Bug contourf

Description

I'm running into an error after upgrading the proplot version . The error occurs when using contourf, the error message is KeyError: 'red'

Steps to reproduce

# your code here
# we should be able to copy-paste this into python and exactly reproduce your bug

import proplot as plot
import matplotlib.pyplot as plt
import numpy as np

f, axs = plot.subplots(ncols=1, nrows=1) 

#generate random data
data=np.ones((180,360))
data[20:80,80:150]+=1

x=np.arange(0,360,1)
y=np.arange(0,180,1)

axs.contourf(x,y,data,cmap=plt.cm.coolwarm)

plt.show()

Expected behavior: A simple contourf plot

Actual behavior: An error message occured and the plot was not produced

Equivalent steps in matplotlib

Pretty sure its Proplot. With matplotlib this works

import matplotlib.pyplot as plt
import numpy as np

f,axs=plt.subplots()

#generate random data
data=np.ones((180,360))
data[20:80,80:150]+=1
x=np.arange(0,360,1)
y=np.arange(0,180,1)

axs.contourf(x,y,data,cmap=plt.cm.coolwarm)
plt.show()

Proplot version

Matplotlib: 3.4.1
Proplot: 0.9.1

[matthias-k]

I think I ran into the same bug in a slightly different setting, which might help to debug the issue:

import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import proplot as pplt

f, axs = pplt.subplots(
    nrows=1, ncols=1,
)

data = np.random.randn(2, 100)
axs[0].hist2d(
        data[0], data[1],
        cmap=plt.cm.gray_r
)
axs[0].arrow(-1, 0, 1, 0, color='blue')

axs[0].hist2d(
        data[0], data[1],
        cmap=plt.cm.gray_r
)

Here, the second hist2d function also throws the 'red' error. Interestingly, the example works if I

• don't use cmap=plt.cm.gray_r with either of the two hist2d calls
• or remove the first hist2d call completely
• or use plt.subplots instead of plt.subplots
• or create the axes simply via ax=plt.gca()

So, to summarize, you really need to call hist2d twice, with a call to arrow in between, and all of that on an axes object that was created via pplt.subplots. The second hist2d call can also go to a different axes, it doesn't have to be the same as the first axes.

[matthias-k]

I did some more digging. I know that this code worked with an older version of proplot, probably around v6.4. With matplotlib 3.4.2, the bug is already present in version 0.7, and earlier versions of proplot seem to be incompatible with matplotlib 3.4.2. Therefore I installed matplotlib 3.3.4, and found that the bug is not yet present in 0.6.4, but present in 0.7.0.

To pinpoint the source of the bug further, I ran git bisect between 0.6.4 and 0.7.0. Unfortunately, there were a lot of commits with other bugs (usually 'tuple' object has no attribute 'cmap'), so git bisect could not exactly locate the commit that introduced the bug, but it claims it could reduce it to only two commits:

There are only 'skip'ped commits left to test.
The first bad commit could be any of:
2f87b04a5ac06ceeff355c60d6b3e0be984060be
3d8e7dd0ce16e1f9ac98b430899ecead94648959
We cannot bisect more!
bisect run cannot continue any more

So apparently, it's either 2f87b04 or 3d8e7dd

Let me know if there is more that I can do to help!

[lukelbd]

Thanks for the report @raphiclimate and the help @matthias-k.

It looks like the segmentdata dictionary is for some mysterious reason cleaned out in the colormap instances stored under the cm module: plt.cm.coolwarm._segmentdata returns {}. For now, simply using the colormap names instead of these instances should work, e.g. cmap='coolwarm' or cmap='gray'. But I'll look into this.

[lukelbd]

To be clear that's the source of the KeyError -- normally segmentdata contains red, blue, and green channel transitions (or hue, luminance, saturation for a PerceptualColormap). It looks like just importing proplot and nothing else breaks the colormaps stored under the cm module. But the equivalent registered colormaps still work just fine. Strange.

[matthias-k]

@lukelbd thanks for the quick response. Your workaround saved me a lot of work (I have a submission deadline coming up and was already thinking about setting up a new enviroment with the old proplot version just for adapting a few figures ;)).

An additional clarification: I guess it's not only the import. Otherwise I would expect that (in my example from above) already the first hist2d call fails, and not only the second one after the arrow call.

[lukelbd]

No problem. Thanks for your clarification -- after that I found the solution relatively quickly (182a186).

The Colormap constructor function converts input native matplotlib colormaps to proplot's colormap subclasses (i.e. ListedColormap to DiscreteColormap, LinearSegmentedColormap to ContinuousColormap). There was a bug in the conversion process that caused erasure of the original segmentdata. This happened the first time the native matplotlib colormap was passed through Colormap (i.e., the first hist2d call). To fix this, I simply make sure to copy the native colormap segment data dictionary before using it for the new colormap.

Thanks @matthias-k for digging further and @lukelbd for addressing the issue so quickly!