Why UltraPlot?

Matplotlib is an extremely versatile plotting package used by scientists and engineers far and wide. However, matplotlib can be cumbersome or repetitive for users who…

  • Make highly complex figures with many subplots.

  • Want to finely tune their annotations and aesthetics.

  • Need to make new figures nearly every day.

UltraPlot’s core mission is to provide a smoother plotting experience for matplotlib’s most demanding users. We accomplish this by expanding upon matplotlib’s object-oriented interface. UltraPlot makes changes that would be hard to justify or difficult to incorporate into matplotlib itself, owing to differing design choices and backwards compatibility considerations.

This page enumerates these changes and explains how they address the limitations of matplotlib’s default interface. To start using these features, see the usage introduction and the user guide.

Less typing, more plotting

Limitation

Matplotlib users often need to change lots of plot settings all at once. With the default interface, this requires calling a series of one-liner setter methods.

This workflow is quite verbose – it tends to require “boilerplate code” that gets copied and pasted a hundred times. It can also be confusing – it is often unclear whether properties are applied from an Axes setter (e.g. set_xlabel() and set_xticks()), an XAxis or YAxis setter (e.g. set_major_locator() and set_major_formatter()), a Spine setter (e.g. set_bounds()), or a “bulk” property setter (e.g. tick_params()), or whether one must dig into the figure architecture and apply settings to several different objects. It seems like there should be a more unified, straightforward way to change settings without sacrificing the advantages of object-oriented design.

Changes

UltraPlot includes the format() command to resolve this. Think of this as an expanded and thoroughly documented version of the update() command. format() can modify things like axis labels and titles and apply new “rc” settings to existing axes. It also integrates with various constructor functions to help keep things succinct. Further, the format() and format() commands can be used to format() several subplots at once.

Together, these features significantly reduce the amount of code needed to create highly customized figures. As an example, it is trivial to see that…

import ultraplot as uplt
fig, axs = uplt.subplots(ncols=2)
axs.format(color='gray', linewidth=1)
axs.format(xlim=(0, 100), xticks=10, xtickminor=True, xlabel='foo', ylabel='bar')

is much more succinct than…

import matplotlib.pyplot as plt
import matplotlib.ticker as mticker
import matplotlib as mpl
with mpl.rc_context(rc={'axes.linewidth': 1, 'axes.edgecolor': 'gray'}):
    fig, axs = plt.subplots(ncols=2, sharey=True)
    axs[0].set_ylabel('bar', color='gray')
    for ax in axs:
        ax.set_xlim(0, 100)
        ax.xaxis.set_major_locator(mticker.MultipleLocator(10))
        ax.tick_params(width=1, color='gray', labelcolor='gray')
        ax.tick_params(axis='x', which='minor', bottom=True)
        ax.set_xlabel('foo', color='gray')

Class constructor functions

Limitation

Matplotlib and cartopy define several classes with verbose names like MultipleLocator, FormatStrFormatter, and LambertAzimuthalEqualArea. They also keep them out of the top-level package namespace. Since plotting code has a half life of about 30 seconds, typing out these extra class names and import statements can be frustrating.

Parts of matplotlib’s interface were designed with this in mind. Backend classes, native axes projections, axis scales, colormaps, box styles, arrow styles, and arc styles are referenced with “registered” string names, as are basemap projections. So, why not “register” everything else?

Changes

In UltraPlot, tick locators, tick formatters, axis scales, property cycles, colormaps, normalizers, and cartopy projections are all “registered”. This is accomplished by defining “constructor functions” and passing various keyword arguments through these functions.

The constructor functions also accept intuitive inputs alongside “registered” names. For example, a scalar passed to Locator returns a MultipleLocator, a lists of strings passed to Formatter returns a FixedFormatter, and Cycle and Colormap accept colormap names, individual colors, and lists of colors. Passing the relevant class instance to a constructor function simply returns it, and all the registered classes are available in the top-level namespace – so class instances can be directly created with e.g. uplt.MultipleLocator(...) or uplt.LogNorm(...) rather than relying on constructor functions.

The below table lists the constructor functions and the keyword arguments that use them.

Links

  • For more on axes projections, see this page.

  • For more on axis locators, see this page.

  • For more on axis formatters, see this page.

  • For more on axis scales, see this page.

  • For more on datetime locators and formatters, see this page.

  • For more on colormaps, see this page.

  • For more on normalizers, see this page.

  • For more on color cycles, see this page.

