FIX: Image scaling for large dynamic range ints

[jklymak]

PR Summary

This fixes #10072.

The problem was that for a uint32 image, the dynamic range was too high for a float32 to properly represent. Simply changing to float64 fixes the issue. I suppose someone could use a unit64 image, but...

Test coming: code:

import matplotlib.pyplot as plt
import numpy as np

img = np.load('/Users/jklymak/downloads/test_image.npy')
img = np.array([[1, 2, 3], [1, 2, 10000000000]], dtype=np.uint32)
#Display full image
plt.imshow(img, interpolation = 'nearest')
plt.clim(0, 5.)
plt.show()

PR Checklist

• Has Pytest style unit tests
• Code is PEP 8 compliant
• New features are documented, with examples if plot related
• Documentation is sphinx and numpydoc compliant
• Added an entry to doc/users/next_whats_new/ if major new feature (follow instructions in README.rst there)
• Documented in doc/api/api_changes.rst if API changed in a backward-incompatible way

[anntzer]

That's not going to help the efforts on trying to reduce the memory footprint of imshow... (#6952, #8143, etc.). Not a blocker, but should be kept in mind.

[jklymak]

Yeah, understood. This is the dumbest solution.

I don't quite understand why this rescaling is necessary, despite the nice comments....

[jklymak]

OK, I only used the bigger float64 if the amin-amax > 1e8.

[anntzer]

I think(?) you probably care more about maxabs/minabs.

[jklymak]

Hmmm, not at all sure, because I once understood all the low-level reps of floats etc, but not so good at it anymore. Happy if someone wants to pick this up and do it better.

But, I think the error comes in when we do:

matplotlib/lib/matplotlib/image.py

Lines 389 to 390 in 884060a

	if a_min != a_max:
	A_scaled /= ((a_max - a_min) / 0.8)

and the resulting float doesn't preserve all the bits in the unit32.

[tacaswell]

This is a reasonable approach.

The scaling is to be able to track the over/under pixels and get around ringing in the interpolations.

[tacaswell]

Discussed this on the phone call. Will investigate

• delaying conversion to float as long as possible
• avoiding float128 due to platform-dependence
• add a test

[jklymak]

Test added, and changes requested above made... Thanks!

[dopplershift]

Working and slow is better than broken and fast. Happy to merge a future PR that keeps things working and makes it faster.