-
Notifications
You must be signed in to change notification settings - Fork 0
/
improved_wgan.py
234 lines (202 loc) · 8.08 KB
/
improved_wgan.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
'''
This is the training code for the WGAN-GP.
This is largely copied from https://github.com/eriklindernoren/Keras-GAN/blob/master/wgan_gp/wgan_gp.py
I have changed the optimiser, the architectures of the generator and discriminator, the dataset used,
and otherwise streamlined the code somewhat and made it fit with my setup.
'''
from keras.layers.merge import _Merge
from keras.layers import Input
from keras.models import Model
from keras.optimizers import Adam
from functools import partial
import keras.backend as K
from data_prep import prepare_images, prepare_mnist, prepare_cifar10, prepare_anime_images
from alt_gen import make_mnist_generator, make_cifar_generator, make_anime_generator
from generator import make_generator
from discriminator import make_discriminator, make_poke_discriminator
import matplotlib.pyplot as plt
import os
import sys
import numpy as np
# parameters to tune
MAX_ITERATIONS = 100000
N_CRITIC = 5
BATCH_SIZE = 64
SAMPLE_INTERVAL = 50
LOG_FILE = 'logs/dummy_logs.txt'
CRITIC_WEIGHTS_SAVE_LOC = 'weights/imp_wgan_dummy_critic.h5'
GENERATOR_WEIGHTS_SAVE_LOC = 'weights/imp_wgan_dummy_gen.h5'
# the below should be a folder
IMAGES_SAVE_DIR = "results"
GRADIENT_PENALTY_WEIGHT = 10
mode = 'pokemon'
if len(sys.argv) > 1:
if sys.argv[1] == 'mnist':
mode = 'mnist'
elif sys.argv[1] == 'cifar':
mode = 'cifar'
elif sys.argv[1] == 'anime':
mode = 'anime'
elif sys.argv[1] == 'pokemon-alt':
mode = 'alt'
class RandomWeightedAverage(_Merge):
"""Provides a (random) weighted average between real and generated image samples"""
def _merge_function(self, inputs):
alpha = K.random_uniform((BATCH_SIZE, 1, 1, 1))
return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])
def gradient_penalty_loss(y_true, y_pred, averaged_samples, gradient_penalty_weight):
"""
Computes gradient penalty based on prediction and weighted real / fake samples
"""
gradients = K.gradients(y_pred, averaged_samples)[0]
# compute the euclidean norm by squaring ...
gradients_sqr = K.square(gradients)
# ... summing over the rows ...
gradients_sqr_sum = K.sum(gradients_sqr,
axis=np.arange(1, len(gradients_sqr.shape)))
# ... and sqrt
gradient_l2_norm = K.sqrt(gradients_sqr_sum)
# compute lambda * (1 - ||grad||)^2 still for each single sample
gradient_penalty = gradient_penalty_weight * K.square(1 - gradient_l2_norm)
# return the mean as loss over all the batch samples
return K.mean(gradient_penalty)
def sample_images(g_model, epoch, noise, image_save_dir, greyscale=False):
r, c = 5, 5
gen_imgs = g_model.predict(noise)
# Rescale images 0 - 1
gen_imgs = 0.5 * (gen_imgs + 1)
gen_imgs = np.clip(gen_imgs, 0, 1)
fig, axs = plt.subplots(r, c)
cnt = 0
for i in range(r):
for j in range(c):
if greyscale:
axs[i,j].imshow(gen_imgs[cnt, :, :, 0], cmap='gray')
else:
axs[i,j].imshow(gen_imgs[cnt, :, :, :])
axs[i,j].axis('off')
cnt += 1
fig.savefig(os.path.join(image_save_dir, "image_%d.png" % epoch))
plt.close()
def wasserstein_loss(y_true, y_pred):
return K.mean(y_true * y_pred)
def mean_loss(y_true, y_pred):
return K.mean(y_pred)
optimizer = Adam(0.0005, beta_1=0.0, beta_2=0.9999)
input_dim = 100
# Build the generator and critic
generator = make_generator()
image_shape = (128, 128, 3)
if mode == 'mnist':
generator = make_mnist_generator()
image_shape = (28, 28, 1)
elif mode == 'cifar':
generator = make_cifar_generator()
image_shape = (32, 32, 3)
elif mode == 'anime':
generator = make_anime_generator()
image_shape = (48, 48, 3)
input_dim = 40
elif mode == 'alt':
image_shape = (48, 48, 3)
generator = make_anime_generator()
input_dim = 40
CONST_NOISE = np.random.normal(0, 1, (25, input_dim))
