You are reading an old version of the documentation (v2.2.2). For the latest version see
Version 2.2.2
Fork me on GitHub

Table Of Contents

API Changes

Log of changes to Matplotlib that affect the outward-facing API. If updating Matplotlib breaks your scripts, this list may help you figure out what caused the breakage and how to fix it by updating your code.

For new features that were added to Matplotlib, please see What’s new in Matplotlib.

API Changes in 2.2.0

New dependency

kiwisolver is now a required dependency to support the new constrained_layout, see Constrained Layout Guide for more details.


Classes, functions, and methods

The unused and untested Artist.onRemove and Artist.hitlist methods have been deprecated.

The now unused mlab.less_simple_linear_interpolation function is deprecated.

The unused ContourLabeler.get_real_label_width method is deprecated.

The unused FigureManagerBase.show_popup method is deprecated. This introduced in e945059b327d42a99938b939a1be867fa023e7ba in 2005 but never built out into any of the backends.

backend_tkagg.AxisMenu is deprecated, as it has become unused since the removal of “classic” toolbars.

Changed function signatures

kwarg fig to GridSpec.get_subplot_params is deprecated, use figure instead.

Using pyplot.axes with an Axes as argument is deprecated. This sets the current axes, i.e. it has the same effect as For clarity should be preferred over plt.axes(ax).

Using strings instead of booleans to control grid and tick visibility is deprecated. Using "on", "off", "true", or "false" to control grid and tick visibility has been deprecated. Instead, use normal booleans (True/False) or boolean-likes. In the future, all non-empty strings may be interpreted as True.

When given 2D inputs with non-matching numbers of columns, plot currently cycles through the columns of the narrower input, until all the columns of the wider input have been plotted. This behavior is deprecated; in the future, only broadcasting (1 column to n columns) will be performed.


The rcParams["backend.qt4"] and rcParams["backend.qt5"] rcParams were deprecated in version 2.2. In order to force the use of a specific Qt binding, either import that binding first, or set the QT_API environment variable.

Deprecation of the nbagg.transparent rcParam. To control transparency of figure patches in the nbagg (or any other) backend, directly set figure.patch.facecolor, or the figure.facecolor rcParam.

Deprecated Axis.unit_data

Use Axis.units (which has long existed) instead.


Function Signatures

Contouring no longer supports legacy corner masking. The deprecated ContourSet.vmin and ContourSet.vmax properties have been removed.

Passing None instead of "none" as format to errorbar is no longer supported.

The bgcolor keyword argument to Axes has been removed.

Modules, methods, and functions

The, mpl_toolkits.exceltools and mpl_toolkits.gtktools modules have been removed. remains available at

The mpl_toolkits.mplot3d.art3d.iscolor function has been removed.

The Axes.get_axis_bgcolor, Axes.set_axis_bgcolor, Bbox.update_from_data, Bbox.update_datalim_numerix, MaxNLocator.bin_boundaries methods have been removed.

mencoder can no longer be used to encode animations.

The unused FONT_SCALE and fontd attributes of the RendererSVG class have been removed.

color maps

The spectral colormap has been removed. The Vega* colormaps, which were aliases for the tab* colormaps, have been removed.


The following deprecated rcParams have been removed:

  • axes.color_cycle (see axes.prop_cycle),
  • legend.isaxes,
  • svg.embed_char_paths (see svg.fonttype),
  • text.fontstyle, text.fontangle, text.fontvariant, text.fontweight, text.fontsize (renamed to, etc.),
  • tick.size (renamed to tick.major.size).

Only accept string-like for Categorical input

Do not accept mixed string / float / int input, only strings are valid categoricals.

Removal of unused imports

Many unused imports were removed from the codebase. As a result, trying to import certain classes or functions from the “wrong” module (e.g. Figure from matplotlib.backends.backend_agg instead of matplotlib.figure) will now raise an ImportError.

Axes3D.get_xlim, get_ylim and get_zlim now return a tuple

They previously returned an array. Returning a tuple is consistent with the behavior for 2D axes.

Exception type changes

If MovieWriterRegistry can’t find the requested MovieWriter, a more helpful RuntimeError message is now raised instead of the previously raised KeyError.

auto_adjust_subplotpars now raises ValueError instead of RuntimeError when sizes of input lists don’t match

Figure.set_figwidth and Figure.set_figheight default forward to True

matplotlib.Figure.set_figwidth and matplotlib.Figure.set_figheight had the kwarg forward=False by default, but Figure.set_size_inches now defaults to forward=True. This makes these functions conistent.

Do not truncate svg sizes to nearest point

There is no reason to size the SVG out put in integer points, change to out putting floats for the height, width, and viewBox attributes of the svg element.

Fontsizes less than 1 pt are clipped to be 1 pt.

