What's new in Matplotlib 3.4.0

For a list of all of the issues and pull requests since the last revision, see the GitHub Stats.

Table of Contents

Figure and Axes creation / management

New subfigure functionality

New figure.Figure.add_subfigure and figure.Figure.subfigures functionalities allow creating virtual figures within figures. Similar nesting was previously done with nested gridspecs (see Nested Gridspecs). However, this did not allow localized figure artists (e.g., a colorbar or suptitle) that only pertained to each subgridspec.

The new methods figure.Figure.add_subfigure and figure.Figure.subfigures are meant to rhyme with figure.Figure.add_subplot and figure.Figure.subplots and have most of the same arguments.

See Figure subfigures for further details.


The subfigure functionality is experimental API as of v3.4.

(Source code, png, pdf)


Single-line string notation for subplot_mosaic

Figure.subplot_mosaic and pyplot.subplot_mosaic now accept a single-line string, using semicolons to delimit rows. Namely,


may be written as the shorter:


(Source code, png, pdf)


Changes to behavior of Axes creation methods (gca, add_axes, add_subplot)

The behavior of the functions to create new Axes (pyplot.axes, pyplot.subplot, figure.Figure.add_axes, figure.Figure.add_subplot) has changed. In the past, these functions would detect if you were attempting to create Axes with the same keyword arguments as already-existing Axes in the current Figure, and if so, they would return the existing Axes. Now, pyplot.axes, figure.Figure.add_axes, and figure.Figure.add_subplot will always create new Axes. pyplot.subplot will continue to reuse an existing Axes with a matching subplot spec and equal kwargs.

Correspondingly, the behavior of the functions to get the current Axes (pyplot.gca, figure.Figure.gca) has changed. In the past, these functions accepted keyword arguments. If the keyword arguments matched an already-existing Axes, then that Axes would be returned, otherwise new Axes would be created with those keyword arguments. Now, the keyword arguments are only considered if there are no Axes at all in the current figure. In a future release, these functions will not accept keyword arguments at all.

add_subplot/add_axes gained an axes_class parameter

In particular, mpl_toolkits Axes subclasses can now be idiomatically used using, e.g., fig.add_subplot(axes_class=mpl_toolkits.axislines.Axes)

Subplot and subplot2grid can now work with constrained layout

constrained_layout depends on a single GridSpec for each logical layout on a figure. Previously, pyplot.subplot and pyplot.subplot2grid added a new GridSpec each time they were called and were therefore incompatible with constrained_layout.

Now subplot attempts to reuse the GridSpec if the number of rows and columns is the same as the top level GridSpec already in the figure, i.e., plt.subplot(2, 1, 2) will use the same GridSpec as plt.subplot(2, 1, 1) and the constrained_layout=True option to Figure will work.

In contrast, mixing nrows and ncols will not work with constrained_layout: plt.subplot(2, 2, 1) followed by plt.subplots(2, 1, 2) will still produce two GridSpecs, and constrained_layout=True will give bad results. In order to get the desired effect, the second call can specify the cells the second Axes is meant to cover: plt.subplots(2, 2, (2, 4)), or the more Pythonic plt.subplot2grid((2, 2), (0, 1), rowspan=2) can be used.

Plotting methods

axline supports transform parameter

axline now supports the transform parameter, which applies to the points xy1, xy2. The slope (if given) is always in data coordinates.

For example, this can be used with ax.transAxes for drawing lines with a fixed slope. In the following plot, the line appears through the same point on both Axes, even though they show different data limits.

fig, axs = plt.subplots(1, 2)

for i, ax in enumerate(axs):
    ax.axline((0.25, 0), slope=2, transform=ax.transAxes)
    ax.set(xlim=(i, i+5), ylim=(i, i+5))

(Source code, png, pdf)


New automatic labeling for bar charts

A new Axes.bar_label method has been added for auto-labeling bar charts.


Example of the new automatic labeling.

A list of hatches can be specified to bar and barh

Similar to some other rectangle properties, it is now possible to hand a list of hatch styles to bar and barh in order to create bars with different hatch styles, e.g.

