Review 2024

[neptunes5thmoon]

• [ ] kernel doesn't show up in vscode, can be fixed by running python -m ipykernel install --user --name "04-instance-segmentation"
• [x] still a TODO about colormaps in Section 0
• [x] rename woodshole to tissuenet_data
• [x] the scaffolding for Task 1.3 has no create_sdt_target and it's unclear from the task that it should be float32
• [x] typo Task 1.3: __get_item__ -> __getitem__
• [x] after the dataset in Task 1.3 I started doubting myself when I saw the visualization, I think showing channel 0 is safer, maybe even adding the masks in there to
• [x] task 1.4 unclear from task what epoch in train should be
• [ ] task 2.1 I think this is hard, especially since the documentation for maximum_filter isn't all that good for explaining what it does. Maybe we can give them the line for how to compute max_filtered. Or we put a checkpoint right after so they don't get stuck
• [x] task 3.2 confusing that there's no coding to do
• [x] the results for affinities aren't that great? Maybe we can have them do a leaderboard or set an expectation for what they can reasonably expect?

[pattonw]

Thanks for the feedback.

• kernel not showing up: I need to play around with this more, I'm not sure why its not showing up. I did install ipykernel. Is there a reason we need to call ipykernel install as well as pip install ipykernel?
• maximum filter task: this seems both too complex and too simple to me since all they really have to do is call the given function, but is made complicated by the unhelpful documentation

As for the affinities section, I have added a few basics that get the affinities up to doing pretty decently. 1) weight balancing for 0/1 affinities, pretty necessary for neighboring affs since its almost all 1s 2) long range affinities (this makes it much easier to see visually that the task is different to a simple foreground/background segmentation) 3) mutex watershed call for getting the final segmentation instead of seeded watershed from distance to boundary, which sort of defeats the purpose of affinities unless you do a second agglomeration step.

I think this section could still use some better images and explanations

[neptunes5thmoon]

I think the magic line doesn't install ipykernel but rather registers the kernel somehow. Not sure exactly though, I just copied it from some other exercise. This seems to corroborate my assumptions: https://ipython.readthedocs.io/en/stable/install/kernel_install.html

Maybe checkpoint right after is a good compromise for the maximum filter part? In Milan people were struggling with this iirc (but also we were pretty rushed).

I'll try to find time to look over the affinities part again tomorrow