Opposite transform of DivisiblePad?

[jmlipman]

Hi,

I'm using DivisiblePad to pad images on the fly. It's very handy because I have variable-size images and all I care is that they can fit into the model. At test time I also have variable-size images, and I need to undo this operation, i.e., the model gets a padded image and the prediction needs to be unpadded (cropped) to be compared with its ground truth. CenterSpatialCrop could be useful but it only allows for a fixed roi_size for every image. Is there any nice way to solve this?

Thanks!

[KumoLiu]

Hi @jmlipman, thanks for your interest here.

I think you mean inverse and DivisiblePad is invertible, you could refer to this tutorial.

Hope it helps, thanks!

[jmlipman]

I think this is what I was looking for. Thank you!