Tensorflow segmentation

[philippotto]

Open Issues:

• Delete old z-slice predictions from the predictions map to free up space
• Multiple Magic Wand "instances" can be spawned
• Magic Wand only works in "positive" z-direction (maybe fine for now)
• Should the Magic Wand be disabled by default and only be enabled via dev console or feature flag?
• Update the tensorflow model to the fully trained new version

URL of deployed dev instance (used for testing):

• https://tensorflowsegmentation.webknossos.xyz

Steps to test:

• Open a volume tracing.
• Use Ctrl + Click on a segment to trigger the Magic Wand. A toast should open and inform you of the process.
• Use the brush to correct the prediction and move in positive z-direction through the dataset.
• The magic wand mode should always predict the next 3 slices, then wait until the user moved "forward" and then continue predicting.
• Close the modal to stop the magic wand mode.
• The magic wand should also work on the XZ and YZ viewports, but the current model doesn't perform very well on these (will be improved with the new model).
• Press Alt in any mode to move the viewport in a photoshop-like manner.

Issues:

• fixes Finalize first iteration of magic wand #3960

---

- [ ] Updated changelog
- [ ] Updated migration guide if applicable

• Needs datastore update after deployment
• Ready for review
• add new routes to wk-connect or document the need

[daniel-wer]

@philippotto I'd say this is ready for a first review round. I cannot assign you as a reviewer, because you originally created this PR ^^ I'm especially looking for ways to improve the code in the volumetracing_saga. It may make sense to talk about this in person after you had a quick glance at the code in question :)

I've also updated the PR description and listed the remaining ToDos, feel free to add more stuff there.

[philippotto]

Nice 👍 The code looks really good. I left some suggestions/questions.

Also, I'd suggest to extract the entire infer-saga into an own file. Then, the helper functions within that saga can also be extracted to the top-level so that the main saga is not so long :)

[philippotto]

Just a reminder to set this to false.

[daniel-wer]

@philippotto I've finished up my Todos, maybe you could have another look at the code and do a final test run?
The redux-saga update was not too involved and I tested the main functionality of our sagas afterwards, everything worked as expected 🤞

[philippotto]

Nice! Code refactoring looks very good 👍 Testing worked well, too. Only thing, I noticed is that the auto-annotation stops if I'm moving too fast along the w direction (e.g., z). I thought you said that this wouldn't happen? However, I'm fine with this.

You can approve this PR now ^^ I can't because I created it.

[daniel-wer]

I thought you said that this wouldn't happen?

Yes that's what I thought, but it does indeed stop if the user moves too far ahead. Let's rethink what should happen if the user moves too far away while using the auto brush as part of the next iteration :)