Test validation augmentation

[NathanMolinier]

Description

This PR allow to run augmentations during validation to save a different set of weights

[NathanMolinier]

Results on AMOS22: Models are trained solely on AMOS22 CT. The probability of data augmentation being applied to validation samples is referenced after validation_aug_XX

[Hendrik-code]

Results on AMOS22: Models are trained solely on AMOS22 CT. The probability of data augmentation being applied to validation samples is referenced after validation_aug_XX

I think I need more explanation
what is OSSTA?
but sounds nice, looks like 0.5 yields the overall best results, right?

[NathanMolinier]

This is One Sequence to Segment Them All (OSSTA) the MICCAI version. I realize that I did train the Augmentation validation with the parameters of the current transform GPU that include Paul's code. My codebase was a bit messy, I trained on different branches. Maybe I can retrain it, but the code should work

[Hendrik-code]

This is One Sequence to Segment Them All (OSSTA) the MICCAI version. I realize that I did train the Augmentation validation with the parameters of the current transform GPU that include Paul's code. My codebase was a bit messy, I trained on different branches. Maybe I can retrain it, but the code should work

can you run OSSTA with the validation setups again? Because OSSTA currently yields the overall best results without the vcalidation thing, so your results could be a fluke as far as we know.

Code looks good to me, but I want to merge only if we definitely have positive evidance for this.