Contour Plots

[marcharper]

Add contour plots, likely using Affine2D and matplotlib's built-in contour functions.

[btweinstein]

Any progress on the contour plots? I am going to submit a publication that features python ternary soon, and some contour plots would be helpful. Should I just do this myself, or have you made progress?

[marcharper]

I've played with it -- so far I've tried to use mpl's transformations and contour functions. The main difference versus ternary's current API is how mpl expects the data (numpy matrices in a specific format), which is somewhat unusual for ternary data. This is the easiest way to add contour functionality, with the downside being a lack of control of how the contour plots are actually made (other than passing through standard mpl parameters). If that's good enough for your purposes I can try to patch it in sometime soon.

I've also played with passing through to mpl's heatmapping functions as well in a similar manner (so that heatmaps can also be generated from arbitrary data rather than specifying values on a partition as the library current expects).

[btweinstein]

I am finding that a listed colormap with a small amount of colors works well enough. Don't worry about it! Thanks for the info though.

[olivier-s-j]

@marcharper would you be so kind to share the code you mention to create a contour plot?

[btweinstein]

I recommend just creating a listed colormap with a couple of colors, that creates nice looking contours. I can dig out the code if this reply is not sufficient

[olivier-s-j]

Thx for the help. I've tired it with a discrete colormap but it does not give the result I am looking for. Code for contour lines would be helpful.

[marcharper]

@ojtwist Sorry for not responding to the request above. All I did back then was to project the data with ternary's functions and then use matplotlib's contour function. You could try something similar. I had issues getting the axes right.

Another approach would be to manually compute the contours and use ternary's curve plotting functions.

The "right" way to do this is using matplotlib's transformations (indeed the entire library could benefit), but that's a lot more work, and like @btweinstein I haven't personally encountered a situation where a contour plot would have been better than a heatmap.

[tloredo]

I haven't personally encountered a situation where a contour plot would have been better than a heatmap

For statistical inference, it's common to show the boundaries of credible regions, regions that contain a specified fraction of the probability in a ternary distribution. From a statistical perspective, what's interesting is the probability in a region, not the density. So from a statistical perspective, contours are very interesting. I think this would be a valuable addition.

[marcharper]

I'm sure there are useful scenarios for contour plots, but I don't have a personal need for them at the moment nor a lot of extra time to spend implementing this functionality. I'm happy to review and merge a PR that adds some or all of the functionality.

In the meantime, if you are able to compute the boundaries of the credible region(s) then you could easily plot them using tax.plot(). I would also guess that a heatmap would suffice to illustrate the credible regions.

[tloredo]

I wasn't really asking you to implement contours (though I was hopeful!), just making the point that there are well-motivated and not-uncommon use cases. I may look into this.

I would also guess that a heatmap would suffice to illustrate the credible regions.

My first reaction to that was no, a heatmap of the PDF (what is typically available) isn't readily useful for this. However, one could build a mapping from density to P(region with larger density) and plot a heatmap of that. It would essentially portray credible region boundaries for all choices of contained probability. I've never seen anyone do that, but it's an interesting idea that I'll try to look into.

[btweinstein]

I was able to create "contours" by using a colormap with only a couple of colors, i.e. in the below image. It's not an ideal solution, but it may be helpful for you.

[tloredo]

Thanks; the same idea had occurred to me!