Automatic dimensions and spacing

Limitation

Matplotlib plots tend to require “tweaking” when you have more than one subplot in the figure. This is partly because you must specify the physical dimensions of the figure, despite the fact that…

  1. The subplot aspect ratio is generally more relevant than the figure aspect ratio. A default aspect ratio of 1 is desirable for most plots, and the aspect ratio must be held fixed for geographic and polar projections and most imshow() plots.

  2. The subplot width and height control the “apparent” size of lines, markers, text, and other plotted content. If the figure size is fixed, adding more subplots will decrease the average subplot size and increase the “apparent” sizes. If the subplot size is fixed instead, this can be avoided.

Matplotlib also includes “tight layout” and “constrained layout” algorithms that can help users avoid having to tweak GridSpec spacing parameters like left, bottom, and wspace. However, these algorithms are disabled by default and somewhat cumbersome to configure. They also cannot apply different amounts of spacing between different subplot row and column boundaries.

Changes

By default, UltraPlot fixes the physical dimensions of a reference subplot rather than the figure. The reference subplot dimensions are controlled with the refwidth, refheight, and refaspect Figure keywords, with a default behavior of refaspect=1 and refwidth=2.5 (inches). If the data aspect ratio of the reference subplot is fixed (as with geographic, polar, imshow(), and heatmap() plots) then this is used instead of refaspect.

Alternatively, you can independently specify the width or height of the figure with the figwidth and figheight parameters. If only one is specified, the other is adjusted to preserve subplot aspect ratios. This is very often useful when preparing figures for submission to a publication. To request figure dimensions suitable for submission to a specific publication, use the journal keyword.

By default, UltraPlot also uses its own tight layout algorithm – preventing text labels from overlapping with subplots. This algorithm works with the GridSpec subclass rather than GridSpec, which provides the following advantages:

  • The GridSpec subclass interprets spacing parameters with font size-relative units rather than figure size-relative units. This is more consistent with the tight layout pad arguments (which, like matplotlib, are specified in font size-relative units) and obviates the need to adjust spaces when the figure size or font size changes.

  • The GridSpec subclass permits variable spacing between rows and columns, and the tight layout algorithm takes this into account. Variable spacing is critical for making outer colorbars and legends and axes panels without “stealing space” from the parent subplot – these objects usually need to be spaced closer to their parents than other subplots.

  • You can override particular spacing parameters and leave the tight layout algorithm to adjust the unspecified spacing parameters. For example, passing right=1 to add_subplots() fixes the right margin at 1 font size-width while the others are adjusted automatically.

  • Only one GridSpec is permitted per figure, considerably simplifying the tight layout algorithm calculations. This restriction is enforced by requiring successive add_subplot() calls to imply the same geometry and include only subplot specs generated from the same GridSpec.

Links

Working with multiple subplots

Limitation

When working with multiple subplots in matplotlib, the path of least resistance often leads to redundant figure elements. Namely…

  • Repeated axis tick labels.

  • Repeated axis labels.

  • Repeated colorbars.

  • Repeated legends.

These sorts of redundancies are very common even in publications, where they waste valuable page space. It is also generally necessary to add “a-b-c” labels to figures with multiple subplots before submitting them to publications, but matplotlib has no built-in way of doing this.

Changes

UltraPlot makes it easier to work with multiple subplots and create clear, concise figures.

  • Axis tick labels and axis labels are automatically shared and aligned between subplot in the same GridSpec row or column. This is controlled by the sharex, sharey, spanx, spany, alignx, and aligny figure keywords.

  • The figure colorbar() and legend`() commands can easily draw colorbars and legends intended to reference more than one subplot in arbitrary contiguous rows and columns. See the next section for details.

  • A-b-c labels can be added to subplots simply using the rc.abc setting – for example, uplt.rc['abc'] = 'A.' or axs.format(abc='A.'). This is possible because add_subplot() assigns a unique number() to every new subplot.

  • The format() command can easily format multiple subplots at once or add colorbars, legends, panels, twin axes, or inset axes to multiple subplots at once. A SubplotGrid is returned by subplots(), and can be indexed like a list or a 2D array.

  • The panel_axes() (shorthand panel()) commands draw thin panels along the edges of subplots. This can be useful for plotting 1D summary statistics alongside 2D plots. You can also add twin axes and panel axes to several subplots at once using SubplotGrid commands.