FreeType doesn’t allow fonts to get smaller than 1 pt, so all Agg backends were silently rounding up to 1 pt. PDF (other vector backends?) were letting us write fonts that were less than 1 pt, but they could not be placed properly because position information comes from FreeType. This change makes it so no backends can use fonts smaller than 1 pt, consistent with FreeType and ensuring more consistent results across backends.

Changes to Qt backend class MRO

To support both Agg and cairo rendering for Qt backends all of the non-Agg specific code previously in backend_qt5agg.FigureCanvasQTAggBase has been moved to backend_qt5.FigureCanvasQT so it can be shared with the cairo implementation. The FigureCanvasQTAggBase.paintEvent(), FigureCanvasQTAggBase.blit(), and FigureCanvasQTAggBase.print_figure() methods have moved to FigureCanvasQTAgg.paintEvent(), FigureCanvasQTAgg.blit(), and FigureCanvasQTAgg.print_figure(). The first two methods assume that the instance is also a QWidget so to use FigureCanvasQTAggBase it was required to multiple inherit from a QWidget sub-class.

Having moved all of its methods either up or down the class hierarchy FigureCanvasQTAggBase has been deprecated. To do this with out warning and to preserve as much API as possible, backend_qt5.FigureCanvasQTAggBase now inherits from backend_qt5.FigureCanvasQTAgg.

The MRO for FigureCanvasQTAgg and FigureCanvasQTAggBase used to be




respectively. They are now




Axes.imshow clips RGB values to the valid range

When Axes.imshow is passed an RGB or RGBA value with out-of-range values, it now logs a warning and clips them to the valid range. The old behaviour, wrapping back in to the range, often hid outliers and made interpreting RGB images unreliable.

GTKAgg and GTKCairo backends deprecated

The GTKAgg and GTKCairo backends have been deprecated. These obsolete backends allow figures to be rendered via the GTK+ 2 toolkit. They are untested, known to be broken, will not work with Python 3, and their use has been discouraged for some time. Instead, use the GTK3Agg and GTK3Cairo backends for rendering to GTK+ 3 windows.

API Changes in 2.1.2

Figure.legend no longer checks for repeated lines to ignore

matplotlib.Figure.legend used to check if a line had the same label as an existing legend entry. If it also had the same line color or marker color legend didn’t add a new entry for that line. However, the list of conditions was incomplete, didn’t handle RGB tuples, didn’t handle linewidths or linestyles etc.

This logic did not exist in Axes.legend. It was included (erroneously) in Matplotlib 2.1.1 when the legend argument parsing was unified [#9324]( This change removes that check in Axes.legend again to restore the old behavior.

This logic has also been dropped from Figure.legend, where it was previously undocumented. Repeated lines with the same label will now each have an entry in the legend. If you do not want the duplicate entries, don’t add a label to the line, or prepend the label with an underscore.

API Changes in 2.1.1

Default behavior of log scales reverted to clip <= 0 values

The change it 2.1.0 to mask in logscale by default had more disruptive changes than anticipated and has been reverted, however the clipping is now done in a way that fixes the issues that motivated changing the default behavior to 'mask'.

As a side effect of this change, error bars which go negative now work as expected on log scales.

API Changes in 2.1.0

Default behavior of log scales changed to mask <= 0 values

Calling matplotlib.axes.Axes.set_xscale or matplotlib.axes.Axes.set_yscale now uses ‘mask’ as the default method to handle invalid values (as opposed to ‘clip’). This means that any values <= 0 on a log scale will not be shown.

Previously they were clipped to a very small number and shown.

matplotlib.cbook.CallbackRegistry.process() suppresses exceptions by default

Matplotlib uses instances of CallbackRegistry as a bridge between user input event from the GUI and user callbacks. Previously, any exceptions raised in a user call back would bubble out of of the process method, which is typically in the GUI event loop. Most GUI frameworks simple print the traceback to the screen and continue as there is not always a clear method of getting the exception back to the user. However PyQt5 now exits the process when it receives an un-handled python exception in the event loop. Thus, process() now suppresses and prints tracebacks to stderr by default.

What process() does with exceptions is now user configurable via the exception_handler attribute and kwarg. To restore the previous behavior pass None

cb = CallbackRegistry(exception_handler=None)

A function which take and Exception as its only argument may also be passed

def maybe_reraise(exc):
    if isinstance(exc, RuntimeError):
        raise exc

cb = CallbackRegistry(exception_handler=maybe_reraise)

Improved toggling of the axes grids

The g key binding now switches the states of the x and y grids independently (by cycling through all four on/off combinations).

The new G key binding switches the states of the minor grids.

Both bindings are disabled if only a subset of the grid lines (in either direction) is visible, to avoid making irreversible changes to the figure.

Removal of warning on empty legends

plt.legend used to issue a warning when no labeled artist could be found. This warning has been removed.

More accurate legend autopositioning

Automatic positioning of legends now prefers using the area surrounded by a Line2D rather than placing the legend over the line itself.

