[BUG] Incorrect Spatial Dimension Assignment for apply_filter_bank() output (Fix Included)

[mfs4rd]

Summary

I'm using this repo to deconvolve brightfield images, however there is a mistake in waveorder.filter that prevents the code from working.

Code for reproduction
Deconvolution Parameters

transfer_function_arguments = {
    'wavelength_illumination': 0.532,
    'z_padding': 0,
    'numerical_aperture_detection': 1.25,
    'numerical_aperture_illumination': 0.5,
}

simulation_arguments = {
    "zyx_shape": (128, 256, 256),
    "yx_pixel_size": 0.1625,
    "z_pixel_size": 0.5,
    "index_of_refraction_media": 1.33,
}

Input images have a dimension of 245x2048x2048

Error Traceback

Traceback (most recent call last):
  File ".../CytoLand/data/far-red/hcs.py", line 157, in <module>
    main()
  File ".../CytoLand/data/far-red/hcs.py", line 91, in main
    img_phase = phase_thick_3d.apply_inverse_transfer_function(
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../miniconda3/envs/viscy/lib/python3.11/site-packages/waveorder/models/phase_thick_3d.py", line 264, in apply_inverse_transfer_function
    f_real = apply_filter_bank(inverse_filter[None, None], zyx[None])[0]
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../miniconda3/envs/viscy/lib/python3.11/site-packages/waveorder/filter.py", line 114, in apply_filter_bank
    padded_output_spectrum[output_channel_idx] += stretched_multiply(
RuntimeError: The size of tensor a (256) must match the size of tensor b (2048) at non-singleton dimension 2

Error Cause
The reason for this error is found here https://github.com/mehta-lab/waveorder/blob/aa7050f973f36dc204ea495e105b5432223dc68d/waveorder/filter.py#L77

The spatial dimensions for padded_output_spectrum is being determined by the shape of the filter, not the padded_input_spectrum. This means that when the result of streched_multipy() is added to the output array at line 114 there is a dimension missmatch.

Fix
Remove line 77 and add spatial_dims = padded_input_spectrum.shape[1:] after line 91. This allows the deconvolution to run successfully

Environment:

Operating system: Ubuntu 24.04
Python version: 3.11
Repo Version: 2.2.1

[talonchandler]

Thanks for reporting @mfs4rd!

On main I found that line 77 was already gone, so I poked through git history and I think I fixed this bug in this commit as part of #440. Are you able to retry from main?

We'll cut another another release soon...apologies that this bug shipped in the release you're using. Let me know how you make out.