-
Notifications
You must be signed in to change notification settings - Fork 3
/
model.py
258 lines (181 loc) · 10.3 KB
/
model.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
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
import os
from glob import glob
import cv2
from utils.data_generator import Data_Generator
from dataset import Dataset_Color
from PIL import Image
class Generator():
'''
The pix2pix generator is a U-net, i.e. a variant of an auto-encoder where layers are symmetrically stacked one to the other.
It learns a mapping between gray scale to RGB space, conditionally to a grayscale image.
References:
https://arxiv.org/pdf/1611.07004.pdf
https://arxiv.org/pdf/1505.04597.pdf
'''
def generator(self,c):
with tf.variable_scope('generator'):
self.initializer = tf.truncated_normal_initializer(stddev=0.02)
#Encoder
enc0 = slim.conv2d(c,64,[3,3],padding="SAME",
biases_initializer=None,activation_fn= tf.nn.leaky_relu,
weights_initializer=self.initializer)
enc0 = tf.space_to_depth(enc0,2)
enc1 = slim.conv2d(enc0,128,[3,3],padding="SAME",
activation_fn= tf.nn.leaky_relu,normalizer_fn=slim.batch_norm,
weights_initializer=self.initializer)
enc1 = tf.space_to_depth(enc1,2)
enc2 = slim.conv2d(enc1,256,[3,3],padding="SAME",
normalizer_fn=slim.batch_norm,activation_fn= tf.nn.leaky_relu,
weights_initializer= self.initializer)
enc2 = tf.space_to_depth(enc2,2)
enc3 = slim.conv2d(enc2,512,[3,3],padding="SAME",
normalizer_fn=slim.batch_norm,activation_fn= tf.nn.leaky_relu,
weights_initializer=self.initializer)
enc3 = tf.space_to_depth(enc3,2)
#Decoder
gen0 = slim.conv2d(
enc3,num_outputs=512,kernel_size=[3,3],
padding="SAME",normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.elu, weights_initializer=self.initializer)
gen0 = tf.depth_to_space(gen0,2)
gen1 = slim.conv2d(
tf.concat([gen0,enc2],3),num_outputs=256,kernel_size=[3,3],
padding="SAME",normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.elu,weights_initializer=self.initializer)
gen1 = tf.depth_to_space(gen1,2)
gen2 = slim.conv2d(
tf.concat([gen1,enc1],3),num_outputs=128,kernel_size=[3,3],
padding="SAME",normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.elu,weights_initializer=self.initializer)
gen2 = tf.depth_to_space(gen2,2)
gen3 = slim.conv2d(
tf.concat([gen2,enc0],3),num_outputs=64,kernel_size=[3,3],
padding="SAME",normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.elu, weights_initializer=self.initializer)
gen3 = tf.depth_to_space(gen3,2)
g_out = slim.conv2d(
gen3,num_outputs=3,kernel_size=[1,1],padding="SAME",
biases_initializer=None,activation_fn=tf.nn.tanh,
weights_initializer=self.initializer)
return g_out
class Discriminator():
'''
The discriminator is a classical ConvNet classifier.
It learns to discriminate fake-colored image vs. originally colored images
'''
def discriminator(self, c, reuse = False):
with tf.variable_scope('discriminator'):
self.initializer = tf.truncated_normal_initializer(stddev=0.02)
conv1 = slim.conv2d(c,32,[3,3],padding="SAME",scope='d0',
biases_initializer=None,activation_fn= tf.nn.leaky_relu,stride=[2,2],
reuse=reuse,weights_initializer=self.initializer)
conv2 = slim.conv2d(conv1,64,[3,3],padding="SAME",scope='d1',
normalizer_fn=slim.batch_norm,activation_fn= tf.nn.leaky_relu,stride=[2,2],
reuse=reuse,weights_initializer=self.initializer)
conv3 = slim.conv2d(conv2,128,[3,3],padding="SAME",scope='d2',
normalizer_fn=slim.batch_norm,activation_fn= tf.nn.leaky_relu,stride=[2,2],
reuse=reuse,weights_initializer=self.initializer)
conv4 = slim.conv2d(conv3,256,[3,3],padding="SAME",scope='d3',
normalizer_fn=slim.batch_norm,activation_fn= tf.nn.leaky_relu,stride=[2,2],
reuse=reuse,weights_initializer=self.initializer)
dis_full = slim.fully_connected(slim.flatten(conv4),1024,activation_fn= tf.nn.leaky_relu,scope='dl',
reuse=reuse, weights_initializer=self.initializer)
d_out = slim.fully_connected(dis_full,1,activation_fn=tf.nn.sigmoid,scope='do',
reuse=reuse, weights_initializer=self.initializer)
return d_out
class GAN():
'''
The Generative Adversarial Network framework we used to train our color generator
'''
def __init__(self, config):
self.config = config
self.data_init()
self.model_init()
def data_init(self):
print("\nData init")
