def D_model(Height, Width, channel=3):
base = 32
inputs = Input((Height, Width, channel))
x = Conv2D(base, (5, 5), padding='same', strides=(2,2), name='d_conv1',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(inputs)
x = LeakyReLU(alpha=0.2)(x)
#x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='d_conv1_bn')(x)
x = Conv2D(base*2, (5, 5), padding='same', strides=(2,2), name='d_conv2',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
#x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='d_conv2_bn')(x)
x = Conv2D(base*4, (5, 5), padding='same', strides=(2,2), name='d_conv3',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
#x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='d_conv3_bn')(x)
x = Conv2D(base*8, (5, 5), padding='same', strides=(2,2), name='d_conv4',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
#x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='d_conv4_bn')(x)
x = Flatten()(x)
#x = Dense(4096, activation='relu', name='d_dense1',
# kernel_initializer=RN(mean=0.0, stddev=0.02), bias_initializer=Constant())(x)
x = Dense(1, activation='sigmoid', name='d_out',
kernel_initializer=RN(mean=0.0, stddev=0.02), bias_initializer=Constant())(x)
model = Model(inputs=inputs, outputs=x, name='D')
return model
def G_model(Height, Width, channel=3):
inputs = Input((100, ))
in_h = int(Height / 16)
in_w = int(Width / 16)
d_dim = 256
base = 128
x = Dense(in_h * in_w * d_dim,
name='g_dense1',
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(inputs)
x = Reshape((in_h, in_w, d_dim), input_shape=(d_dim * in_h * in_w, ))(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_dense1_bn')(x)
# 1/8
x = Conv2DTranspose(base * 4, (5, 5),
name='g_conv1',
padding='same',
strides=(2, 2),
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv1_bn')(x)
# 1/4
x = Conv2DTranspose(base * 2, (5, 5),
name='g_conv2',
padding='same',
strides=(2, 2),
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv2_bn')(x)
# 1/2
x = Conv2DTranspose(base, (5, 5),
name='g_conv3',
padding='same',
strides=(2, 2),
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv3_bn')(x)
# 1/1
x = Conv2DTranspose(channel, (5, 5),
name='g_out',
padding='same',
strides=(2, 2),
kernel_initializer=RN(mean=0.0, stddev=0.02),
bias_initializer=Constant())(x)
x = Activation('tanh')(x)
model = Model(inputs=inputs, outputs=x, name='G')
return model
def G_model():
inputs = Input([100, ], name="x")
con_x = Input([num_classes, ], name="con_x")
con_x2 = Input([img_height, img_width, num_classes], name="con_x2")
#con_x = K.zeros([None, num_classes, 1, 1])
#print(con_x.shape)
#con_x = np.zeros([len(_con_x), num_classes, 1, 1], dtype=np.float32)
#con_x[np.arange(len(_con_x)), _con_x] = 1
x = concatenate([inputs, con_x], axis=-1)
in_h = int(img_height / 16)
in_w = int(img_width / 16)
d_dim = 256
base = 128
x = Dense(in_h * in_w * d_dim, name='g_dense1',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = Reshape((in_h, in_w, d_dim), input_shape=(d_dim * in_h * in_w,))(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_dense1_bn')(x)
# 1/8
x = Conv2DTranspose(base*4, (5, 5), name='g_conv1', padding='same', strides=(2,2),
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv1_bn')(x)
# 1/4
x = Conv2DTranspose(base*2, (5, 5), name='g_conv2', padding='same', strides=(2,2),
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv2_bn')(x)
# 1/2
x = Conv2DTranspose(base, (5, 5), name='g_conv3', padding='same', strides=(2,2),
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = Activation('relu')(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv3_bn')(x)
# 1/1
x = Conv2DTranspose(channel, (5, 5), name='g_out', padding='same', strides=(2,2),
kernel_initializer=RN(mean=0.0, stddev=0.02), bias_initializer=Constant())(x)
x = Activation('tanh')(x)
#con_x = np.zerns([len(_con_x), num_classes, img_height, img_width], dtype=np.float32)
#con_x[np.arange(len(_con_x)), _con_x] = 1
x2 = concatenate([x, con_x2], axis=-1)
model = Model(inputs=[inputs, con_x, con_x2], outputs=[x], name='G')
gan_g_model = Model(inputs=[inputs, con_x, con_x2], outputs=[x2], name='GAN_G')
return model, gan_g_model
def D_model():
base = 32
inputs = Input([img_height, img_width, channel + num_classes])
x = Conv2D(base, (5, 5), padding='same', strides=(2,2), name='d_conv1',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(inputs)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(base*2, (5, 5), padding='same', strides=(2,2), name='d_conv2',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(base*4, (5, 5), padding='same', strides=(2,2), name='d_conv3',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(base*8, (5, 5), padding='same', strides=(2,2), name='d_conv4',
kernel_initializer=RN(mean=0.0, stddev=0.02), use_bias=False)(x)
x = LeakyReLU(alpha=0.2)(x)
x = Flatten()(x)
x = Dense(1, activation='sigmoid', name='d_out',
kernel_initializer=RN(mean=0.0, stddev=0.02), bias_initializer=Constant())(x)
model = Model(inputs=inputs, outputs=x, name='D')
return model
def classifying(Height, Width, channel=3, num_classes=10):
inputs_x = Input((Height, Width, channel),
name='classifying_X') # generator出力を取得
x = Flatten()(inputs_x)
x = Dense(num_classes,
activation='sigmoid',
name='d_out',
kernel_initializer=RN(mean=0.0, stddev=0.02),
bias_initializer=Constant())(x)
model = Model(inputs=inputs_x, outputs=x, name='D')
return model
def conv_x(Height, Width, channel=3):
inputs_z = Input((Height, Width, channel), name='Z') # 入力画像を取得
x = Conv2D(32, (5, 5),
padding='same',
strides=(2, 2),
name='g_conv1',
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(inputs_z)
x = InstanceNormalization()(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv1_bn')(x)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(64, (5, 5),
padding='same',
strides=(2, 2),
name='g_conv2',
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv2_bn')(x)
x = InstanceNormalization()(x)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(128, (5, 5),
padding='same',
strides=(2, 2),
name='g_conv3',
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='d_conv3_bn')(x)
x = InstanceNormalization()(x)
x = LeakyReLU(alpha=0.2)(x)
x = Conv2D(256, (5, 5),
padding='same',
strides=(2, 2),
name='g_conv4',
kernel_initializer=RN(mean=0.0, stddev=0.02),
use_bias=False)(x)
x = BatchNormalization(momentum=0.9, epsilon=1e-5, name='g_conv4_bn')(x)
x = InstanceNormalization()(x)
x = LeakyReLU(alpha=0.2)(x)
model = Model(inputs=[inputs_z], outputs=[x], name='G')
return model