(Source code, png, pdf)


Setting BarContainer orientation

BarContainer now accepts a new string argument orientation. It can be either 'vertical' or 'horizontal', default is None.

Contour plots now default to using ScalarFormatter

Pass fmt="%1.3f" to the contouring call to restore the old default label format.

Axes.errorbar cycles non-color properties correctly

Formerly, Axes.errorbar incorrectly skipped the Axes property cycle if a color was explicitly specified, even if the property cycler was for other properties (such as line style). Now, Axes.errorbar will advance the Axes property cycle as done for Axes.plot, i.e., as long as all properties in the cycler are not explicitly passed.

For example, the following will cycle through the line styles:

x = np.arange(0.1, 4, 0.5)
y = np.exp(-x)
offsets = [0, 1]

plt.rcParams['axes.prop_cycle'] = plt.cycler('linestyle', ['-', '--'])

fig, ax = plt.subplots()
for offset in offsets:
    ax.errorbar(x, y + offset, xerr=0.1, yerr=0.3, fmt='tab:blue')

(Source code, png, pdf)


errorbar errorevery parameter matches markevery

Similar to the markevery parameter to plot, the errorevery parameter of errorbar now accept slices and NumPy fancy indexes (which must match the size of x).

(Source code, png, pdf)


hexbin supports data reference for C parameter

As with the x and y parameters, Axes.hexbin now supports passing the C parameter using a data reference.

data = {
    'a': np.random.rand(1000),
    'b': np.random.rand(1000),
    'c': np.random.rand(1000),

fig, ax = plt.subplots()
ax.hexbin('a', 'b', C='c', data=data, gridsize=10)

(Source code, png, pdf)


Support callable for formatting of Sankey labels

The format parameter of matplotlib.sankey.Sankey can now accept callables.

This allows the use of an arbitrary function to label flows, for example allowing the mapping of numbers to emoji.

(Source code, png, pdf)


Axes.spines access shortcuts

Axes.spines is now a dedicated container class Spines for a set of Spines instead of an OrderedDict. On top of dict-like access, Axes.spines now also supports some pandas.Series-like features.

Accessing single elements by item or by attribute:


Accessing a subset of items:

ax.spines[['top', 'right']].set_visible(False)

Accessing all items simultaneously:


New stairs method and StepPatch artist

pyplot.stairs and the underlying artist StepPatch provide a cleaner interface for plotting stepwise constant functions for the common case that you know the step edges. This supersedes many use cases of pyplot.step, for instance when plotting the output of numpy.histogram.

For both the artist and the function, the x-like edges input is one element longer than the y-like values input

(Source code, png, pdf)


See Stairs Demo for examples.

Added orientation parameter for stem plots

By default, stem lines are vertical. They can be changed to horizontal using the orientation parameter of Axes.stem or pyplot.stem:

(Source code, png, pdf)


Angles on Bracket arrow styles

Angles specified on the Bracket arrow styles (]-[, ]-, -[, or |-| passed to arrowstyle parameter of FancyArrowPatch) are now applied. Previously, the angleA and angleB options were allowed, but did nothing.

(Source code, png, pdf)


TickedStroke patheffect

The new TickedStroke patheffect can be used to produce lines with a ticked style. This can be used to, e.g., distinguish the valid and invalid sides of the constraint boundaries in the solution space of optimizations.


Colors and colormaps

Collection color specification and mapping

Reworking the handling of color mapping and the keyword arguments for facecolor and edgecolor has resulted in three behavior changes:

  1. Color mapping can be turned off by calling Collection.set_array(None). Previously, this would have no effect.
  2. When a mappable array is set, with facecolor='none' and edgecolor='face', both the faces and the edges are left uncolored. Previously the edges would be color-mapped.
  3. When a mappable array is set, with facecolor='none' and edgecolor='red', the edges are red. This addresses Issue #1302. Previously the edges would be color-mapped.

