GAN

cell_size = 8
im_size = (8, 8)
input_size = (*im_size, 3)
gen_params = 50

def discriminator():
    m = keras.models.Sequential()

    m.add(keras.layers.UpSampling2D(input_shape = input_size))
    m.add(keras.layers.Conv2D(cell_size,
                              5,
#                               input_shape = input_size,
                              padding = 'same',
#                              activation=keras.activations.selu,
#                               kernel_initializer = 'lecun_normal'
                             ))
    m.add(keras.layers.Dropout(0.4))
    m.add(keras.layers.BatchNormalization(momentum=0.9))
    m.add(keras.layers.LeakyReLU(alpha=0.2))


    m.add(keras.layers.Conv2D(cell_size,
                              1,
                              input_shape = input_size,
                              padding = 'same',
#                              activation=keras.activations.selu,
#                               kernel_initializer = 'lecun_normal'
                             ))
    m.add(keras.layers.Dropout(0.4))
    m.add(keras.layers.BatchNormalization(momentum=0.9))
    m.add(keras.layers.LeakyReLU(alpha=0.2))

    m.add(keras.layers.Conv2D(cell_size,
                              1,
                              input_shape = input_size,
                              padding = 'same',
#                              activation=keras.activations.selu,
#                               kernel_initializer = 'lecun_normal'
                             ))
    m.add(keras.layers.Dropout(0.4))
    m.add(keras.layers.BatchNormalization(momentum=0.9))
    m.add(keras.layers.LeakyReLU(alpha=0.2))

    m.add(keras.layers.Flatten())
    m.add(keras.layers.Dense(1, activation=keras.activations.sigmoid))
    return m


def generator():
    m = keras.models.Sequential()

    m.add(keras.layers.Dense(8 * 8 * 3,
                             input_dim=gen_params,
#                              activation=keras.activations.selu,
#                               kernel_initializer = 'lecun_normal'
                             ))

    m.add(keras.layers.BatchNormalization(momentum=0.9))
    m.add(keras.layers.LeakyReLU(alpha=0.2))
    m.add(keras.layers.Dropout(0.4))
    m.add(keras.layers.Reshape((8, 8, 3)))


    m.add(keras.layers.Conv2DTranspose(4,
                                       1,
                                       padding = 'same',
#                                        activation = keras.activations.selu,
                                       kernel_initializer = 'lecun_normal'))
    m.add(keras.layers.BatchNormalization(momentum=0.9))
    m.add(keras.layers.LeakyReLU(alpha=0.2))
    m.add(keras.layers.Dropout(0.4))

    m.add(keras.layers.Conv2DTranspose(input_size[-1],
                                   5,
                                   padding = 'same',
                                   activation = keras.activations.sigmoid,
#                               kernel_initializer = 'lecun_normal'
                             ))
    assert np.array_equal(m.layers[-1].output_shape[1:], input_size), m.summary()

    return m