Prototype for annotator plugins

[constantinpape]

Minimal prototype for annotator plugin.

This is an attempt to refactor the current implementations of annotation tools into proper napari plugins.
It contains a minimal working example for the 2d annotation tool (not yet all features added).

Design

• The base class sam_annotator._annotator._AnnotatorBase initializes all shared functionality for the annotation plugins.
• The annotation tools inherit from this class and add their specific functionality, e.g. sam_annotator.annotator.Annotator2d.
  • (There we could also implement a function as entry point from python / the CLI so that the tool can be started without going through the napari plugin interface).
• In addition, all widgets will be refactored to _widgets.py, and some more clean up.

This would address #200 and #206.

[constantinpape]

Need to update test/test_gui.py to match these changes!

[constantinpape]

I have now ported all annotation tools to the new plugin design and everything is working!
The overall design approach is summarized here: https://github.com/computational-cell-analytics/micro-sam/blob/b440dcb493502721ff99ba0f4be1077905d0e892/doc/development.md

Next steps:

• Before merging this PR we will merge the current dev into master and make a 0.4.0 release (to publish the recent improvements to the training methodology).
• The core plugin functionality in this PR is working, but there are a lot of smaller follow-ups to beautify the plugin and extend functionality. We will work on this for version 0.5.0.
  • This will also include support for the new decoder based instance segmentation in napari, see Implement new segmentation plugin #305.

[GenevieveBuckley]

This class based creation of dock widgets is a bit more heavy duty that the approach I tried (involving a magicgui container stuffed with the existing widgets). I hope/think that will make it more robust.

I'm broadly supportive of this design. Note that I've mostly looked at the code, and haven't spent time playing with it in the GUI, eg: to check things like whether the embeddings don't get recomputed unnecessarily, or that computing embeddings, then changing the selected input image produces something sensible, etc.

We've discussed this previously, and I still agree that refactoring to reduce duplication (eg: combining the 2D and 3D annotation widgets, etc.) is a good long term goal. Not in this PR, but something later. I think that structure you've made with this design will probably help with that long term goal. It seems a little easier to understand how the pieces fit together here.

[constantinpape]

Thanks for the review @GenevieveBuckley !

I'm broadly supportive of this design. Note that I've mostly looked at the code, and haven't spent time playing with it in the GUI, eg: to check things like whether the embeddings don't get recomputed unnecessarily, or that computing embeddings, then changing the selected input image produces something sensible, etc.

I have tried each tool a couple of times and everything seems to work. Of course there might still be some more subtle bugs or issue (e.g. that everything is set correctly when updating the image etc.).
I will do some more testing of this later, and we will also test it more in the lab before doing an actual release.

We've discussed this previously, and I still agree that refactoring to reduce duplication (eg: combining the 2D and 3D annotation widgets, etc.) is a good long term goal. Not in this PR, but something later. I think that structure you've made with this design will probably help with that long term goal. It seems a little easier to understand how the pieces fit together here.

Yes, there's probably some more room to combine functionality here in follow up work.

Regarding your more in-detail review and next steps:

• I have addressed the comments about the tracking util functions and the tests.
• I will make dedicated issues out of TODOs when finalizing this PR (and open a few more follow up issues).
• And I will do that as soon as we have made a new release with the updated training functionality and new fine-tuned models. (Or have merged the current dev into master as preparation for this, hopefully by the end of this week.)

[GenevieveBuckley]

I have tried each tool a couple of times and everything seems to work. Of course there might still be some more subtle bugs or issue (e.g. that everything is set correctly when updating the image etc.).

That's great! 😄 Yes, of course unexpected issues can always pop up later. It sounds like this is a good balance between doing enough real world testing to catch any obvious problems, but not too much that holds the project back.

[GenevieveBuckley]

I see this PR is now out of draft state. I don't have any substantive new comments to add since our last discussion.

[constantinpape]

I see this PR is now out of draft state. I don't have any substantive new comments to add since our last discussion.

Thanks @GenevieveBuckley. I moved this out of draft because I have merged dev into master now and made a new release and wanted to wait with merging this one until that.
I will clean up and merge this PR today.