I was reading this paper on automatic image resizing with a trainable Resizer/encoder yesterday and I am curious how do they back-propagate over the Resize as it is not trainable.
Something like this looks very interesting, and could be and of the shelf add on to fastai.
I merely skimmed through the paper so I’m most likely mistaken, but I got the impression that their method learns how to resize an image (which is differentiable), not what size to resize it to.