Scale flows to correct shape

[mrariden]

Per #774, when resize is not None, masks get resized but the flows don't in dynamics.compute_masks(). This isn't immediately an issue, except that if no putative cellpix are found (here), zero-filled arrays of the original size are returned.

The result is that, depending on if putative cellpix are found, iterating through the images (here) produces a list of inconsistent np array shapes. Finally, trying to convert the list-of-arrays into a stack of arrays fails at line 663.

This fix reshapes the flows to the correct shape.

Should we think about tests that address this?

[codecov]

Codecov Report

Merging #775 (00d0ed1) into main (95f8d98) will increase coverage by 0.02%.
Report is 33 commits behind head on main.
The diff coverage is 78.94%.

@@            Coverage Diff             @@
##             main     #775      +/-   ##
==========================================
+ Coverage   63.39%   63.41%   +0.02%     
==========================================
  Files          14       14              
  Lines        2964     2966       +2     
==========================================
+ Hits         1879     1881       +2     
  Misses       1085     1085              

Files	Coverage Δ
cellpose/models.py	86.12% <100.00%> (ø)
cellpose/dynamics.py	56.79% <76.47%> (+0.23%)	⬆️

[carsen-stringer]

technically the correct thing to do is to leave them in the unresized form, so let's return a resized array if cellpix is none

[mrariden]

I think I see, my solution works but it shouldn't because compute_masks should respect the resize or rescale arguments? (I am still unsure what the convention is for rescale, vs resize, reshape, resample.)

Here's the data in case I didn't describe it well enough before. This is the state 5 loops through the iterator on line 655 in _run_cp()

The input dP being processed has a shape of (2, 59, 77, 77) after tiling:

>>>[dPi.shape for dPi in dP[0][0:5]]
[(77, 77), (77, 77), (77, 77), (77, 77), (77, 77)]

>>>[dPi.shape for dPi in dP[1][0:5]]
[(77, 77), (77, 77), (77, 77), (77, 77), (77, 77)]

And the output masks are reshaped because of this line:

>>>[m.shape for m in masks]
[(232, 232), (232, 232), (232, 232), (232, 232), (232, 232)]

But the p array is not reshaped and has inconsistent size as a result:

>>>[pi.shape for pi in p]
[(2, 232, 232), (2, 232, 232), (2, 77, 77), (2, 232, 232), (2, 77, 77)]

So, I have two thoughts, (1) should the resize be the opposite here to be resize = [shape[1], shape[2]] if resample else None (remove the not) ? And (2), I think we should remove the resizing from compute_masks() since that seems to be a different function.

[mrariden]

Okay new solution refactors dynamics.compute_masks() to dynamics.reshape_and_compute_masks() which calls dynamics.compute_masks() which has no internal reshaping behavior.