Cleanup of stock sample data

The sample data of stocks has been cleaned up to remove redundancies and increase portability. The AAPL.dat.gz, INTC.dat.gz and aapl.csv files have been removed entirely and will also no longer be available from matplotlib.cbook.get_sample_data. If a CSV file is required, we suggest using the msft.csv that continues to be shipped in the sample data. If a NumPy binary file is acceptable, we suggest using one of the following two new files. The aapl.npy.gz and goog.npy files have been replaced by aapl.npz and goog.npz, wherein the first column’s type has changed from to np.datetime64 for better portability across Python versions. Note that Matplotlib does not fully support np.datetime64 as yet.

Updated qhull to 2015.2

The version of qhull shipped with Matplotlib, which is used for Delaunay triangulation, has been updated from version 2012.1 to 2015.2.

Improved Delaunay triangulations with large offsets

Delaunay triangulations now deal with large x,y offsets in a better way. This can cause minor changes to any triangulations calculated using Matplotlib, i.e. any use of matplotlib.tri.Triangulation that requests that a Delaunay triangulation is calculated, which includes matplotlib.pyplot.tricontour, matplotlib.pyplot.tricontourf, matplotlib.pyplot.tripcolor, matplotlib.pyplot.triplot, matplotlib.mlab.griddata and mpl_toolkits.mplot3d.axes3d.Axes3D.plot_trisurf.

Use backports.functools_lru_cache instead of functools32

It’s better maintained and more widely used (by pylint, jaraco, etc).

cbook.is_numlike only performs an instance check

is_numlike() now only checks that its argument is an instance of (numbers.Number, np.Number). In particular, this means that arrays are now not num-like.

Elliptical arcs now drawn between correct angles

The matplotlib.patches.Arc patch is now correctly drawn between the given angles.

Previously a circular arc was drawn and then stretched into an ellipse, so the resulting arc did not lie between theta1 and theta2.

-d$backend no longer sets the backend

It is no longer possible to set the backend by passing -d$backend at the command line. Use the MPLBACKEND environment variable instead.

Path.intersects_bbox always treats the bounding box as filled

Previously, when Path.intersects_bbox was called with filled set to False, it would treat both the path and the bounding box as unfilled. This behavior was not well documented and it is usually not the desired behavior, since bounding boxes are used to represent more complex shapes located inside the bounding box. This behavior has now been changed: when filled is False, the path will be treated as unfilled, but the bounding box is still treated as filled. The old behavior was arguably an implementation bug.

When Path.intersects_bbox is called with filled set to True (the default value), there is no change in behavior. For those rare cases where Path.intersects_bbox was called with filled set to False and where the old behavior is actually desired, the suggested workaround is to call Path.intersects_path with a rectangle as the path:

from matplotlib.path import Path
from matplotlib.transforms import Bbox, BboxTransformTo
rect = Path.unit_rectangle().transformed(BboxTransformTo(bbox))
result = path.intersects_path(rect, filled=False)

WX no longer calls generates IdleEvent events or calls idle_event

Removed unused private method _onIdle from FigureCanvasWx.

The IdleEvent class and FigureCanvasBase.idle_event method will be removed in 2.2

Correct scaling of magnitude_spectrum()

The functions matplotlib.mlab.magnitude_spectrum() and matplotlib.pyplot.magnitude_spectrum() implicitly assumed the sum of windowing function values to be one. In Matplotlib and Numpy the standard windowing functions are scaled to have maximum value of one, which usually results in a sum of the order of n/2 for a n-point signal. Thus the amplitude scaling magnitude_spectrum() was off by that amount when using standard windowing functions (Bug 8417 ). Now the behavior is consistent with matplotlib.pyplot.psd() and scipy.signal.welch(). The following example demonstrates the new and old scaling:

import matplotlib.pyplot as plt
import numpy as np

tau, n = 10, 1024  # 10 second signal with 1024 points
T = tau/n  # sampling interval
t = np.arange(n)*T

a = 4  # amplitude
x = a*np.sin(40*np.pi*t)  # 20 Hz sine with amplitude a

# New correct behavior: Amplitude at 20 Hz is a/2
plt.magnitude_spectrum(x, Fs=1/T, sides='onesided', scale='linear')

# Original behavior: Amplitude at 20 Hz is (a/2)*(n/2) for a Hanning window
w = np.hanning(n)  # default window is a Hanning window
plt.magnitude_spectrum(x*np.sum(w), Fs=1/T, sides='onesided', scale='linear')

Change to signatures of bar() & barh()

For 2.0 the default value of *align* changed to 'center'. However this caused the signature of bar() and barh() to be misleading as the first parameters were still left and bottom respectively:

bar(left, height, *, align='center', **kwargs)
barh(bottom, width, *, align='center', **kwargs)

despite behaving as the center in both cases. The methods now take *args, **kwargs as input and are documented to have the primary signatures of:

bar(x, height, *, align='center', **kwargs)
barh(y, width, *, align='center', **kwargs)

Passing left and bottom as keyword arguments to bar() and barh() respectively will warn. Support will be removed in Matplotlib 3.0.

