Dev

[hvgazula]

closes
#278
#283
#285
#289
#291
#296
#297
#302

[satra]

@hvgazula - happy to review, but perhaps see if you can fix the tests that are failing.

[satra]

also quick note as we improve the codebase, let's make sure functions/methods have type annotation and have docstrings.

[hvgazula]

Noted.

[hvgazula]

@hvgazula - happy to review, but perhaps see if you can fix the tests that are failing.

It is tensorflow again 😬 . I limited it to 2.15.* for the time being.

[hvgazula]

@satra Please see this https://github.com/hvgazula/nobrainer/actions/runs/8335069499/job/22809853267 This is a successful run of the ci.yml after restricting tensorflow. I am not sure why these checks failed in this repo. Please let me know how to proceed.

[satra]

sent a separate PR. if that passes, we can merge that and then update this.

yes tensorflow 2.16 has been released with keras 3. i would start a separate branch to update to that and restrict it to 2.16+.

[hvgazula]

The unetr is a 147M parameter model. The memory on GH runner is too limited to run this test. For the time being, I am disabling this test.

[satra]

is the unetr that big even for a block size of 32? since this is not actually creating a model could you exercise the code with a smaller model?

[hvgazula]

I tried 48 and my MBP M2 (16GB) struggled massively. The reduction in parameters was in the order of thousands. So, I don't think going down to 32 will help but I can try.

[hvgazula]

Here

[satra]

does it make sense that a reduction of 8 in each of 3 dimensions only reduces number of parameters by 3 million?