Julien Cohen-Adad

thank you Antoine-- we will keep sct_deepseg_sc for a while, given the _very_ long inference time of nnUnet, so don't worry about the deprecation...

I keep thinking we should re-visit the approach for segmentation. What we did in `sct_deepseg_sc` was to crop the image around the cord before training/inferring. Not for you @naga-karthik 😬,...

@joshuacwnewton thank you so much for offering to help! 🎉 As you rightfully said, some CLIs are still using our custom parser via `msct_parser`, and we are (slowly) transitioning to...

> Might be good to hammer down how we want to communicate verbosity, and what types of output correspond to each level. 💯 for the sake of simplicity, i would...

for these cases i would merge verbose 1 and 2, ie: if -v, then do what is currently under verbose 1 and 2

Indeed, but in this case there is an additional layer of complexity, because we might not _need_ the verbose=0 as initially designed. When I first designed this v=0, 1, 2...

> > So, we could also change the current behavior, and have: > > > > * v=0: logging.level=info (default behaviour) > > * v=1: logging.level=debug, with output of plotting...

i like that approach https://github.com/neuropoly/spinalcordtoolbox/issues/2676#issuecomment-723089582, it does indeed have the merit of being intuitive. The downside is that it implies large breakage in cross-compatibility.

How about also adding Wandb image logging in this trainer, to also address #2490?