Font cache as json

The font cache is now saved as json, rather than a pickle.

Invalid (Non-finite) Axis Limit Error

When using set_xlim() and set_ylim(), passing non-finite values now results in a ValueError. The previous behavior resulted in the limits being erroneously reset to (-0.001, 0.001).

scatter and Collection offsets are no longer implicitly flattened

Collection (and thus both 2D scatter and 3D scatter) no longer implicitly flattens its offsets. As a consequence, scatter’s x and y arguments can no longer be 2+-dimensional arrays.


GraphicsContextBase’s linestyle property.

The GraphicsContextBase.get_linestyle and GraphicsContextBase.set_linestyle methods, which had no effect, have been deprecated. All of the backends Matplotlib ships use GraphicsContextBase.get_dashes and GraphicsContextBase.set_dashes which are more general. Third-party backends should also migrate to the *_dashes methods.


matplotlib.testing.noseclasses is deprecated and will be removed in 2.3

EngFormatter num arg as string

Passing a string as num argument when calling an instance of matplotlib.ticker.EngFormatter is deprecated and will be removed in 2.3.

mpl_toolkits.axes_grid module

All functionally from mpl_toolkits.axes_grid can be found in either mpl_toolkits.axes_grid1 or mpl_toolkits.axisartist. Axes classes from mpl_toolkits.axes_grid based on Axis from mpl_toolkits.axisartist can be found in mpl_toolkits.axisartist.

Axes collision in Figure.add_axes

Adding an axes instance to a figure by using the same arguments as for a previous axes instance currently reuses the earlier instance. This behavior has been deprecated in Matplotlib 2.1. In a future version, a new instance will always be created and returned. Meanwhile, in such a situation, a deprecation warning is raised by AxesStack.

This warning can be suppressed, and the future behavior ensured, by passing a unique label to each axes instance. See the docstring of add_axes() for more information.

Additional details on the rationale behind this deprecation can be found in #7377 and #9024.

Former validators for contour.negative_linestyle

The former public validation functions validate_negative_linestyle and validate_negative_linestyle_legacy will be deprecated in 2.1 and may be removed in 2.3. There are no public functions to replace them.


Many unused or near-unused matplotlib.cbook functions and classes have been deprecated: converter, tostr, todatetime, todate, tofloat, toint, unique, is_string_like, is_sequence_of_strings, is_scalar, Sorter, Xlator, soundex, Null, dict_delall, RingBuffer, get_split_ind, wrap, get_recursive_filelist, pieces, exception_to_str, allequal, alltrue, onetrue, allpairs, finddir, reverse_dict, restrict_dict, issubclass_safe, recursive_remove, unmasked_index_ranges.

Code Removal

Moved to Renamed because it now handles Qt5 as well.

Previously Deprecated methods

The GraphicsContextBase.set_graylevel, FigureCanvasBase.onHilite and mpl_toolkits.axes_grid1.mpl_axes.Axes.toggle_axisline methods have been removed.

The ArtistInspector.findobj method, which was never working due to the lack of a get_children method, has been removed.

The deprecated point_in_path, get_path_extents, point_in_path_collection, path_intersects_path, convert_path_to_polygons, cleanup_path and clip_path_to_rect functions in the matplotlib.path module have been removed. Their functionality remains exposed as methods on the Path class.

The deprecated Artist.get_axes and Artist.set_axes methods have been removed

The matplotlib.backends.backend_ps.seq_allequal function has been removed. Use np.array_equal instead.

The deprecated matplotlib.rcsetup.validate_maskedarray, matplotlib.rcsetup.deprecate_savefig_extension and matplotlib.rcsetup.validate_tkpythoninspect functions, and associated savefig.extension and tk.pythoninspect rcparams entries have been removed.

The kwarg resolution of matplotlib.projections.polar.PolarAxes has been removed. It has deprecation with no effect from version 0.98.x.


Support for setting an Axes’s aspect to "normal" has been removed, in favor of the synonym "auto".

shading kwarg to pcolor

The shading kwarg to pcolor has been removed. Set edgecolors appropriately instead.

Functions removed from the lines module

The matplotlib.lines module no longer imports the pts_to_prestep, pts_to_midstep and pts_to_poststep functions from matplotlib.cbook.

PDF backend functions

The methods embedTeXFont and tex_font_mapping of matplotlib.backqend_pdf.PdfFile have been removed. It is unlikely that external users would have called these methods, which are related to the font system internal to the PDF backend.


Remove the delaunay triangulation code which is now handled by Qhull via matplotlib.tri.

API Changes in 2.0.1

Extensions to matplotlib.backend_bases.GraphicsContextBase

To better support controlling the color of hatches, the method matplotlib.backend_bases.GraphicsContextBase.set_hatch_color was added to the expected API of GraphicsContext classes. Calls to this method are currently wrapped with a try:...except Attribute: block to preserve back-compatibility with any third-party backends which do not extend GraphicsContextBase.