Links

Simpler colorbars and legends

Limitation

In matplotlib, it can be difficult to draw legend()s along the outside of subplots. Generally, you need to position the legend manually and tweak the spacing to make room for the legend.

Also, colorbar()s drawn along the outside of subplots with e.g. fig.colorbar(..., ax=ax) need to “steal” space from the parent subplot. This can cause asymmetry in figures with more than one subplot. It is also generally difficult to draw “inset” colorbars in matplotlib and to generate outer colorbars with consistent widths (i.e., not too “skinny” or “fat”).

Changes

UltraPlot includes a simple framework for drawing colorbars and legends that reference individual subplots and multiple contiguous subplots.

  • To draw a colorbar or legend on the outside of a specific subplot, pass an “outer” location (e.g. loc='l' or loc='left') to colorbar() or legend().

  • To draw a colorbar or legend on the inside of a specific subplot, pass an “inner” location (e.g. loc='ur' or loc='upper right') to colorbar() or legend().

  • To draw a colorbar or legend along the edge of the figure, use colorbar() and legend. The col, row, and span keywords control which GridSpec rows and columns are spanned by the colorbar or legend.

Since GridSpec permits variable spacing between subplot rows and columns, “outer” colorbars and legends do not alter subplot spacing or add whitespace. This is critical e.g. if you have a colorbar between columns 1 and 2 but nothing between columns 2 and 3. Also, Figure and Axes colorbar widths are now specified in physical units rather than relative units, which makes colorbar thickness independent of subplot size and easier to get just right.

Links

  • For more on single-subplot colorbars and legends, see this page.

  • For more on multi-subplot colorbars and legends, see this page.

  • For new colorbar features, see this page.

  • For new legend features, see this page.

Improved plotting commands

Limitation

A few common plotting tasks take a lot of work using matplotlib alone. The seaborn, xarray, and pandas packages offer improvements, but it would be nice to have this functionality built right into matplotlib’s interface.

Changes

UltraPlot uses the PlotAxes subclass to add various seaborn, xarray, and pandas features to existing matplotlib plotting commands along with several additional features designed to make things easier.

The following features are relevant for “1D” PlotAxes commands like line() (equivalent to plot()) and scatter():

The following features are relevant for “2D” PlotAxes commands like pcolor() and contour():

  • The treatment of data arguments passed to the 2D PlotAxes commands is standardized. This makes them more flexible and arguably more intuitive to use than their matplotlib counterparts.

  • The cmap and norm keyword arguments are interpreted by the Colormap and Norm constructor functions. This permits succinct and flexible colormap and normalizer application.

  • The colorbar keyword draws on-the-fly colorbars using the result of the plotting command. Note that “inset” colorbars can also be drawn, analogous to “inset” legends.

  • The contour(), contourf(), pcolormesh(), and pcolor() commands all accept a labels keyword. This draws contour and grid box labels on-the-fly. Labels are automatically colored black or white according to the luminance of the underlying grid box or filled contour.

  • The default vmin and vmax used to normalize colormaps now excludes data outside the x and y axis bounds xlim and ylim if they were explicitly fixed. This can be disabled by setting rc['cmap.inbounds'] to False or by passing inbounds=False to PlotAxes commands.

  • The DiscreteNorm normalizer is paired with most colormaps by default. It can easily divide colormaps into distinct levels, similar to contour plots. This can be disabled by setting rc['cmap.discrete'] to False or by passing discrete=False to PlotAxes commands.

  • The DivergingNorm normalizer is perfect for data with a natural midpoint and offers both “fair” and “unfair” scaling. The SegmentedNorm normalizer can generate uneven color gradations useful for unusual data distributions.

  • The heatmap() command invokes pcolormesh() then applies an equal axes apect ratio, adds ticks to the center of each gridbox, and disables minor ticks and gridlines. This can be convenient for things like covariance matrices.

  • Coordinate centers passed to commands like pcolor() are automatically translated to “edges”, and coordinate edges passed to commands like contour() are automatically translated to “centers”. In matplotlib, pcolor simply truncates and offsets the data when it receives centers.

  • Commands like pcolor(), contourf() and colorbar() automatically fix an irritating issue where saved vector graphics appear to have thin white lines between filled contours, grid boxes, and colorbar segments. This can be disabled by passing edgefix=False to PlotAxes commands.

Links

  • For the 1D plotting features, see this page.

  • For the 2D plotting features, see this page.

  • For treatment of 1D data arguments, see this page.

  • For treatment of 2D data arguments, see this page.

