segretini-matplottini

A collection of useful Matplolib and Seaborn tips & tricks collected over years of colorful plot-making.

Installation

Simply run the following.

git clone [email protected]:AlbertoParravicini/segretini-matplottini.git
pip install segretini-matplottini

Repository Structure

• segretini_matplottini/utils/plot_utils.py contains many useful functions commonly used during plotting, to add labels above bars or writing nice labels with exponential notation.
• segretini_matplottini/utils/data_utils.py contains functions useful for data preprocessing, such as removing outliers, computing and speedups.
• segretini_matplottini/plots contains custom plotting functions that can be used like standard Seaborn plots.
• notebooks contains examples that show how to create very complex custom prompts. Look at them if you are trying to replicate some specific feature, like having separate bar groups or adding fancy custom annotations.
• notebooks/plot_scratchbook.py contains common commands used to setup plots: creating plots, customizing tick labels, adding custom legends, etc. If you don't remember how to customize some part of your plot, this is a good starting point.
• data contains files used for plots. For the most part, you can ignore it.
• plots is where all the plots are stored. If you find something you like, the code to replicate it is in notebooks.

Tips and Tricks

An ever-growing collection of tips I've found or discovered along the way, together with some nice resources I like a lot.

Resources

• Subtleties of Color: a six-parts guide on how to pick nice colors and create great palettes, from the Earth Observatory NASA blog.
• Adobe Color: a free web tool to create palettes of different types (shades, complementary, etc.). It also has accessibility tools to test for color blindness safety.

Colors

Picking the right colors is hard! I found the following tips to be very helpful.

• Not everyone sees colors in the same way: most reviewers will look at your papers after printing them in black & white. Always check for that! You can do it with matplotlib (check out hex_color_to_grayscale in src/plot_utils.py) or doing a print preview after saving your plot as PDF. If colors are too similar, try adjusting the L (lightness) in the HSL representation, or the B (brightness) in the HSB representation. Also, a lot of people are color blind, and there are many types of color blindness! It's always better to double check.

• Hidden color biases: people tend to associate implicit meanings to colors. To simplify the matter a lot, green is usually associated to positive things, while red is bad. If your plot is not explicitely comparing classes (for example, you want to show the speed of your algorithm on different datasets), just go for a neutral/positive palette, using colors such as gree, blue, and light pink.

• Add redundant information: if you are plotting many different classes, and use one color per class, it can be difficult to distinguish among them. Instead, add some kind of redundant information. In scatterplots and lineplots you can use different markers (circles, diamonds, etc.), while in barplots you can use different hatches (//// or \\) or add labels to each class.

Update 2022-09-25

I'm revamping the structure of the repository, to make it easier to integrate in other repositories and create pretty plots.

• I also added a new legend style, and squashed many bugs in plot_utils.py

Update 2021-08-28

Update style of Ridgeplot to be readable in black & white. Added large layout to Ridgeplot

Updates to plot_utils.py: added option to directly provide vertical coordinates to add_labels. Added better outlier removal based on interquantile range (the same approach used to find outliers in box-plots)

Added notebooks/plot_performance_scaling.py: this plot shows the relative performance increase of processors, memory and interconnection technologies from 1996 to 2021. Shamefully copied from AI and Memory Wall by Amir Gholami. This plot shows how to use dates for the x-axis, and do fairly complex visualization on log-scale axes (i.e. linear regressions on data with exponential increase).

Update 2021-03-22

Updated Ridgeplot to have confidence intervals and be more user-friendly (notebooks/plot_ridgeplot.py). Added some general tips about choosing colors.

Update 2021-03-20

Added Correlation Scatterplot, find it in notebooks/plot_correlation_scatterplot.py. Minor updates to plot_utils.py: new palettes, improved robustness of get_exp_label, minor bugfixes.

Update 2020-11-25

Added Roofline Plot, find it in notebooks/plot_roofline.py.