DCGANで顔生成 - 日々の記録

顔生成した.

discriminatorとgenerator

def generator_model(input_shape):
    input_ch = 1024
    inputs = Input(shape=input_shape)
    x = Dense(input_ch*8*8)(inputs)
    x = ReLU()(x)
    x = Reshape((8, 8, input_ch))(x)
    x = BatchNormalization(momentum=0.9)(x)
    x = ReLU()(x)
    x = Deconv2D(input_ch, (4, 4), strides=(2, 2), padding="SAME")(x)
    x = BatchNormalization(momentum=0.9)(x)
    x = ReLU()(x)
    x = Deconv2D(input_ch//2, (4, 4), strides=(2, 2), padding="SAME")(x)
    x = BatchNormalization(momentum=0.9)(x)
    x = ReLU()(x)
    x = Deconv2D(input_ch//4, (4, 4), strides=(2, 2), padding="SAME")(x)
    x = BatchNormalization(momentum=0.9)(x)
    x = ReLU()(x)
    x = Deconv2D(input_ch//8, (4, 4), strides=(2, 2), padding="SAME")(x)
    x = BatchNormalization(momentum=0.9)(x)
    x = ReLU()(x)
    x = Deconv2D(3, (3, 3), strides=(1, 1), padding="SAME")(x)
    out = Activation('tanh')(x)
    return Model(inputs=inputs, outputs=out)


def discriminator_model(input_shape):
    ch_list = [64, 128, 128, 256, 256, 512]
    kernel_list = [4, 3, 4, 3, 4, 3]
    stride_list = [2, 1, 2, 1, 2, 1]
    inputs = Input(shape=input_shape)
    x = Conv2D(
        64, (3, 3), strides=1, padding='same', input_shape=input_shape
    )(inputs)
    for cl, kl, sl in zip(ch_list, kernel_list, stride_list):
        x = noise.GaussianNoise(0)(x)
        x = LeakyReLU()(x)
        x = Conv2D(
            cl, kl, strides=sl, padding='same', input_shape=input_shape
        )(x)
    x = noise.GaussianNoise(0)(x)
    x = LeakyReLU()(x)
    flatten = Flatten()(x)
    x2 = Dense(1)(flatten)
    act = Activation('sigmoid')(x2)
    return Model(inputs=inputs, outputs=act)

迷走した結果こんな感じになった. discriminatorにbatchnormalization入れるならreal imageとfake imageを別々のバッチに入れるとよいらしい.

d_loss_real = discriminator.train_on_batch(
   image_batch, [0.9]*(BATCH_SIZE))
d_loss_fake = discriminator.train_on_batch(
    generated_images, [0]*(BATCH_SIZE))

学習を安定化させるテクニックにノイジーなラベルを使うとよいとあるので0.9にしてみた.
Dにガウシアンノイズ入れたりしたけど意味あるのかこれ.

qiita.com
bamos.github.io