mnist_func.py

input = layers.Input(input_shape)

x_1 = layers.Conv2D(32, kernel_size=(3, 3), activation="relu")(input)
x_1 = layers.MaxPooling2D(pool_size=(2, 2))(x_1)
x_1 = layers.Conv2D(64, kernel_size=(3, 3), activation="relu")(x_1)
x_1 = layers.MaxPooling2D(pool_size=(2, 2))(x_1)
x_1 = layers.Flatten()(x_1)

x_2 = layers.Conv2D(16, kernel_size=(3, 3), activation="relu")(input)
x_2 = layers.MaxPooling2D(pool_size=(2, 2))(x_2)
x_2 = layers.Conv2D(32, kernel_size=(7, 7), activation="relu")(x_2)
x_2 = layers.MaxPooling2D(pool_size=(2, 2))(x_2)
x_2 = layers.Flatten()(x_2)

x = layers.concatenate([x_1, x_2])
x = layers.Dropout(0.5)(x)
out = layers.Dense(num_classes, activation="softmax")(x)

model = keras.Model(input, out)

model.summary()