nets.py 文件源码

def build_discriminator( shape, build_disc=True ) :
    '''
    Build discriminator.
    Set build_disc=False to build an encoder network to test
    the encoding/discrimination capability with autoencoder...
    '''
    def conv2d( x, filters, shape=(4, 4), **kwargs ) :
        '''
        I don't want to write lengthy parameters so I made a short hand function.
        '''
        x = Conv2D( filters, shape, strides=(2, 2),
            padding='same',
            kernel_initializer=Args.kernel_initializer,
            **kwargs )( x )
        #x = MaxPooling2D()( x )
        x = BatchNormalization(momentum=Args.bn_momentum)( x )
        x = LeakyReLU(alpha=Args.alpha_D)( x )
        return x

    # https://github.com/tdrussell/IllustrationGAN
    # As proposed by them, unlike GAN hacks, MaxPooling works better for anime dataset it seems.
    # However, animeGAN doesn't use it so I'll keep it more similar to DCGAN.

    face = Input( shape=shape )
    x = face

    # Warning: Don't batchnorm the first set of Conv2D.
    x = Conv2D( 64, (4, 4), strides=(2, 2),
        padding='same',
        kernel_initializer=Args.kernel_initializer )( x )
    x = LeakyReLU(alpha=Args.alpha_D)( x )
    # 32x32

    x = conv2d( x, 128 )
    # 16x16

    x = conv2d( x, 256 )
    # 8x8

    x = conv2d( x, 512 )
    # 4x4

    if build_disc:
        x = Flatten()(x)
        # add 16 features. Run 1D conv of size 3.
        #x = MinibatchDiscrimination(16, 3)( x )

        #x = Dense(1024, kernel_initializer=Args.kernel_initializer)( x )
        #x = LeakyReLU(alpha=Args.alpha_D)( x )

        # 1 when "real", 0 when "fake".
        x = Dense(1, activation='sigmoid',
            kernel_initializer=Args.kernel_initializer)( x )
        return models.Model( inputs=face, outputs=x )
    else:
        # build encoder.
        x = Conv2D(Args.noise_shape[2], (4, 4), activation='tanh')(x)
        return models.Model( inputs=face, outputs=x )