Keras image classifier using VGG16

image_size = 128
mean_pixel = [103.939, 116.779, 123.68]


def get_img(animal, index):
    file_path = 'data/train/%ss/%s%03d.jpg' % (animal, animal, index)
    img = cv2.resize(cv2.imread(file_path), (image_size, image_size))
    for c in range(3):
        img[:, :, c] = img[:, :, c] - mean_pixel[c]
    return img


def get_data():
    X = []
    Y = []
    for i in range(0, 10000):
        X.append(get_img("cat", i))
        Y.append([0, 1, 0, 0, 0, 0, 0, 0])
        X.append(get_img("dog", i))
        Y.append([1, 0, 0, 0, 0, 0, 0, 0])
    return np.array(X), np.array(Y)


if __name__ == "__main__":
    vgg16 = VGG16(weights='imagenet', include_top=False)

    input = Input(shape=(image_size, image_size, 3), name='image_input')
    output_vgg16_conv = vgg16(input)

    x = Flatten(name='flatten')(output_vgg16_conv)
    x = Dense(4096, activation='relu', name='fc1')(x)
    x = Dense(4096, activation='relu', name='fc2')(x)
    x = Dense(8, activation='softmax', name='predictions')(x)

    my_model = Model(input=input, output=x)
    my_model.summary()

    weights_path = "my_vgg16_weights.h5"

    if os.path.isfile(weights_path):
        my_model.load_weights(weights_path)

    checkpoint = ModelCheckpoint(weights_path, monitor='val_acc', verbose=0, save_best_only=False, mode='max')
    my_model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])

    X, Y = get_data()
    my_model.fit(X, Y, epochs=20, verbose=1, callbacks=[checkpoint], shuffle=True)

    cat_img = get_img("cat", 10001)
    cat_im = np.expand_dims(cat_img, axis=0)
    print(my_model.predict(cat_im))

    dog_img = get_img("dog", 10001)
    dog_im = np.expand_dims(dog_img, axis=0)
    print(my_model.predict(dog_im))