New lazy to_multiscale() for labels

[LucaMarconato]

Closes #1069

This PR (still work-in-progress) aims at implementing a function to generate a pyramidal image given a dask array, by relying on ome-zarr internal APIs and by extending functionalities currently not available via ome_zarr.scale.Scaler. Such a function is also provided by the multiscale-spatial-image package, but as explained in the linked issue we hit a performance roadblock.

ome-zarr-py does not support downscaling of the z dimension, nor downscaling via scale factors specified per dimension (both cases are supported by multiscale-spatial-image); these cases will be covered by this PR. The implementation could also live in ome-zarr-py (CC @will-moore @joshmoore @jo-mueller), I'd be happy if someone later can pick up the code so I can remove it from spatialdata and use it from upstream.

Side technical note:

• the previous implementation in spatialdata uses xarray_coarsen (from multiscale-spatial-image) for images, (which is fully lazy), and itk_nearest (from multiscale-spatial-image) for labels (which is not lazy and therefore problematic). Since downscaling for images doesn't give problems, we could use the old implementation, but I'm reimplementing it with ome-zarr-py APIs anyway. So far it seems that both implementations lead to identical results when the scale factors are all 2.

[jo-mueller]

Hi @LucaMarconato ,

PR ome/ome-zarr-py#516 also addresses the issues with the downscaling in z (now built in for all methods) 👍

[LucaMarconato]

That's perfect! Then I'll use that PR instead.

[codecov]

Codecov Report

❌ Patch coverage is 97.97980% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.93%. Comparing base (e88c24e) to head (0d4ae59).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/spatialdata/models/pyramids_utils.py	94.87%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1068      +/-   ##
==========================================
+ Coverage   91.89%   91.93%   +0.04%     
==========================================
  Files          49       51       +2     
  Lines        7598     7664      +66     
==========================================
+ Hits         6982     7046      +64     
- Misses        616      618       +2     

Files with missing lines	Coverage Δ
src/spatialdata/_io/io_raster.py	93.84% <100.00%> (-0.10%)	⬇️
src/spatialdata/datasets.py	100.00% <100.00%> (ø)
src/spatialdata/models/chunks_utils.py	100.00% <100.00%> (ø)
src/spatialdata/models/models.py	88.42% <100.00%> (ø)
src/spatialdata/models/pyramids_utils.py	94.87% <94.87%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[LucaMarconato]

The PR is ready. I'll merge and test it with downstream libraries, as it will provide a significant performance boost that I want to make available soon. Additionally, I've created an issue (#1070) to align the code with the z-support from ome-zarr-py once the corresponding PR is merged (ome/ome-zarr-py#516).

[LucaMarconato]

I actually now also added the per-dimension scaling to the tests. I put some plots here to show it in action. Here the actual numerical values of the pixels differ between resize and coarse (expected). So I relaxed the tests.