🔨 Rework PPs, Use transformsV2, & Fix corruptions

[o-laurent]

• Improve the corruption transforms
• Rework postprocessing methods
• Use transformsV2
• Fix the security issue

[codecov]

Codecov Report

Attention: Patch coverage is 99.27007% with 2 lines in your changes missing coverage. Please review.

Project coverage is 99.15%. Comparing base (b3fe75e) to head (de6d239).
Report is 58 commits behind head on main.

✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
torch_uncertainty/routines/segmentation.py	66.66%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #145      +/-   ##
==========================================
- Coverage   99.19%   99.15%   -0.05%     
==========================================
  Files         144      144              
  Lines        7192     7307     +115     
  Branches      925      942      +17     
==========================================
+ Hits         7134     7245     +111     
- Misses         26       28       +2     
- Partials       32       34       +2     

Flag	Coverage Δ
cpu	?
pytest	99.15% <99.27%> (-0.05%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[o-laurent]

@alafage we need to state that the calibration dataloader parameters are going to be those of the validation (or test) set, then, notably for MC-BatchNorm, otherwise you wouldn't know how to set the parameters of this method. Ideally we should find a workaround because it makes no sense to define the MCBatchNorm parameter in the datamodule; maybe temporarily overwrite the dataloader's batch_size in fit?
Moreover, we should warn the user when taking the test dataset as a calibration set, and ideally provide a way to split the sets in a more principled way from the start.

[o-laurent]

I've made changes accordingly. Tell me what you think about them when you have time @alafage. Basically, I believe that the post-processing set (not only a calibration set anymore) should be handled at the datamodule level instead of the routine.

[alafage]

So we don’t have pretrained weights?

[o-laurent]

We have some, but I don't think that we can sell our library on this for now. Maybe it would be good to improve our number of pre-trained models for the submission.