Rescale image data if outside float32 precision

[Czaki]

References and relevant issues

closes #6533

Description

Add check if image data could be visualized using vispy and if not, then rescale data.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (dab2a86) 92.29% compared to head (d146dd0) 92.27%.
Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6537      +/-   ##
==========================================
- Coverage   92.29%   92.27%   -0.03%     
==========================================
  Files         604      605       +1     
  Lines       54037    54185     +148     
==========================================
+ Hits        49876    49999     +123     
- Misses       4161     4186      +25     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jnahlers]

Thanks @Czaki, this method looks like it handles the toy example in the 'Steps to Reproduce' of #6533. However, I am still having trouble with my actual images. An example is this image.

[Czaki]

@jnahlers Your case is different from described in the original issue. In this case, is small contrast limits for big values. But it looks like I could easily fix the code to work with your data.

[jnahlers]

It looks like there is still some kind of overflow happening? Check out this array, specifically around index 50 (i.e. [50,:,:]).

[Czaki]

Now It should be better. I forgot to cast infinity to float32.

[psobolewskiPhD]

@Czaki I think that test fail is real related to the uninitialized float64 empty array. Not sure if switching those arrays to zeroes would just mask something?

[Czaki]

@Czaki I think that test fail is real related to the uninitialized float64 empty array. Not sure if switching those arrays to zeroes would just mask something?

No. The error was the result of mistaking scale and offset. For testability, I move logic to CoercedContrastLimits.

[jnahlers]

@Czaki I am still having trouble with the examples I gave above . Is this the issue of small contrast limits for big values you mentioned above?

[Czaki]

@jnahlers It looks like I have made a bug when merging main into this PR. I have fixed it now.

[Czaki]

As far as I could check, the tests are failing now because of NaN. So it looks like using np.empty on float is a bad idea. So it should be either a predefined array or int data.

Random is bad because its slowdown test suite.

cc @psobolewskiPhD

[psobolewskiPhD]

Even with int could we get something strange, given that fact that it's uninitialized?
Maybe best to use zeros?

Going forward, could we make a set of arrays of different size/shapes using random.random and then reuse them everywhere else?

[Czaki]

No. Int cannot be something strange.

We could create preset of data using fixtures.

[jnahlers]

@Czaki Ping me if you want to me try the branch again, right now I'm still seeing black on the examples. Thanks for you work on this!

[Czaki]

@Czaki Ping me if you want to me try the branch again, right now I'm still seeing black on the examples. Thanks for you work on this!

I have removed an important change when removing debug print....

[Czaki]

This is how it looks on my machine:

napari_scale-2024-01-30_11.59.36.mp4

I do not understand the source of vertical artifacts. But as it looks like thumbnail do not have such artifacts, maybe rescale up could be done smarter.

[Czaki]

@jnahlers Is it ok now?

[jnahlers]

@jnahlers Is it ok now?

Looks good!