Cartopy and basemap integration

Limitation

There are two widely-used engines for working with geographic data in matplotlib: cartopy and basemap. Using cartopy tends to be verbose and involve boilerplate code, while using basemap requires plotting with a separate Basemap object rather than the Axes. They both require separate import statements and extra lines of code to configure the projection.

Furthermore, when you use cartopy and basemap plotting commands, “map projection” coordinates are the default coordinate system rather than longitude-latitude coordinates. This choice is confusing for many users, since the vast majority of geophysical data are stored with longitude-latitude (i.e., “Plate Carrée”) coordinates.

Changes

UltraPlot can succinctly create detailed geographic plots using either cartopy or basemap as “backends”. By default, cartopy is used, but basemap can be used by passing backend='basemap' to axes-creation commands or by setting rc['geo.backend'] to 'basemap'. To create a geographic plot, simply pass the PROJ name to an axes-creation command, e.g. fig, ax = uplt.subplots(proj='pcarree') or fig.add_subplot(proj='pcarree'). Alternatively, use the Proj constructor function to quickly generate a Projection or Basemap instance.

Requesting geographic projections creates a GeoAxes with unified support for cartopy and basemap features via the format() command. This lets you quickly modify geographic plot features like latitude and longitude gridlines, gridline labels, continents, coastlines, and political boundaries. The syntax is conveniently analogous to the syntax used for format() and format().

The GeoAxes subclass also makes longitude-latitude coordinates the “default” coordinate system by passing transform=ccrs.PlateCarree() or latlon=True to PlotAxes commands (depending on whether cartopy or basemap is the backend). And to enforce global coverage over the poles and across longitude seams, you can pass globe=True to 2D PlotAxes commands like contour() and pcolormesh().

Links

Pandas and xarray integration

Limitation

Scientific data is commonly stored in array-like containers that include metadata – namely, DataArrays, DataFrames, and Series. When matplotlib receives these objects, it ignores the associated metadata. To create plots that are labeled with the metadata, you must use the plot(), plot(), and plot() commands instead.

This approach is fine for quick plots, but not ideal for complex ones. It requires learning a different syntax from matplotlib, and tends to encourage using the pyplot interface rather than the object-oriented interface. The plot commands also include features that would be useful additions to matplotlib in their own right, without requiring special containers and a separate interface.

Changes

UltraPlot reproduces many of the plot(), plot(), and plot() features directly on the PlotAxes commands. This includes grouped or stacked bar plots and layered or stacked area plots from two-dimensional input data, auto-detection of diverging datasets for application of diverging colormaps and normalizers, and on-the-fly colorbars and legends using colorbar and legend keywords.

UltraPlot also handles metadata associated with DataArray, DataFrame, Series, and Quantity objects. When a plotting command receives these objects, it updates the axis tick labels, axis labels, subplot title, and colorbar and legend labels from the metadata. For Quantity arrays (including Quantity those stored inside DataArray containers), a unit string is generated from the pint.Unit according to the rc.unitformat setting (note UltraPlot also automatically calls setup_matplotlib() whenever a Quantity is used for x and y coordinates and removes the units from z coordinates to avoid the stripped-units warning message). These features can be disabled by setting rc.autoformat to False or passing autoformat=False to any plotting command.

Links

Aesthetic colors and fonts

Limitation

A common problem with scientific visualizations is the use of “misleading” colormaps like 'jet'. These colormaps have jarring jumps in hue, saturation, and luminance that can trick the human eye into seeing non-existing patterns. It is important to use “perceptually uniform” colormaps instead. Matplotlib comes packaged with a few of its own, plus the ColorBrewer colormap series, but external projects offer a larger variety of aesthetically pleasing “perceptually uniform” colormaps that would be nice to have in one place.

Matplotlib also “registers” the X11/CSS4 color names, but these are relatively limited. The more numerous and arguably more intuitive XKCD color survey names can only be accessed with the 'xkcd:' prefix. As with colormaps, there are also external projects with useful color names like open color.

Finally, matplotlib comes packaged with DejaVu Sans as the default font. This font is open source and include glyphs for a huge variety of characters. However in our opinion, it is not very aesthetically pleasing. It is also difficult to switch to other fonts on limited systems or systems with fonts stored in incompatible file formats (see below).

Changes

UltraPlot adds new colormaps, colors, and fonts to help you make more aesthetically pleasing figures.

Links

  • For more on colormaps, see this page.

