Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.76%. Comparing base (393315b) to head (667c689).
Report is 2 commits behind head on main.

@@            Coverage Diff             @@
##             main     #346      +/-   ##
==========================================
- Coverage   84.27%   83.76%   -0.51%     
==========================================
  Files           8        8              
  Lines        1558     1540      -18     
==========================================
- Hits         1313     1290      -23     
- Misses        245      250       +5     

Apologies for chiming in here again, but I'm curious about the decision to remove percentiles_for_norm, and I think it would be helpful to get some pointers on how its functionality should be replaced once I pull in these changes. I've found its capacity to do per-channel normalization quite helpful when there are extreme differences in the intensities of different markers. For example, using this:

sdata.pl.render_images(
    '78_image',
    channel=["CD45", "PanCK", "DAPI"]
).pl.show(coordinate_systems='78')

I'd get an image like this:

Because of data like that, I've routinely started plotting like this:

sdata.pl.render_images(
    '78_image',
    percentiles_for_norm=(0, 100),
    channel=["CD45", "PanCK", "DAPI"]
).pl.show(coordinate_systems='78')

which for this example gives:

How would you achieve something like this using just the norm flag?

Hm, interesting. Does passing a matplotlib Normalise achieve a different outcome? I've also modified the norm parameter to no longer be a flag but instead sth that you can pass a matplotlib.colors.Normalise object.

One caveat that I see in that, this way, all channels are normalised based on the same vmin/vmax which might not necessarily work if your channels have different value ranges. We have a mostly undocumented transfunc parameter in one of the functions (currently on the go, will check later).

Is the goal to scale all channels to [0,1]? I'd be happy to build out that parameter for this purpose

yeah exactly, I think having some control over which channels get which vmin/vmax would be super helpful. Even with percentiles_for_norm I've sometimes wondered how it would look if I could set a list of (pmin, pmax) tuples with the same length as channel, to have some more fine-grained control over which channels to highlight. I suppose something similar could be achieved if it was possible to set a list of Normalize objects with the same length as channel. But achieving the task of percentiles_for_norm(0, 100) (which I think would still be the most common use-case for me) would get a lot more wordy since I'd first have to figure out the per-channel vmin/vmax corresponding to each channels min and max, and instantiate the appropriate Normalize objects to hand off to render_images.