General QOL improvements

[Sllambias]

Todo:

• [x] (WILL NOT IMPLEMENT) Add ThroughputMonitor REASON: No added value and probably moderate overhead. Measure FLOPs does what we want.

• [x] Add measure_flops This is now printed when models are instantiated (also pr. layer and also including # params)

• [x] Add option to overwrite in inference (so we can e.g. continue from broken runs) We can now use --overwrite if we wish to overwrite predictions AND results_json. Otherwise only non-existing will be predicted/evaluated

• [x] Remove additional large keys in plans (new_sizes and new_spacings for now) - which means they will also not be in the ckpt and in wandb etc

• [x] Add lm.hparams to Augmenter Now a part of the hparams.yaml and ckpt and wandb

• [x] Recursive Find Function will now let you know if it doesn't find anything (e.g. if you've got a typo in the Planner or Manager specified in the CLI) Guess this is nice to find typos and bugs mainly

• [x] Evaluator bug fix (correct results, but weird looking output for models with > 10 classes)

• [x] WandB image plots Very beautiful. Need to figure out what to do exactly with classification. For now it is not enabled.

• [x] (WILL NOT IMPLEMENT) Per label metrics during training REASON: known Torchmetrics bug, waiting for them to fix it https://github.com/Lightning-AI/torchmetrics/issues/2282

• [x] Binary evaluation in inference Run yucca_evaluation with "--as_binary" to achieve this.

• [x] Surface Dice (see: https://github.com/MIC-DKFZ/MedNeXt/blob/c5ed3f38b56d58c80581c75fea856865f42ddb75/nnunet_mednext/evaluation/region_based_evaluation.py#L112) Run yucca_evaluation with "--surface_eval" to achieve this

To be fully tested with (along with pytests):

• [x] 3D
• [x] 2D
• [x] Segmentation
• [x] Classification
• [x] Self-Supervised
• [x] Logging disabled

[Sllambias]

@asbjrnmunk and @jakobamb if you've any QOL improvements you would like to see that are simple/medium difficult to implement add here

[asbjrnmunk]

ad measure flops: When should this be run? It is a full forward pass, so it takes some time and would probably be a bit cumbersome to run always.

Otherwise sounds good.

[asbjrnmunk]

@Sllambias for flop counting you can see https://github.com/asbjrnmunk/agave/blob/main/src/count_flops.py :)

[Sllambias]

reviews of course conditioned on passing automatic test and the manual tests i've written above