Description

faressc

opened on Aug 15, 2024

Issue body actions

Bug summary

Plotting an equal numpy.array of dtype float64 and float32 with pyplot.imshow and a defined colors.LogNorm as norm will produce different colors in the resulting image.

Code for reproduction

import matplotlib.pyplot as plt
import matplotlib.colors as colors
import numpy as np

norm = colors.LogNorm(vmin=1e-10, vmax=1)
test_numpy = np.array([[1e-4]])
test_numpy_32 = np.array([[1e-4]], dtype=np.float32)

for test in [test_numpy, test_numpy_32]:
    plt.imshow(test, norm=norm)
    plt.title(f"numpy array of dtype: {test.dtype}")
    plt.colorbar()
    plt.show()

Actual outcome

The resulting plot for the dtype float64 numpy.array looks like this:

and the plot for the dtype float32 numpy.array looks like this:

Expected outcome

I would expect both outputs to have the same color.

Additional information

No response

Operating system

Arch

Matplotlib Version

3.9.1.post1

Matplotlib Backend

module://matplotlib_inline.backend_inline

Python version

Python 3.12.4

Jupyter version

No response

Installation

pip

Activity

1. tacaswell commented on Aug 15, 2024

   

   tacaswell

   on Aug 15, 2024

   Member

   More actions

   I can confirm this, however it is related to having singular arrays:

   import matplotlib.pyplot as plt
   import matplotlib.colors as colors
   import numpy as np
   
   norm = colors.LogNorm(vmin=1e-10, vmax=1)
   test_numpy = np.array([[1e-4, 1e-5, 1e-1]]).T
   test_numpy_32 = np.array([[1e-4, 1e-5, 1e-1]], dtype=np.float32).T
   
   fig, axl = plt.subplots(1, 2)
   
   for ax, test in zip(axl, [test_numpy, test_numpy_32]):
       im = ax.imshow(test, norm=norm)
       ax.set_title(f"numpy array of dtype: {test.dtype}")
       fig.colorbar(im)
   
   plt.show()

   works correctly.
2. tacaswell

   added this to the v3.9.3 milestone on Aug 15, 2024
3. tacaswell commented on Aug 15, 2024

   

   tacaswell

   on Aug 15, 2024

   Member

   More actions

   ax.imshow(test, norm=norm, interpolation_stage='rgba')

   Also fixes the problem so it is definitely in our data-value resampling code.
4. tacaswell

   modified the milestones: v3.9.3, v3.10.0 on Aug 15, 2024
5. tacaswell commented on Aug 15, 2024

   

   tacaswell

   on Aug 15, 2024

   Member

   More actions

   I think this is fixed by #28122 but that seems too big of a change to backport.
6. timhoffm commented on Aug 15, 2024

   

   timhoffm

   on Aug 15, 2024

   Member

   More actions

   I suspect that it may be an effect of the limited precision. float32 can only store 7-8 significant digits, but the log scale is set to 10 decades. Depending on how the exact calculation is done internally, the precision limit can translate to a macroscopic effect. A quick test would be to check if this goes away when limiting the log range to 6 or 7 decades.
7. ksunden commented on Aug 15, 2024

   

   ksunden

   on Aug 15, 2024

   Member

   More actions

   Indeed, changing the bounds improves it significantly (I zoomed in pretty far, didn't try too many range sizes):

   
   
8. jklymak commented on Aug 15, 2024

   

   jklymak

   on Aug 15, 2024

   Member

   More actions

   This is indeed fixed on master for both interpolation_stage='rgba' (the new default) and interpolation_stage='data' (the current default). I don't think we are going to backport either #28122 or #28061, but I also don't think we should fix this just for a point release as this would be a long-standing behaviour. Currently the workaround is interpolation_stage='rgba'. I'll close as already fixed on master
9. jklymak

   closed this as not plannedon Aug 15, 2024
10. faressc commented on Aug 15, 2024

    

    faressc

    on Aug 15, 2024

    Author

    More actions

    Thanks for the quick answers and happy to hear that it has been already fixed. I'll use the workaround interpolation_stage='rgba' for now :)
11. ksunden

    changed the title [-][Bug]: Plotting numpy.arrray of dtype float32 with pyplot.imshow and specified colors.LogNorm produces wrong colors[/-] [+][Bug]: Plotting numpy.array of dtype float32 with pyplot.imshow and specified colors.LogNorm produces wrong colors[/+] on Nov 1, 2024

Issue actions