This value can be accessed in the backends via matplotlib.backend_bases.GraphicsContextBase.get_hatch_color (which was added in 2.0 see Extension to matplotlib.backend_bases.GraphicsContextBase) and should be used to color the hatches.

In the future there may also be hatch_linewidth and hatch_density related methods added. It is encouraged, but not required that third-party backends extend GraphicsContextBase to make adapting to these changes easier.

afm.get_fontconfig_fonts returns a list of paths and does not check for existence

afm.get_fontconfig_fonts used to return a set of paths encoded as a {key: 1, ...} dict, and checked for the existence of the paths. It now returns a list and dropped the existence check, as the same check is performed by the caller (afm.findSystemFonts) as well.

bar now returns rectangles of negative height or width if the corresponding input is negative used to normalize the coordinates of the rectangles that it created, to keep their height and width positives, even if the corresponding input was negative. This normalization has been removed to permit a simpler computation of the correct sticky_edges to use.

Do not clip line width when scaling dashes

The algorithm to scale dashes was changed to no longer clip the scaling factor: the dash patterns now continue to shrink at thin line widths. If the line width is smaller than the effective pixel size, this may result in dashed lines turning into solid gray-ish lines. This also required slightly tweaking the default patterns for ‘–’, ‘:’, and ‘.-‘ so that with the default line width the final patterns would not change.

There is no way to restore the old behavior.

Deprecate ‘Vega’ color maps

The “Vega” colormaps are deprecated in Matplotlib 2.0.1 and will be removed in Matplotlib 2.2. Use the “tab” colormaps instead: “tab10”, “tab20”, “tab20b”, “tab20c”.

API Changes in 2.0.0

Deprecation and removal

Color of Axes

The axisbg and axis_bgcolor properties on Axes have been deprecated in favor of facecolor.

GTK and GDK backends deprecated

The GDK and GTK backends have been deprecated. These obsolete backends allow figures to be rendered via the GDK API to files and GTK2 figures. They are untested and known to be broken, and their use has been discouraged for some time. Instead, use the GTKAgg and GTKCairo backends for rendering to GTK2 windows.

WX backend deprecated

The WX backend has been deprecated. It is untested, and its use has been discouraged for some time. Instead, use the WXAgg backend for rendering figures to WX windows.

CocoaAgg backend removed

The deprecated and not fully functional CocoaAgg backend has been removed.

round removed from TkAgg Backend

The TkAgg backend had its own implementation of the round function. This was unused internally and has been removed. Instead, use either the round builtin function or numpy.round.

‘hold’ functionality deprecated

The ‘hold’ keyword argument and all functions and methods related to it are deprecated, along with the ‘axes.hold’ rcParams entry. The behavior will remain consistent with the default hold=True state that has long been in place. Instead of using a function or keyword argument (hold=False) to change that behavior, explicitly clear the axes or figure as needed prior to subsequent plotting commands.

Artist.update has return value

The methods matplotlib.artist.Artist.set, matplotlib.Artist.update, and the function matplotlib.artist.setp now use a common codepath to look up how to update the given artist properties (either using the setter methods or an attribute/property).

The behavior of matplotlib.Artist.update is slightly changed to return a list of the values returned from the setter methods to avoid changing the API of matplotlib.Artist.set and matplotlib.artist.setp.

The keys passed into matplotlib.Artist.update are now converted to lower case before being processed, to match the behavior of matplotlib.Artist.set and matplotlib.artist.setp. This should not break any user code because there are no set methods with capitals in their names, but this puts a constraint on naming properties in the future.

Legend initializers gain edgecolor and facecolor kwargs

The Legend background patch (or ‘frame’) can have its edgecolor and facecolor determined by the corresponding keyword arguments to the matplotlib.legend.Legend initializer, or to any of the methods or functions that call that initializer. If left to their default values of None, their values will be taken from matplotlib.rcParams. The previously-existing framealpha kwarg still controls the alpha transparency of the patch.

Qualitative colormaps

Colorbrewer’s qualitative/discrete colormaps (“Accent”, “Dark2”, “Paired”, “Pastel1”, “Pastel2”, “Set1”, “Set2”, “Set3”) are now implemented as ListedColormap instead of LinearSegmentedColormap.

To use these for images where categories are specified as integers, for instance, use:

plt.imshow(x, cmap='Dark2', norm=colors.NoNorm())

Change in the draw_image backend API

The draw_image method implemented by backends has changed its interface.

This change is only relevant if the backend declares that it is able to transform images by returning True from option_scale_image. See the draw_image docstring for more information.

matplotlib.ticker.LinearLocator algorithm update

The matplotlib.ticker.LinearLocator is used to define the range and location of axis ticks when the user wants an exact number of ticks. LinearLocator thus differs from the default locator MaxNLocator, for which the user specifies a maximum number of intervals rather than a precise number of ticks.

