-
Notifications
You must be signed in to change notification settings - Fork 2
/
gan.py
189 lines (149 loc) · 8.65 KB
/
gan.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
from keras.models import Model, Sequential
from keras.layers import Input, Dense, Reshape, concatenate
from keras.layers.core import Activation, Flatten
from keras.layers.normalization import BatchNormalization
from keras.layers.convolutional import UpSampling2D, Conv2D, MaxPooling2D
from keras.optimizers import SGD
from support.image_save import image_for_snapshot , image_from_array
from support.glove import Glove
from keras import backend as K
import numpy as np
from PIL import Image
import os
class Gan(object):
def __init__(self):
K.set_image_dim_ordering('tf')
self.generator = None
self.discriminator = None
self.model = None
self.img_width = 16
self.img_height = 16
self.img_channels = 3
self.random_input_dim = 20
self.text_input_dim = 100
self.config = None
self.glove_source_dir_path = './glove'
self.glove_model = Glove()
def create_model(self):
init_img_width = self.img_width // 4
init_img_height = self.img_height // 4
random_input = Input(shape=(self.random_input_dim,))
text_input1 = Input(shape=(self.text_input_dim,))
random_dense = Dense(self.random_input_dim)(random_input)
text_layer1 = Dense(1024)(text_input1)
merged = concatenate([random_dense, text_layer1])
generator_layer = Activation('tanh')(merged)
generator_layer = Dense(128 * init_img_width * init_img_height)(generator_layer)
generator_layer = BatchNormalization()(generator_layer)
generator_layer = Activation('tanh')(generator_layer)
generator_layer = Reshape((init_img_width, init_img_height, 128),
input_shape=(128 * init_img_width * init_img_height,))(generator_layer)
generator_layer = UpSampling2D(size=(2, 2))(generator_layer)
generator_layer = Conv2D(64, kernel_size=5, padding='same')(generator_layer)
generator_layer = Activation('tanh')(generator_layer)
generator_layer = UpSampling2D(size=(2, 2))(generator_layer)
generator_layer = Conv2D(self.img_channels, kernel_size=5, padding='same')(generator_layer)
generator_output = Activation('tanh')(generator_layer)
self.generator = Model([random_input, text_input1], generator_output)
self.generator.compile(loss='mean_squared_error', optimizer="SGD")
# print('generator: ', self.generator.summary())
text_input2 = Input(shape=(self.text_input_dim,))
text_layer2 = Dense(1024)(text_input2)
img_input2 = Input(shape=(self.img_width, self.img_height, self.img_channels))
img_layer2 = Conv2D(64, kernel_size=(5, 5), padding='same')(
img_input2)
img_layer2 = Activation('tanh')(img_layer2)
img_layer2 = MaxPooling2D(pool_size=(2, 2))(img_layer2)
img_layer2 = Conv2D(128, kernel_size=5)(img_layer2)
img_layer2 = Activation('tanh')(img_layer2)
img_layer2 = MaxPooling2D(pool_size=(2, 2))(img_layer2)
img_layer2 = Flatten()(img_layer2)
img_layer2 = Dense(1024)(img_layer2)
merged = concatenate([img_layer2, text_layer2])
discriminator_layer = Activation('tanh')(merged)
discriminator_layer = Dense(1)(discriminator_layer)
discriminator_output = Activation('sigmoid')(discriminator_layer)
self.discriminator = Model([img_input2, text_input2], discriminator_output)
d_optim = SGD(lr=0.0005, momentum=0.9, nesterov=True)
self.discriminator.compile(loss='binary_crossentropy', optimizer=d_optim)
# print('discriminator: ', self.discriminator.summary())
model_output = self.discriminator([self.generator.output, text_input1])
self.model = Model([random_input, text_input1], model_output)
self.discriminator.trainable = False
g_optim = SGD(lr=0.0005, momentum=0.9, nesterov=True)
self.model.compile(loss='binary_crossentropy', optimizer=g_optim)
# print('generator-discriminator: ', self.model.summary())
def load_model(self):