# we currently use the same discriminator across all
critic = make_discriminator(image_shape, batchnorm=False)
if mode == 'pokemon':
optimizer = Adam(0.0001, beta_1=0.5, beta_2=0.9)
critic = make_poke_discriminator(image_shape, batchnorm=False)
# Freeze generator's layers while training critic
generator.trainable = False
for l in generator.layers:
l.trainable = False
# Image input (real sample)
real_img = Input(shape=image_shape)
# Noise input
z_disc = Input(shape=(input_dim,))
# Generate image based of noise (fake sample)
fake_img = generator(z_disc)
# Discriminator determines validity of the real and fake images
fake = critic(fake_img)
valid = critic(real_img)
# Construct weighted average between real and fake images
interpolated_img = RandomWeightedAverage()([real_img, fake_img])
# Determine validity of weighted sample
validity_interpolated = critic(interpolated_img)
# Use Python partial to provide loss function with additional
# 'averaged_samples' argument
partial_gp_loss = partial(gradient_penalty_loss,
averaged_samples=interpolated_img,
gradient_penalty_weight=GRADIENT_PENALTY_WEIGHT)
partial_gp_loss.__name__ = 'gradient_penalty' # Keras requires function name
critic_model = Model(inputs=[real_img, z_disc],
outputs=[valid, fake, validity_interpolated])
critic_model.compile(loss=[wasserstein_loss,
wasserstein_loss,
partial_gp_loss],
optimizer=optimizer)
# For the generator we freeze the critic's layers
critic.trainable = False
for l in critic.layers:
l.trainable = False
generator.trainable = True
for l in generator.layers:
l.trainable = True
# Sampled noise for input to generator
z_gen = Input(shape=(input_dim,))
# Generate images based of noise
img = generator(z_gen)
# Discriminator determines validity
gen_crit_out = critic(img)
# Defines generator model
generator_model = Model(inputs=[z_gen], outputs=[gen_crit_out])
generator_model.compile(loss=wasserstein_loss, optimizer=optimizer)
with open(LOG_FILE, 'w') as f:
f.write("")
# Load the dataset
prepare_function = lambda x: prepare_images('data', x, (image_shape[0], image_shape[1])) # TODO fix this!
if mode == 'mnist':
prepare_function = prepare_mnist
elif mode == 'cifar':
prepare_function = prepare_cifar10
elif mode == 'anime':
prepare_function = lambda x: prepare_anime_images('data', x, (image_shape[0], image_shape[1]))
datagen = prepare_function(BATCH_SIZE)
# Adversarial ground truths
valid = np.ones((BATCH_SIZE, 1), dtype=np.float32)
fake = -valid
dummy = -np.zeros((BATCH_SIZE, 1), dtype=np.float32) # Dummy gt for gradient penalty
a_epoch = 0
for epoch in range(0, MAX_ITERATIONS+1):
for _ in range(N_CRITIC):
# get real images
try:
imgs = next(datagen)[0]
except StopIteration:
datagen = prepare_function(BATCH_SIZE)
imgs = next(datagen)[0]
# if we run out of data, generate more.
if imgs.shape[0] != BATCH_SIZE:
datagen = prepare_function(BATCH_SIZE)
imgs = next(datagen)[0]
imgs = (imgs.astype(np.float32) - 0.5) * 2.0
# Sample generator input
noise = np.random.normal(0, 1, (BATCH_SIZE, input_dim))
# Train the critic
d_loss = critic_model.train_on_batch([imgs, noise],
[valid, fake, dummy])
g_loss = generator_model.train_on_batch(np.random.normal(0, 1, (BATCH_SIZE, input_dim)), valid)
# Plot the progress
print ("%d [D loss: %f] [G loss: %f]" % (epoch, d_loss[0], g_loss))
# we also record it in a file
with open(LOG_FILE, 'a+') as f:
f.write('%d %f %f\n' % (epoch, d_loss[0], g_loss))
# If at save interval => save generated image samples
if epoch % SAMPLE_INTERVAL == 0:
sample_images(generator, epoch, CONST_NOISE, IMAGES_SAVE_DIR, greyscale=(mode == 'mnist'))
critic.save_weights(CRITIC_WEIGHTS_SAVE_LOC)
generator.save_weights(GENERATOR_WEIGHTS_SAVE_LOC)
sample_images(generator, MAX_ITERATIONS, CONST_NOISE, IMAGES_SAVE_DIR, greyscale=(mode == 'mnist'))
critic.save_weights(CRITIC_WEIGHTS_SAVE_LOC)
generator.save_weights(GENERATOR_WEIGHTS_SAVE_LOC)