The view range algorithm in matplotlib.ticker.LinearLocator has been changed so that more convenient tick locations are chosen. The new algorithm returns a plot view range that is a multiple of the user-requested number of ticks. This ensures tick marks will be located at whole integers more consistently. For example, when both y-axes of a``twinx`` plot use matplotlib.ticker.LinearLocator with the same number of ticks, their y-tick locations and grid lines will coincide.

matplotlib.ticker.LogLocator gains numticks kwarg

The maximum number of ticks generated by the LogLocator can now be controlled explicitly via setting the new ‘numticks’ kwarg to an integer. By default the kwarg is None which internally sets it to the ‘auto’ string, triggering a new algorithm for adjusting the maximum according to the axis length relative to the ticklabel font size.

matplotlib.ticker.LogFormatter: two new kwargs

Previously, minor ticks on log-scaled axes were not labeled by default. An algorithm has been added to the LogFormatter to control the labeling of ticks between integer powers of the base. The algorithm uses two parameters supplied in a kwarg tuple named ‘minor_thresholds’. See the docstring for further explanation.

To improve support for axes using SymmetricLogLocator, a ‘linthresh’ kwarg was added.

New defaults for 3D quiver function in

Matplotlib has both a 2D and a 3D quiver function. These changes affect only the 3D function and make the default behavior of the 3D function match the 2D version. There are two changes:

  1. The 3D quiver function previously normalized the arrows to be the same length, which makes it unusable for situations where the arrows should be different lengths and does not match the behavior of the 2D function. This normalization behavior is now controlled with the normalize keyword, which defaults to False.
  2. The pivot keyword now defaults to tail instead of tip. This was done in order to match the default behavior of the 2D quiver function.

To obtain the previous behavior with the 3D quiver function, one can call the function with

ax.quiver(x, y, z, u, v, w, normalize=True, pivot='tip')

where “ax” is an Axes3d object created with something like

import mpl_toolkits.mplot3d.axes3d
ax = plt.sublot(111, projection='3d')

Stale figure behavior

Attempting to draw the figure will now mark it as not stale (independent if the draw succeeds). This change is to prevent repeatedly trying to re-draw a figure which is raising an error on draw. The previous behavior would only mark a figure as not stale after a full re-draw succeeded.

The spectral colormap is now nipy_spectral

The colormaps formerly known as spectral and spectral_r have been replaced by nipy_spectral and nipy_spectral_r since Matplotlib 1.3.0. Even though the colormap was deprecated in Matplotlib 1.3.0, it never raised a warning. As of Matplotlib 2.0.0, using the old names raises a deprecation warning. In the future, using the old names will raise an error.

Default install no longer includes test images

To reduce the size of wheels and source installs, the tests and baseline images are no longer included by default.

To restore installing the tests and images, use a setup.cfg with

tests = True
toolkits_tests = True

in the source directory at build/install time.

Changes in 1.5.3

ax.plot(..., marker=None) gives default marker

Prior to 1.5.3 kwargs passed to plot were handled in two parts – default kwargs generated internal to plot (such as the cycled styles) and user supplied kwargs. The internally generated kwargs were passed to the matplotlib.lines.Line2D.__init__ and the user kwargs were passed to ln.set(**kwargs) to update the artist after it was created. Now both sets of kwargs are merged and passed to __init__. This change was made to allow None to be passed in via the user kwargs to mean ‘do the default thing’ as is the convention through out mpl rather than raising an exception.

Unlike most Line2D setter methods set_marker did accept None as a valid input which was mapped to ‘no marker’. Thus, by routing this marker=None through __init__ rather than set(...) the meaning of ax.plot(..., marker=None) changed from ‘no markers’ to ‘default markers from rcparams’.

This is change is only evident if mpl.rcParams['lines.marker'] has a value other than 'None' (which is string 'None' which means ‘no marker’).

Changes in 1.5.2

Default Behavior Changes

Changed default autorange behavior in boxplots

Prior to v1.5.2, the whiskers of boxplots would extend to the minimum and maximum values if the quartiles were all equal (i.e., Q1 = median = Q3). This behavior has been disabled by default to restore consistency with other plotting packages.

To restore the old behavior, simply set autorange=True when calling plt.boxplot.

Changes in 1.5.0

Code Changes

Reversed matplotlib.cbook.ls_mapper, added ls_mapper_r

Formerly, matplotlib.cbook.ls_mapper was a dictionary with the long-form line-style names ("solid") as keys and the short forms ("-") as values. This long-to-short mapping is now done by ls_mapper_r, and the short-to-long mapping is done by the ls_mapper.

Prevent moving artists between Axes, Property-ify Artist.axes, deprecate Artist.{get,set}_axes

This was done to prevent an Artist that is already associated with an Axes from being moved/added to a different Axes. This was never supported as it causes havoc with the transform stack. The apparent support for this (as it did not raise an exception) was the source of multiple bug reports and questions on SO.