Transparency (alpha) can be set as an array in collections

Previously, the alpha value controlling transparency in collections could be specified only as a scalar applied to all elements in the collection. For example, all the markers in a scatter plot, or all the quadrilaterals in a pcolormesh plot, would have the same alpha value.

Now it is possible to supply alpha as an array with one value for each element (marker, quadrilateral, etc.) in a collection.

(Source code, png, pdf)


pcolormesh has improved transparency handling by enabling snapping

Due to how the snapping keyword argument was getting passed to the Agg backend, previous versions of Matplotlib would appear to show lines between the grid edges of a mesh with transparency. This version now applies snapping by default. To restore the old behavior (e.g., for test images), you may set rcParams["pcolormesh.snap"] (default: True) to False.

(Source code, png, pdf)


Note that there are lines between the grid boundaries of the main plot which are not the same transparency. The colorbar also shows these lines when a transparency is added to the colormap because internally it uses pcolormesh to draw the colorbar. With snapping on by default (below), the lines at the grid boundaries disappear.

(Source code, png, pdf)


IPython representations for Colormap objects

The matplotlib.colors.Colormap object now has image representations for IPython / Jupyter backends. Cells returning a colormap on the last line will display an image of the colormap.

In[1]: cmap = plt.get_cmap('viridis').with_extremes(bad='r', under='g', over='b')

In[2]: cmap
viridis colormap

Colormap.set_extremes and Colormap.with_extremes

Because the Colormap.set_bad, Colormap.set_under and Colormap.set_over methods modify the colormap in place, the user must be careful to first make a copy of the colormap if setting the extreme colors e.g. for a builtin colormap.

The new Colormap.with_extremes(bad=..., under=..., over=...) can be used to first copy the colormap and set the extreme colors on that copy.

The new Colormap.set_extremes method is provided for API symmetry with Colormap.with_extremes, but note that it suffers from the same issue as the earlier individual setters.

Get under/over/bad colors of Colormap objects

matplotlib.colors.Colormap now has methods get_under, get_over, get_bad for the colors used for out-of-range and masked values.

New cm.unregister_cmap function

cm.unregister_cmap allows users to remove a colormap that they have previously registered.

New CenteredNorm for symmetrical data around a center

In cases where data is symmetrical around a center, for example, positive and negative anomalies around a center zero, CenteredNorm is a new norm that automatically creates a symmetrical mapping around the center. This norm is well suited to be combined with a divergent colormap which uses an unsaturated color in its center.

(Source code, png, pdf)


If the center of symmetry is different from 0, it can be set with the vcenter argument. To manually set the range of CenteredNorm, use the halfrange argument.

See Colormap Normalization for an example and more details about data normalization.

New FuncNorm for arbitrary normalizations

The FuncNorm allows for arbitrary normalization using functions for the forward and inverse.

(Source code, png, pdf)


See Colormap Normalization for an example and more details about data normalization.

GridSpec-based colorbars can now be positioned above or to the left of the main axes

... by passing location="top" or location="left" to the colorbar() call.

Titles, ticks, and labels

supxlabel and supylabel

It is possible to add x- and y-labels to a whole figure, analogous to FigureBase.suptitle using the new FigureBase.supxlabel and FigureBase.supylabel methods.

(Source code, png, pdf)


Shared-axes subplots tick label visibility is now correct for top or left labels

When calling subplots(..., sharex=True, sharey=True), Matplotlib automatically hides x tick labels for Axes not in the first column and y tick labels for Axes not in the last row. This behavior is incorrect if rcParams specify that Axes should be labeled on the top (rcParams["xtick.labeltop"] = True) or on the right (rcParams["ytick.labelright"] = True).

