lanczos_resize_edgetpu.py

def create_lanczos_resize_func_edgetpu(images_shape: Tuple[int], dim: int, factor: int, edgetpu=True):
    name = "lanczos_resize_%s_%d_%d" % ("_".join(map(str, images_shape)), dim, factor)
    ctor = lambda: create_lanczos_resize_func(images_shape, dim, factor)
    
    return create_func_edgetpu(images_shape, ctor, name, edgetpu)


def create_lanczos_resize_func(images_shape: Tuple[int], dim: int, factor: int):
    kernel = create_lanczos_resize_kernel(dim)
    
    @tf.function(input_signature=[tf.TensorSpec(images_shape, tf.float32)])
    def lanczos_func(images):
        # Note how we stride: every `factor` pixel is chosen
        return tf.nn.depthwise_conv2d(images, kernel, strides=[1, factor, factor, 1], padding="SAME")
    
    return lanczos_func


def create_lanczos_resize_kernel(dim: int):
    x = np.linspace(-dim, dim, dim * 2 + 1)
    y = np.linspace(-dim, dim, dim * 2 + 1)
    mx, my = np.meshgrid(x, y)
    # Our "constant" is 3
    kernel = np.sinc(mx / (dim / 3))*np.sinc(mx / dim) * np.sinc(my / (dim / 3)) * np.sinc(my / dim)
    kernel /= kernel.sum()
    return tf.tile(tf.constant(kernel.astype(np.float32))[:, :, None, None], [1, 1, 3, 1])


lanczos_resize = create_lanczos_resize_func(images.shape, 5, 2)
save_image(lanczos_resize(images), "result_origin.jpg")
lanczos_resize = create_lanczos_resize_func_edgetpu(images.shape, 5, 2)
save_image(lanczos_resize(images), "result_edgetpu.jpg")