For almost all use-cases, the assignment of the axes to an artist should be taken care of by the axes as part of the Axes.add_* method, hence the deprecation of {get,set}_axes.

Removing the set_axes method will also remove the ‘axes’ line from the ACCEPTS kwarg tables (assuming that the removal date gets here before that gets overhauled).

Tightened input validation on ‘pivot’ kwarg to quiver

Tightened validation so that only {‘tip’, ‘tail’, ‘mid’, and ‘middle’} (but any capitalization) are valid values for the ‘pivot’ kwarg in the Quiver.__init__ (and hence Axes.quiver and plt.quiver which both fully delegate to Quiver). Previously any input matching ‘mid.*’ would be interpreted as ‘middle’, ‘tip.*’ as ‘tip’ and any string not matching one of those patterns as ‘tail’.

The value of Quiver.pivot is normalized to be in the set {‘tip’, ‘tail’, ‘middle’} in Quiver.__init__.

Reordered Axes.get_children

The artist order returned by Axes.get_children did not match the one used by Axes.draw. They now use the same order, as Axes.draw now calls Axes.get_children.

Changed behaviour of contour plots

The default behaviour of contour() and contourf() when using a masked array is now determined by the new keyword argument corner_mask, or if this is not specified then the new rcParam contour.corner_mask instead. The new default behaviour is equivalent to using corner_mask=True; the previous behaviour can be obtained using corner_mask=False or by changing the rcParam. The example demonstrates the difference. Use of the old contouring algorithm, which is obtained with corner_mask='legacy', is now deprecated.

Contour labels may now appear in different places than in earlier versions of Matplotlib.

In addition, the keyword argument nchunk now applies to contour() as well as contourf(), and it subdivides the domain into subdomains of exactly nchunk by nchunk quads, whereas previously it was only roughly nchunk by nchunk quads.

The C/C++ object that performs contour calculations used to be stored in the public attribute QuadContourSet.Cntr, but is now stored in a private attribute and should not be accessed by end users.

Added set_params function to all Locator types

This was a bug fix targeted at making the api for Locators more consistent.

In the old behavior, only locators of type MaxNLocator have set_params() defined, causing its use on any other Locator to raise an AttributeError ( aside: set_params(args) is a function that sets the parameters of a Locator instance to be as specified within args). The fix involves moving set_params() to the Locator class such that all subtypes will have this function defined.

Since each of the Locator subtypes have their own modifiable parameters, a universal set_params() in Locator isn’t ideal. Instead, a default no-operation function that raises a warning is implemented in Locator. Subtypes extending Locator will then override with their own implementations. Subtypes that do not have a need for set_params() will fall back onto their parent’s implementation, which raises a warning as intended.

In the new behavior, Locator instances will not raise an AttributeError when set_params() is called. For Locators that do not implement set_params(), the default implementation in Locator is used.

Disallow None as x or y value in ax.plot

Do not allow None as a valid input for the x or y args in ax.plot. This may break some user code, but this was never officially supported (ex documented) and allowing None objects through can lead to confusing exceptions downstream.

To create an empty line use

ln1, = ax.plot([], [], ...)
ln2, = ax.plot([], ...)

In either case to update the data in the Line2D object you must update both the x and y data.

Removed args and kwargs from MicrosecondLocator.__call__

The call signature of __call__() has changed from __call__(self, *args, **kwargs) to __call__(self). This is consistent with the superclass Locator and also all the other Locators derived from this superclass.

No ValueError for the MicrosecondLocator and YearLocator

The MicrosecondLocator and YearLocator objects when called will return an empty list if the axes have no data or the view has no interval. Previously, they raised a ValueError. This is consistent with all the Date Locators.

‘OffsetBox.DrawingArea’ respects the ‘clip’ keyword argument

The call signature was OffsetBox.DrawingArea(..., clip=True) but nothing was done with the clip argument. The object did not do any clipping regardless of that parameter. Now the object can and does clip the child Artists if they are set to be clipped.

You can turn off the clipping on a per-child basis using child.set_clip_on(False).

Add salt to clipPath id

Add salt to the hash used to determine the id of the clipPath nodes. This is to avoid conflicts when two svg documents with the same clip path are included in the same document (see and ), however this means that the svg output is no longer deterministic if the same figure is saved twice. It is not expected that this will affect any users as the current ids are generated from an md5 hash of properties of the clip path and any user would have a very difficult time anticipating the value of the id.

Changed snap threshold for circle markers to inf

When drawing circle markers above some marker size (previously 6.0) the path used to generate the marker was snapped to pixel centers. However, this ends up distorting the marker away from a circle. By setting the snap threshold to inf snapping is never done on circles.

This change broke several tests, but is an improvement.

