質問編集履歴
1
コードの追加
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -46,4 +46,192 @@
|
|
|
46
46
|
- 画像分類などは問題なく作れる
|
|
47
47
|
|
|
48
48
|
インターネットで手当たり次第に調べたり、書籍で参考になりそうな情報を探したり、手を尽くしましたがなぜ全くうまくいかないのかわかりません。問題がレイヤーの組み方なのか、何か決定的に重要な処理を飛ばしてしまっているのか、損失関数などの評価の仕方が悪いのかも見当もつきません。
|
|
49
|
-
なにか見落としているポイントなどがありますでしょうか？
|
|
49
|
+
なにか見落としているポイントなどがありますでしょうか？
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
コード
|
|
53
|
+
```Python
|
|
54
|
+
import numpy as np
|
|
55
|
+
from PIL import Image
|
|
56
|
+
|
|
57
|
+
from keras.datasets import mnist
|
|
58
|
+
from keras.layers import *
|
|
59
|
+
from keras.models import *
|
|
60
|
+
from keras.optimizers import *
|
|
61
|
+
|
|
62
|
+
|
|
63
|
+
# --- プログレスバーを表示するクラス(学習自体には関係ありません) ---------
|
|
64
|
+
class ProgressBar:
|
|
65
|
+
def __init__(self, entireJob):
|
|
66
|
+
self.job = entireJob
|
|
67
|
+
self.width = 40
|
|
68
|
+
def draw(self, progress):
|
|
69
|
+
print( ("\r["+"#"*int((progress+1)*self.width/self.job)+" "*(self.width-int((progress+1)*self.width/self.job) ) +"] %d/%d")%(progress+1,self.job), end="")
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
# --- Generatorモデルの定義 -----------
|
|
73
|
+
class Generator:
|
|
74
|
+
def __init__(self):
|
|
75
|
+
layer0 = Input(shape=(1,1,100))
|
|
76
|
+
|
|
77
|
+
layer1 = UpSampling2D(size=(3,3))(layer0)
|
|
78
|
+
layer1 = Conv2D(
|
|
79
|
+
filters=100,
|
|
80
|
+
kernel_size=(2,2),
|
|
81
|
+
strides=(1,1),
|
|
82
|
+
padding='same',
|
|
83
|
+
activation='relu' )(layer1)
|
|
84
|
+
layer1 = BatchNormalization()(layer1)
|
|
85
|
+
|
|
86
|
+
layer2 = UpSampling2D(size=(3,3))(layer1)
|
|
87
|
+
layer2 = Conv2D(
|
|
88
|
+
filters=100,
|
|
89
|
+
kernel_size=(2,2),
|
|
90
|
+
strides=(1,1),
|
|
91
|
+
padding='same',
|
|
92
|
+
activation='relu' )(layer2)
|
|
93
|
+
layer2 = BatchNormalization()(layer2)
|
|
94
|
+
|
|
95
|
+
layer3 = UpSampling2D(size=(2,2))(layer2)
|
|
96
|
+
layer3 = Conv2D(
|
|
97
|
+
filters=80,
|
|
98
|
+
kernel_size=(3,3),
|
|
99
|
+
strides=(1,1),
|
|
100
|
+
padding='valid',
|
|
101
|
+
activation='elu' )(layer3)
|
|
102
|
+
layer3 = BatchNormalization()(layer3)
|
|
103
|
+
|
|
104
|
+
layer4 = UpSampling2D(size=(2,2))(layer3)
|
|
105
|
+
layer4 = Conv2D(
|
|
106
|
+
filters=50,
|
|
107
|
+
kernel_size=(3,3),
|
|
108
|
+
strides=(1,1),
|
|
109
|
+
padding='same',
|
|
110
|
+
activation='elu' )(layer4)
|
|
111
|
+
layer4 = BatchNormalization()(layer4)
|
|
112
|
+
|
|
113
|
+
layer5 = UpSampling2D(size=(2,2))(layer4)
|
|
114
|
+
layer5 = Conv2D(
|
|
115
|
+
filters=20,
|
|
116
|
+
kernel_size=(4,4),
|
|
117
|
+
strides=(2,2),
|
|
118
|
+
padding='valid',
|
|
119
|
+
activation='elu' )(layer5)
|
|
120
|
+
layer5 = BatchNormalization()(layer5)
|
|
121
|
+
|
|
122
|
+
layer6 = Conv2D(
|
|
123
|
+
filters=1,
|
|
124
|
+
kernel_size=(4,4),
|
|
125
|
+
strides=(1,1),
|
|
126
|
+
padding='valid',
|
|
127
|
+
activation='tanh' )(layer5)
|
|
128
|
+
|
|
129
|
+
self.model = Model(layer0, layer6)
|
|
130
|
+
self.model.summary()
|
|
131
|
+
|
|
132
|
+
# --- Discriminatorモデルの定義 -------
|
|
133
|
+
class Discriminator:
|
|
134
|
+
def __init__(self):
|
|
135
|
+
layer0 = Input(shape=(28,28,1))
|
|
136
|
+
layer1 = Conv2D(
|
|
137
|
+
filters=5,
|
|
138
|
+
kernel_size=(3,3),
|
|
139
|
+
strides=(2,2),
|
|
140
|
+
padding='valid',
|
|
141
|
+
activation='elu' )(layer0)
|
|
142
|
+
layer1 = BatchNormalization()(layer1)
|
|
143
|
+
|
|
144
|
+
layer2 = Conv2D(
|
|
145
|
+
filters=10,
|
|
146
|
+
kernel_size=(3,3),
|
|
147
|
+
strides=(2,2),
|
|
148
|
+
padding='valid',
|
|
149
|
+
activation='elu' )(layer1)
|
|
150
|
+
layer2 = BatchNormalization()(layer2)
|
|
151
|
+
|
|
152
|
+
