A Better Default Colormap for Matplotlib | Hacker News Reader

A Better Default Colormap for Matplotlib

bids.github.io

114 pointsmartian23 comments

When it comes to plot colors in matplotlib, it's why I start with

import seaborn as sns

to get a nice plot theme, not "good old 90s" style. See: https://speakerdeck.com/pmigdal/teaching-machine-learning?sl...

This, changes everything.

The theme seaborn uses is actually a direct clone from ggplot2 from R

Yes. Though, I think that 'whitegrid' is a cleaner theme (if there were axes).

For ggplot theme, in matplotlib there is

plt.style.use('ggplot')

'ticks' plus sns.despine() is my favorite.

This will be the default in Matplotlib 2.0, and a beta was just tagged: http://article.gmane.org/gmane.comp.python.matplotlib.announ...

In the talk when he showed the 3 similar maps getting equal score, and then the green one getting a much higher score, that could easily be because people couldn't decide between the 3 similar maps, to they went with the one different map. To do a proper comparison the test should be 1-1 or 3-3, not 3-1.

They wanted green because it's a python library. He goes into the details in the talk

OP here: really worth watching the embedded video on this topic. Viridis is born out of an awesome, intense deep-dive into research on human vision and perception.

I'm mostly using seaborn (https://web.stanford.edu/~mwaskom/software/seaborn/) to wrap matplotlib graphs but great to see a direct alternative into the matplotlib library.

If you're interested in colour maps, cubehelix [1] is a useful scheme to make intensity maps which display well when converted to greyscale.

[1] https://www.mrao.cam.ac.uk/~dag/CUBEHELIX/

I ported these color scales to d3-scale and D3 4.0:

https://github.com/d3/d3-scale/blob/master/README.md#sequent...

I started using viridis a few weeks ago and I love it!

Is it strange that I find the old "JET" the best? It uses the most hues (green, yellow, red and blue are all there), so things are the most distinguishable. The others are either only blue/green/yellow, or only purple/red/yellow.

Quick, which represents a larger value, green or red?

Having a gradient between two colors, from dark to light, makes it easier to determine the scale. Rainbow gradients look nice, but hue transitions are nowhere near as intuitive as luminance.

I think you're both correct, but your points respond to different needs. For some types of results (and audiences and modes of presentation) its more important that the reader be able to determine the approximate value that a color corresponds to than that the reader be able to quickly intuit the progression of values. For me, Jet is better at the former while Viridis et al are better at the latter.

Yes, that's the point. People like it a lot but are worse at actually interpreting it. It's confusing when people have to read lots of confusing data quickly

I dislike that the whole page, including the updates, do not contain any dates...

Good point. I'll ping authors to update page with dates.

Javascript version: https://github.com/politiken-journalism/scale-color-perceptu...

If you use gnuplot, which I do exclusively, you may like this repository:

    https://github.com/Gnuplotting/gnuplot-palettes

I'm pleased. They look extremely similar to the gnuplot and gnuplot2 color maps already present in matplotlib, my favorites.

I also enjoy using Olga Botvinnik's prettyplotlib: github.com/olgabot/prettyplotlib

That was an amusing talk. I had fun watching it.

Reply on news.ycombinator.com