self.dataset = Dataset_Color(self.config)
self.data_generator = Data_Generator(self.config, self.dataset)
# self.prediction_dataset = Dataset_BW(self.config)
def model_init(self):
print("\Model init")
self.D = Discriminator()
self.G = Generator()
self.lambda_L1 = 100
self.condition_in = tf.placeholder(shape=[None,self.config.image_size,self.config.image_size,1],dtype=tf.float32)
self.real_in = tf.placeholder(shape=[None,self.config.image_size,self.config.image_size,3],dtype=tf.float32)
self.Gx = self.G.generator(self.condition_in)
self.Dx = self.D.discriminator(self.real_in)
self.Dg = self.D.discriminator(self.Gx,reuse=True)
self.d_loss = -tf.reduce_mean(tf.log(self.Dx) + tf.log(1.- self.Dg))
self.g_loss = -tf.reduce_mean(tf.log(self.Dg)) + self.lambda_L1 *tf.reduce_mean(tf.abs(self.Gx - self.real_in))
self.trainerD = tf.train.AdamOptimizer(learning_rate= self.config.lr_dis)
self.trainerG = tf.train.AdamOptimizer(learning_rate= self.config.lr_gen)
self.d_grads = self.trainerD.compute_gradients(self.d_loss,slim.get_variables(scope='discriminator'))
self.g_grads = self.trainerG.compute_gradients(self.g_loss, slim.get_variables(scope='generator'))
self.update_D = self.trainerD.apply_gradients(self.d_grads)
self.update_G = self.trainerG.apply_gradients(self.g_grads)
def train(self):
'''
Training loop for the GAN
'''
print('\n\n\n------------ Starting training ------------')
self.init = tf.global_variables_initializer()
self.session_config = tf.ConfigProto()
self.session_config.gpu_options.visible_device_list = self.config.gpu
self.session_config.gpu_options.allow_growth = True
self.session = tf.Session(config=self.session_config)
self.session.run(self.init)
self.var_list = tf.trainable_variables()
self.saver = tf.train.Saver(self.var_list)
# self.saver = tf.train.Saver()
for i in range(self.config.n_epochs):
print ('\n------ Epoch %i ------' % i)
for j in range(self.config.steps_per_epoch):
X_color, X_bw = self.data_generator.generate()
ys = (np.reshape(X_color,[self.config.batch_size,self.config.image_size,self.config.image_size,3])- 0.5) * 2.0
xs = (np.reshape(X_bw,[self.config.batch_size,self.config.image_size,self.config.image_size,1])- 0.5) * 2.0
_, self.dLoss = self.session.run([self.update_D,self.d_loss],feed_dict={self.real_in: ys, self.condition_in:xs})
_, self.gLoss = self.session.run([self.update_G,self.g_loss],feed_dict={self.real_in:ys, self.condition_in:xs})
if j % 10 == 0:
print ("Step: " + str(j) + " Gen Loss: " + str(self.gLoss) + " Disc Loss: " + str(self.dLoss))
if (i+1) % self.config.save_frequency == 0 and i != 0:
if not os.path.exists(self.config.model_dir):
os.makedirs(self.config.model_dir)
self.saver.save(self.session,self.config.model_dir+'/model-'+str(i)+'.cptk')
print ("Saved Model")
def predict(self):
'''
Load model weights and generate colored images from B&W images
'''
if not os.path.exists(self.config.predicted_dir):
os.makedirs(self.config.predicted_dir)
self.init = tf.global_variables_initializer()
self.session_config = tf.ConfigProto()
self.session_config.gpu_options.visible_device_list = self.config.gpu
self.session_config.gpu_options.allow_growth = True
self.session = tf.Session(config=self.session_config)
self.session.run(self.init)
self.var_list = tf.trainable_variables()
self.saver = tf.train.Saver(self.var_list)
ckpt = tf.train.get_checkpoint_state(self.config.model_dir)
self.saver.restore(self.session,ckpt.model_checkpoint_path)
generated_frames = []
samples = os.listdir(self.config.prediction_dir)
print(samples)
_, X_bw = self.dataset.convert_to_arrays(samples, training = False)
xs = (np.reshape(X_bw,[X_bw.shape[0],X_bw.shape[1],X_bw.shape[2],1]) - 0.5) * 2.0
self.sample_G = self.session.run(self.Gx,feed_dict={self.condition_in:xs})
generated_frames.append(self.sample_G)
generated_frames = np.vstack(generated_frames)
for i in range(len(generated_frames)):
red = generated_frames[i][:,:,2].copy()
blue = generated_frames[i][:,:,1].copy()
green = generated_frames[i][:,:,0].copy()
generated_frames[i][:,:,0] = red
generated_frames[i][:,:,1] = blue
generated_frames[i][:,:,2] = green
im = Image.fromarray((((generated_frames[i]) /2 + 0.5)* 256).astype('uint8'))
im.save(self.config.predicted_dir + '%i.jpg' %i)