Cases such as the following are now handled correctly (adjusting visibility as needed on the first row and last column of Axes):

plt.rcParams["xtick.labelbottom"] = False
plt.rcParams["xtick.labeltop"] = True
plt.rcParams["ytick.labelleft"] = False
plt.rcParams["ytick.labelright"] = True

fig, axs = plt.subplots(2, 2, sharex=True, sharey=True)

(Source code, png, pdf)


An iterable object with labels can be passed to Axes.plot

When plotting multiple datasets by passing 2D data as y value to plot, labels for the datasets can be passed as a list, the length matching the number of columns in y.

x = [1, 2, 3]

y = [[1, 2],
     [2, 5],
     [4, 9]]

plt.plot(x, y, label=['low', 'high'])

(Source code, png, pdf)


Fonts and Text

Text transform can rotate text direction

The new Text parameter transform_rotates_text now sets whether rotations of the transform affect the text direction.


Example of the new transform_rotates_text parameter

matplotlib.mathtext now supports overset and underset LaTeX symbols

mathtext now supports overset and underset, called as \overset{annotation}{body} or \underset{annotation}{body}, where annotation is the text "above" or "below" the body.

(Source code, png, pdf)


math_fontfamily parameter to change Text font family

The new math_fontfamily parameter may be used to change the family of fonts for each individual text element in a plot. If no parameter is set, the global value rcParams["mathtext.fontset"] (default: 'dejavusans') will be used.


TextArea/AnchoredText support horizontalalignment

The horizontal alignment of text in a TextArea or AnchoredText may now be specified, which is mostly effective for multiline text:

(Source code, png, pdf)


PDF supports URLs on Text artists

URLs on text.Text artists (i.e., from Artist.set_url) will now be saved in PDF files.

rcParams improvements

New rcParams for dates: set converter and whether to use interval_multiples

The new rcParams["date.converter"] (default: None) allows toggling between matplotlib.dates.DateConverter and matplotlib.dates.ConciseDateConverter using the strings 'auto' and 'concise' respectively.

The new rcParams["date.interval_multiples"] (default: None) allows toggling between the dates locator trying to pick ticks at set intervals (i.e., day 1 and 15 of the month), versus evenly spaced ticks that start wherever the timeseries starts:

dates = np.arange('2001-01-10', '2001-05-23', dtype='datetime64[D]')
y = np.sin(dates.astype(float) / 10)
fig, axs = plt.subplots(nrows=2, constrained_layout=True)

plt.rcParams['date.converter'] = 'concise'
plt.rcParams['date.interval_multiples'] = True
axs[0].plot(dates, y)

plt.rcParams['date.converter'] = 'auto'
plt.rcParams['date.interval_multiples'] = False
axs[1].plot(dates, y)

(Source code, png, pdf)


Date formatters now respect usetex rcParam

The AutoDateFormatter and ConciseDateFormatter now respect rcParams["text.usetex"] (default: False), and will thus use fonts consistent with TeX rendering of the default (non-date) formatter. TeX rendering may also be enabled/disabled by passing the usetex parameter when creating the formatter instance.

In the following plot, both the x-axis (dates) and y-axis (numbers) now use the same (TeX) font:

(Source code, png, pdf)


Setting image.cmap to a Colormap

It is now possible to set rcParams["image.cmap"] (default: 'viridis') to a Colormap instance, such as a colormap created with the new set_extremes above. (This can only be done from Python code, not from the matplotlibrc file.)

Tick and tick label colors can be set independently using rcParams

Previously, rcParams["xtick.color"] (default: 'black') defined both the tick color and the label color. The label color can now be set independently using rcParams["xtick.labelcolor"] (default: 'inherit'). It defaults to 'inherit' which will take the value from rcParams["xtick.color"] (default: 'black'). The same holds for ytick.[label]color. For instance, to set the ticks to light grey and the tick labels to black, one can use the following code in a script:

import matplotlib as mpl

mpl.rcParams['xtick.labelcolor'] = 'lightgrey'
mpl.rcParams['xtick.color'] = 'black'
mpl.rcParams['ytick.labelcolor'] = 'lightgrey'
mpl.rcParams['ytick.color'] = 'black'

Or by adding the following lines to the matplotlibrc file, or a Matplotlib style file:

xtick.labelcolor : lightgrey
xtick.color      : black
ytick.labelcolor : lightgrey
ytick.color      : black

3D Axes improvements

Errorbar method in 3D Axes

The errorbar function Axes.errorbar is ported into the 3D Axes framework in its entirety, supporting features such as custom styling for error lines and cap marks, control over errorbar spacing, upper and lower limit marks.


