🧑‍💻 Define PatchPredictor

[shaneahmed]

• Redesigns PatchPredictor engine using the new EngineABC base class.
• The WSIs are now processed using the same code as for the processing the patches using WSI based dataloader.
• The intermediate output is saved as zarr for the WSIs to resolve memory issues.
• The output of model architectures should now be a dictionary.
• The output can be specified as AnnotationStore for visualisation using TIAViz.

[Abdol]

Thank you @shaneahmed and toolbox team for committing such a good revamp of PatchPredictor. I left only minor feedback.

[measty]

I have done an initial review and commented a few things that need changing.

A few other notes as well: this is awkward to review as some changes are wrapped up in the branch this merges into - some of the issues I have found don't show up in the code changed for this PR cause the actual change is in the dev-update-engine-abc branch.

There will need to be a review of the notebooks before this gets merged, as this will at the very least break the patch-prediction notebook and perhaps others.

Also, is the change of behaviour for a model like "resnet18-kather100k" intentional? previously it would return hard class predictions also, now it does not

[measty]

If preds can be empty list this function wont work. Need to iterate over coords instead which always must be there, and save preds only if they exist. At the moment using patch predictor on a WSI with save_type="AnnotationStore" just saves an empty store

[shaneahmed]

Addressed in this commit 5085035

[measty]

There is still some issues with how it is being saved.

1. the correct filename isn't being passed to engine.save_predictions, so the saved .db is just called the default output.db. It is being set in:
   

   tiatoolbox/tiatoolbox/models/engine/engine_abc.py

   Line 1041 in 45a339b

   out = {image: save_dir / (str(image.stem) + suffix) for image in self.images}

   
   but this is not being used anywhere.
2. The scale_factor is not being set and passed to save_predictions, so the patches are currently saved and can be visualized, but appear wrongly scaled. The scale factor needs to be calculated as model_mpp (from io config) / slide_mpp somewhere in the pipeline and added to kwargs.