layer3 = Conv2D(
|
|
153
|
+
filters=5,
|
|
154
|
+
kernel_size=(3,3),
|
|
155
|
+
strides=(1,1),
|
|
156
|
+
padding='valid',
|
|
157
|
+
activation='relu' )(layer2)
|
|
158
|
+
layer3 = BatchNormalization()(layer3)
|
|
159
|
+
|
|
160
|
+
layer4 = Flatten()(layer3)
|
|
161
|
+
layer4 = Dense(units=30, activation='tanh')(layer4)
|
|
162
|
+
layer4 = BatchNormalization()(layer4)
|
|
163
|
+
|
|
164
|
+
layer5 = Dense(units=1, activation='sigmoid' )(layer4)
|
|
165
|
+
|
|
166
|
+
self.model = Model(layer0, layer5)
|
|
167
|
+
self.model.summary()
|
|
168
|
+
|
|
169
|
+
|
|
170
|
+
|
|
171
|
+
class Main:
|
|
172
|
+
def __init__(self):
|
|
173
|
+
# --- Discriminatorの定義 -----------------
|
|
174
|
+
self.discriminator = Discriminator().model
|
|
175
|
+
self.discriminator.compile(
|
|
176
|
+
optimizer=SGD(learning_rate=1e-4),
|
|
177
|
+
loss='binary_crossentropy',
|
|
178
|
+
metrics=['accuracy'] )
|
|
179
|
+
|
|
180
|
+
# --- GeneratorとDiscriminatorを連結したモデルの定義 ---
|
|
181
|
+
self.generator = Generator().model
|
|
182
|
+
z = Input(shape=(1,1,100))
|
|
183
|
+
img = self.generator(z)
|
|
184
|
+
self.discriminator.trainable = False # Discriminatorを更新しないよう設定
|
|
185
|
+
valid = self.discriminator(img)
|
|
186
|
+
self.combined = Model(z, valid)
|
|
187
|
+
self.combined.compile(
|
|
188
|
+
optimizer=Adam(learning_rate=1e-6),
|
|
189
|
+
loss='binary_crossentropy',
|
|
190
|
+
metrics=['accuracy'] )
|
|
191
|
+
|
|
192
|
+
# --- MNISTデータセットの用意 ---------------
|
|
193
|
+
(x_train, t_train), (x_test, t_test) = mnist.load_data()
|
|
194
|
+
x_train = x_train.reshape(60000, 28, 28, 1)
|
|
195
|
+
x_test = x_test.reshape(10000, 28, 28, 1)
|
|
196
|
+
self.x_train = x_train.astype('float32')
|
|
197
|
+
self.x_test = x_test.astype('float32')
|
|
198
|
+
|
|
199
|
+
# --- 学習 -------------------------------------
|
|
200
|
+
def _train(self, iteration, batch_size):
|
|
201
|
+
progress = ProgressBar(iteration) # プログレスバーを用意
|
|
202
|
+
for i in range(iteration):
|
|
203
|
+
z = np.random.uniform(-1,1,(batch_size//2,1,1,100)) # ノイズベクトルの生成
|
|
204
|
+
f_img = self.generator.predict(z) # f_img(fake_img)の生成
|
|
205
|
+
r_img = self.x_train[np.random.randint(0, 60000, batch_size//2)] # r_img(real_img)を読み込み
|
|
206
|
+
loss_d, acc_d = self.discriminator.train_on_batch(f_img, np.zeros((batch_size//2,1))) # Discriminatorの学習
|
|
207
|
+
loss_d_, acc_d_ = self.discriminator.train_on_batch(r_img, np.ones( (batch_size//2,1))) # acc_d = Discriminatorのaccuracy
|
|
208
|
+
acc_d += acc_d_
|
|
209
|
+
|
|
210
|
+
z = np.random.uniform(-1,1,(batch_size,1,1,100)) # ノイズベクトルの生成
|
|
211
|
+
loss_g, acc_g = self.combined.train_on_batch(z, np.ones((batch_size,1))) # Generatorの学習
|
|
212
|
+
progress.draw(i) # プログレスバーの表示
|
|
213
|
+
print(" Accuracy=(%f,%f)"%(acc_g, acc_d/2), end="")
|
|
214
|
+
|
|
215
|
+
def train(self, iteration, batch_size, epoch):
|
|
216
|
+
for i in range(epoch):
|
|
217
|
+
print("Epoch %d/%d\n"%(i+1, epoch))
|
|
218
|
+
self._train(iteration, batch_size) # _train()をepoch回繰り返します
|
|
219
|
+
|
|
220
|
+
# --- 学習が終わった時の確認用に一枚だけ画像を作ります -------
|
|
221
|
+
def create_image(self):
|
|
222
|
+
z = np.random.uniform(-1,1,(1,1,1,100))
|
|
223
|
+
img = self.generator.predict(z)
|
|
224
|
+
return img.reshape(1,28,28)
|
|
225
|
+
|
|
226
|
+
|
|
227
|
+
if __name__ == "__main__":
|
|
228
|
+
main = Main()
|
|
229
|
+
main.train(iteration=1875, batch_size=32, epoch=1)
|
|
230
|
+
|
|
231
|
+
# --- 画像を表示 -----------------------
|
|
232
|
+
img = main.create_image()
|
|
233
|
+
img = Image.fromarray(np.uint8(img.reshape(28,28) * 255))
|
|
234
|
+
img.show()
|
|
235
|
+
img.save("gan_generated_img.png")
|
|
236
|
+
|
|
237
|
+
```
|