (MNIST)')
plt.xlabel('Epochs')
plt.ylabel('Discriminator Loss (blue), Discriminator Accuracy (orange)')
plt.legend(['Discriminator Loss', 'Discriminator Accuracy'])
plt.show()
plt.plot(G_L)
plt.title('Generator results (MNIST)')
plt.xlabel('Epochs')
plt.ylabel('Generator Loss (blue)')
plt.legend('Generator Loss')
plt.show()
plt.plot(D_L)
plt.title('Discriminator results (CIFAR10)')
plt.xlabel('Epochs')
plt.ylabel('Discriminator Loss (blue), Discriminator Accuracy (orange)')
plt.legend(['Discriminator Loss', 'Discriminator Accuracy'])
plt.show()
plt.plot(G_L)
plt.title('Generator results (CIFAR10)')
plt.xlabel('Epochs')
plt.ylabel('Generator Loss (blue)')
plt.legend('Generator Loss')
plt.show()