batch_shape = (32, 64, 64, 3)

ipt = Input(batch_shape=batch_shape)
x   = Conv2D(6,  (8, 8), strides=(1, 1), activation='relu',  padding='valid',
                               kernel_initializer='he_normal', name='cnn_0')(ipt)
x   = BatchNormalization(name='bn_0')(x)
x   = Dropout(0.1)(x)
x   = Conv2D(12, (8, 8), strides=(2, 2), activation='relu', padding='valid',
                               kernel_initializer='he_normal', name='cnn_1')(x)
x   = BatchNormalization(name='bn_1')(x)
x   = Conv2D(24, (4, 4), strides=(2, 2), activation='relu', padding='valid',
                               kernel_initializer='he_normal', name='cnn_2')(x)
x   = BatchNormalization(name='bn_2')(x)
x   = Conv2D(48, (3, 3), strides=(2, 2), activation='relu', padding='valid',
                               kernel_initializer='he_normal', name='cnn_3')(x)
x   = BatchNormalization(name='bn_3')(x)
x   = Conv2D(12, (1, 1), strides=(1, 1), activation='relu', padding='valid',
                               kernel_initializer='he_normal', name='cnn_4')(x)
x   = BatchNormalization(name='bn_4')(x)
x   = Flatten()(x)
x   = Dense(12, activation='relu')(x)
out = Dense(6, activation='softmax', name='output')(x)

model = Model(ipt, out)
model.compile('adam', 'sparse_categorical_crossentropy', metrics=['accuracy'])