The backward function of unfold is decorated with once_differential, yet I find this link
https://github.com/JiahuiYu/generative_inpainting/blob/master/inpaint_ops.py#L266-L267. It is an inpating model trained with wgan-gp, by extracting patches from an image with tf.extract_image_patches.
Does tf.extract_images_patches support double differential? Why ufold cannot?