config_file_path = './weights/config.npy'
self.config = np.load(config_file_path).item()
self.img_width = self.config['img_width']
self.img_height = self.config['img_height']
self.img_channels = self.config['img_channels']
self.random_input_dim = self.config['random_input_dim']
self.text_input_dim = self.config['text_input_dim']
self.glove_source_dir_path = self.config['glove_source_dir_path']
self.create_model()
self.glove_model.load_glove(self.glove_source_dir_path)
self.generator.load_weights('./weights/gen.h5')
self.discriminator.load_weights('./weights/disc.h5')
def fit(self, image_label_pairs, epochs, batch_size, snapshot_dir_path , snapshot_interval , model_dir_path):
self.config = dict()
self.config['img_width'] = self.img_width
self.config['img_height'] = self.img_height
self.config['random_input_dim'] = self.random_input_dim
self.config['text_input_dim'] = self.text_input_dim
self.config['img_channels'] = self.img_channels
self.config['glove_source_dir_path'] = self.glove_source_dir_path
self.glove_model.load_glove(self.glove_source_dir_path)
config_file_path = './weights/config.npy'
np.save(config_file_path, self.config)
noise = np.zeros((batch_size, self.random_input_dim))
text_batch = np.zeros((batch_size, self.text_input_dim))
self.create_model()
for epoch in range(epochs):
print("Epoch :", epoch)
batch_count = int(image_label_pairs.shape[0] / batch_size)
for batch_index in range(batch_count):
# Step 1: train the discriminator
image_label_pair_batch = image_label_pairs[batch_index * batch_size:(batch_index + 1) * batch_size]
image_batch = []
for index in range(batch_size):
image_label_pair = image_label_pair_batch[index]
normalized_img = image_label_pair[0]
text = image_label_pair[1]
image_batch.append(normalized_img)
text_batch[index, :] = self.glove_model.encode_text(text)
noise[index, :] = np.random.uniform(-1, 1, self.random_input_dim)
image_batch = np.array(image_batch)
generated_images = self.generator.predict([noise, text_batch], verbose=0)
self.discriminator.trainable = True
d_loss = self.discriminator.train_on_batch([np.concatenate((image_batch, generated_images)),
np.concatenate((text_batch, text_batch))],
np.array([1] * batch_size + [0] * batch_size))
# Step 2: train the generator
for index in range(batch_size):
noise[index, :] = np.random.uniform(-1, 1, self.random_input_dim)
self.discriminator.trainable = False
g_loss = self.model.train_on_batch([noise, text_batch], np.array([1] * batch_size))
if epoch % 100 == 0 and (batch_index == 0 or batch_index ==1) :
print("Epoch %d batch %d d_loss : %f" % (epoch, batch_index, d_loss))
print("Epoch %d batch %d g_loss : %f" % (epoch, batch_index, g_loss))
self.save_intermediate_result(generated_images, snapshot_dir_path=snapshot_dir_path,
epoch=epoch, batch_index=batch_index)
self.generator.save_weights('./weights/gen.h5', True)
self.discriminator.save_weights('./weights/disc.h5' , True)
self.generator.save_weights('./weights/gen.h5', True)
self.discriminator.save_weights('./weights/disc.h5' , True)
def generate_image_from_text(self, text):
noise = np.zeros(shape=(1, self.random_input_dim))
encoded_text = np.zeros(shape=(1, self.text_input_dim))
encoded_text[0, :] = self.glove_model.encode_text(text)
noise[0, :] = np.random.uniform(-1, 1, self.random_input_dim)
generated_images = self.generator.predict([noise, encoded_text], verbose=0)
generated_image = generated_images[0]
generated_image = generated_image * 127.5 + 127.5
return Image.fromarray(generated_image.astype(np.uint8))
def save_intermediate_result(self, generated_images, snapshot_dir_path, epoch, batch_index):
image = image_for_snapshot(generated_images)
image_from_array(image).save(
os.path.join(snapshot_dir_path , 'result' + str(epoch) + "-" + str(batch_index) + ".jpg"))