Stem plots in 3D Axes

Stem plots are now supported on 3D Axes. Much like 2D stems, stem3D supports plotting the stems in various orientations:

(Source code, png, pdf)


See also the 3D stem demo.

3D Collection properties are now modifiable

Previously, properties of a 3D Collection that were used for 3D effects (e.g., colors were modified to produce depth shading) could not be changed after it was created.

Now it is possible to modify all properties of 3D Collections at any time.

Panning in 3D Axes

Click and drag with the middle mouse button to pan 3D Axes.

Interactive tool improvements

New RangeSlider widget

widgets.RangeSlider allows for creating a slider that defines a range rather than a single value.

(Source code, png, pdf)


Sliders can now snap to arbitrary values

The Slider UI widget now accepts arrays for valstep. This generalizes the previous behavior by allowing the slider to snap to arbitrary values.

Pausing and Resuming Animations

The animation.Animation.pause and animation.Animation.resume methods allow you to pause and resume animations. These methods can be used as callbacks for event listeners on UI elements so that your plots can have some playback control UI.

Sphinx extensions

plot_directive caption option

Captions were previously supported when using the plot_directive directive with an external source file by specifying content:

.. plot:: path/to/plot.py

    This is the caption for the plot.

The :caption: option allows specifying the caption for both external:

.. plot:: path/to/plot.py
    :caption: This is the caption for the plot.

and inline plots:

.. plot::
    :caption: This is a caption for the plot.

    plt.plot([1, 2, 3])

Backend-specific improvements

Consecutive rasterized draws now merged

Elements of a vector output can be individually set to rasterized, using the rasterized keyword argument, or set_rasterized(). This can be useful to reduce file sizes. For figures with multiple raster elements they are now automatically merged into a smaller number of bitmaps where this will not effect the visual output. For cases with many elements this can result in significantly smaller file sizes.

To ensure this happens do not place vector elements between raster ones.

To inhibit this merging set Figure.suppressComposite to True.

Support raw/rgba frame format in FFMpegFileWriter

When using FFMpegFileWriter, the frame_format may now be set to "raw" or "rgba", which may be slightly faster than an image format, as no encoding/decoding need take place between Matplotlib and FFmpeg.

nbAgg/WebAgg support middle-click and double-click

Double click events are now supported by the nbAgg and WebAgg backends. Formerly, WebAgg would report middle-click events as right clicks, but now reports the correct button type.

nbAgg support binary communication

If the web browser and notebook support binary websockets, nbAgg will now use them for slightly improved transfer of figure display.

Indexed color for PNG images in PDF files when possible

When PNG images have 256 colors or fewer, they are converted to indexed color before saving them in a PDF. This can result in a significant reduction in file size in some cases. This is particularly true for raster data that uses a colormap but no interpolation, such as Healpy mollview plots. Currently, this is only done for RGB images.

Improved font subsettings in PDF/PS

Font subsetting in PDF and PostScript has been re-written from the embedded ttconv C code to Python. Some composite characters and outlines may have changed slightly. This fixes ttc subsetting in PDF, and adds support for subsetting of type 3 OTF fonts, resulting in smaller files (much smaller when using CJK fonts), and avoids running into issues with type 42 embedding and certain PDF readers such as Acrobat Reader.

Kerning added to strings in PDFs

As with text produced in the Agg backend (see the previous what's new entry for examples), PDFs now include kerning in text strings.

Fully-fractional HiDPI in QtAgg

Fully-fractional HiDPI (that is, HiDPI ratios that are not whole integers) was added in Qt 5.14, and is now supported by the QtAgg backend when using this version of Qt or newer.

wxAgg supports fullscreen toggle

The wxAgg backend supports toggling fullscreen using the f shortcut, or the manager function FigureManagerBase.full_screen_toggle.