2025 Review

[neptunes5thmoon]

Setup

• recommend moving data to our s3 bucket and/or the mount (for speed and longevity)

CARE

• students have learned about torch Datasets before - e.g. in segmentation - maybe call back to that instead of re-introducing it as a new concept
• installing tensorboard extension killed my kernel -> could have them install it at the beginning... but also it's probably fine since nothing has a long runtime till then
• need to set workspace interpreter to 05_image_restoration environment for tensorboard plugin to work

N2V

• for the mito confocal could maybe copy more of the introductory stuff about this dataset from COSDD? e.g. autocorrelation - that might help with bringing the group back together as well

COSDD

• maybe put widget code in a python file to the side? I think for code that's included in the notebook the expectation should be for students to look at it line by line and understand what's going on - I don't think that's necessary/helpful here?
• I had some trainings that still looked quite nonsensical after 15 minutes...
• loading pretrained model doesn't work: FileNotFoundError: [Errno 2] No such file or directory: 'checkpoints/mito-confocal-pretrained/training-config.yaml'
• task 8.2 crashes my kernel without an error message (with just 4 images it worked though, maybe a memory problem?)

MicroSplit

• is it correct that there's no images in hte tensorboard logs
• should add instructions to interrupt training after 15-25 minutes

Noise2Noise

• the cell after Task 2 is I think the solution?
• during training: Error rendering output item using 'jupyter-ipywidget-renderer' Cannot read properties of undefined (reading 'ipywidgetsKernel')
• prediction = careamist.predict(source=test_image[2], tile_size=(256, 256), axes="YX", tta_transforms=False)[0] throws an error about stds and means missing from the dataconfig. Not sure how to fix that

[federico-carrara]

Thanks Larissa!

For MicroSplit:

• we don't have images because we use WanDB back home and we usually don't log them... Anyway, it would show just some tiny 64x64 patches which are not very informative most of the time.
• other point is already dealt with.
• data have been uploaded to /mnt, made a PR that changes all data paths.

Tomorrow I'll take care of CARE and N2V! (@Ben-Salmon is it ok for you to focus on COSDD and N2N instead?)