  • For more on color cycles, see this page.

  • For more on fonts, see this page.

  • For importing custom colormaps, colors, and fonts, see this page.

Manipulating colormaps

Limitation

In matplotlib, colormaps are implemented with the LinearSegmentedColormap class (representing “smooth” color gradations) and the ListedColormap class (representing “categorical” color sets). They are somewhat cumbersome to modify or create from scratch. Meanwhile, property cycles used for individual plot elements are implemented with the Cycler class. They are easier to modify but they cannot be “registered” by name like colormaps.

The seaborn package includes “color palettes” to make working with colormaps and property cycles easier, but it would be nice to have similar features integrated more closely with matplotlib’s colormap and property cycle constructs.

Changes

UltraPlot tries to make it easy to manipulate colormaps and property cycles.

UltraPlot also makes all colormap and color cycle names case-insensitive, and colormaps are automatically reversed or cyclically shifted 180 degrees if you append '_r' or '_s' to any colormap name. These features are powered by ColormapDatabase, which replaces matplotlib’s native colormap database.

Links

  • For making new colormaps, see this page.

  • For making new color cycles, see this page.

  • For merging colormaps and cycles, see this page.

  • For modifying colormaps and cycles, see this page.

Physical units engine

Limitation

Matplotlib uses figure-relative units for the margins left, right, bottom, and top, and axes-relative units for the column and row spacing wspace and hspace. Relative units tend to require “tinkering” with numbers until you find the right one. And since they are relative, if you decide to change your figure size or add a subplot, they will have to be readjusted.

Matplotlib also requires users to set the figure size figsize in inches. This may be confusing for users outside of the United States.

Changes

UltraPlot uses physical units for the GridSpec keywords left, right, top, bottom, wspace, hspace, pad, outerpad, and innerpad. The default unit (assumed when a numeric argument is passed) is em-widths. Em-widths are particularly appropriate for this context, as plot text can be a useful “ruler” when figuring out the amount of space you need. UltraPlot also permits arbitrary string units for these keywords, for the Figure keywords figsize, figwidth, figheight, refwidth, and refheight, and in a few other places. This is powered by the physical units engine units(). Acceptable units include inches, centimeters, millimeters, pixels, points, and picas (a table of acceptable units is found here). Note the units() engine also translates rc settings assigned to rc_matplotlib() and rc_UltraPlot, e.g. rc['subplots.refwidth'], rc['legend.columnspacing'], and rc['axes.labelpad'].

Links

Flexible global settings

Limitation

In matplotlib, there are several rcParams that would be useful to set all at once, like spine and label colors. It might also be useful to change these settings for individual subplots rather than globally.

Changes

In UltraPlot, you can use the rc object to change both native matplotlib settings (found in rc_matplotlib) and added UltraPlot settings (found in rc_UltraPlot). Assigned settings are always validated, and “meta” settings like meta.edgecolor, meta.linewidth, and font.smallsize can be used to update many settings all at once. Settings can be changed with uplt.rc.key = value, uplt.rc[key] = value, uplt.rc.update(key=value), using format(), or using context(). Settings that have changed during the python session can be saved to a file with save() (see changed()), and settings can be loaded from files with load().

Links

Loading stuff

Limitation

Matplotlib rcParams can be changed persistently by placing ref:matplotlibrc files in the same directory as your python script. But it can be difficult to design and store your own colormaps and color cycles for future use. It is also difficult to get matplotlib to use custom .ttf and .otf font files, which may be desirable when you are working on Linux servers with limited font selections.

Changes

UltraPlot settings can be changed persistently by editing the default ultraplotrc file in the location given by user_file() (this is usually $HOME/.ultraplot/ultraplotrc) or by adding loose ultraplotrc files to either the current directory or an arbitrary parent directory. Adding files to parent directories can be useful when working in projects with lots of subfolders.

UltraPlot also automatically registers colormaps, color cycles, colors, and font files stored in subfolders named cmaps, cycles, colors, and fonts in the location given by user_folder() (this is usually $HOME/.ultraplot), as well as loose subfolders named ultraplot_cmaps, ultraplot_cycles, ultraplot_colors, and ultraplot_fonts in the current directory or an arbitrary parent directory. You can save colormaps and color cycles to user_folder() simply by passing save=True to Colormap and Cycle. To re-register these files during an active python session, or to register arbitrary input arguments, you can use register_cmaps(), register_cycles(), register_colors(), or register_fonts().

Links