Preserve units with Text position

Previously the ‘get_position’ method on Text would strip away unit information even though the units were still present. There was no inherent need to do this, so it has been changed so that unit data (if present) will be preserved. Essentially a call to ‘get_position’ will return the exact value from a call to ‘set_position’.

If you wish to get the old behaviour, then you can use the new method called ‘get_unitless_position’.

New API for custom Axes view changes

Interactive pan and zoom were previously implemented using a Cartesian-specific algorithm that was not necessarily applicable to custom Axes. Three new private methods, _get_view(), _set_view(), and _set_view_from_bbox(), allow for custom Axes classes to override the pan and zoom algorithms. Implementors of custom Axes who override these methods may provide suitable behaviour for both pan and zoom as well as the view navigation buttons on the interactive toolbars.

MathTex visual changes

The spacing commands in mathtext have been changed to more closely match vanilla TeX.

Improved spacing in mathtext

The extra space that appeared after subscripts and superscripts has been removed.

No annotation coordinates wrap

In #2351 for 1.4.0 the behavior of [‘axes points’, ‘axes pixel’, ‘figure points’, ‘figure pixel’] as coordinates was change to no longer wrap for negative values. In 1.4.3 this change was reverted for ‘axes points’ and ‘axes pixel’ and in addition caused ‘axes fraction’ to wrap. For 1.5 the behavior has been reverted to as it was in 1.4.0-1.4.2, no wrapping for any type of coordinate.


Deprecated GraphicsContextBase.set_graylevel

The GraphicsContextBase.set_graylevel function has been deprecated in 1.5 and will be removed in 1.6. It has been unused. The GraphicsContextBase.set_foreground could be used instead.

deprecated idle_event

The idle_event was broken or missing in most backends and causes spurious warnings in some cases, and its use in creating animations is now obsolete due to the animations module. Therefore code involving it has been removed from all but the wx backend (where it partially works), and its use is deprecated. The animations module may be used instead to create animations.

color_cycle deprecated

In light of the new property cycling feature, the Axes method set_color_cycle is now deprecated. Calling this method will replace the current property cycle with one that cycles just the given colors.

Similarly, the rc parameter axes.color_cycle is also deprecated in lieu of the new axes.prop_cycle parameter. Having both parameters in the same rc file is not recommended as the result cannot be predicted. For compatibility, setting axes.color_cycle will replace the cycler in axes.prop_cycle with a color cycle. Accessing axes.color_cycle will return just the color portion of the property cycle, if it exists.

Timeline for removal has not been set.

Bundled jquery

The version of jquery bundled with the webagg backend has been upgraded from 1.7.1 to 1.11.3. If you are using the version of jquery bundled with webagg you will need to update your html files as such

-    <script src="_static/jquery/js/jquery-1.7.1.min.js"></script>
+    <script src="_static/jquery/js/jquery-1.11.3.min.js"></script>

Code Removed

Removed Image from main namespace

Image was imported from PIL/pillow to test if PIL is available, but there is no reason to keep Image in the namespace once the availability has been determined.

Removed lod from Artist

Removed the method set_lod and all references to the attribute _lod as the are not used anywhere else in the code base. It appears to be a feature stub that was never built out.

Removed Lena images from sample_data

The lena.png and lena.jpg images have been removed from Matplotlib’s sample_data directory. The images are also no longer available from matplotlib.cbook.get_sample_data. We suggest using matplotlib.cbook.get_sample_data('grace_hopper.png') or matplotlib.cbook.get_sample_data('grace_hopper.jpg') instead.


Removed handling of loc as a positional argument to Legend

Legend handlers

Remove code to allow legend handlers to be callable. They must now implement a method legend_artist.


Removed method set_scale. This is now handled via a private method which should not be used directly by users. It is called via Axes.set_{x,y}scale which takes care of ensuring the related changes are also made to the Axes object.

Removed functions with ambiguous argument order from


Removed textcoords and xytext proprieties from Annotation objects.


Both ipython_console_highlighting and ipython_directive have been moved to IPython.

Change your import from ‘matplotlib.sphinxext.ipython_directive’ to ‘IPython.sphinxext.ipython_directive’ and from ‘matplotlib.sphinxext.ipython_directive’ to ‘IPython.sphinxext.ipython_directive’


Deprecated in 2005, use set_color

remove 'faceted' as a valid value for shading in tri.tripcolor

Use edgecolor instead. Added validation on shading to only be valid values.

Remove faceted kwarg from scatter

Remove support for the faceted kwarg. This was deprecated in d48b34288e9651ff95c3b8a071ef5ac5cf50bae7 (2008-04-18!) and replaced by edgecolor.

Remove set_colorbar method from ScalarMappable

Remove set_colorbar method, use colorbar attribute directly.


  • remove get_proxy_renderer method from AbstarctPathEffect class
  • remove patch_alpha and offset_xy from SimplePatchShadow


Contained only a no-longer used port of functionality from PIL

Remove mlab.FIFOBuffer

Not used internally and not part of core mission of mpl.

Remove mlab.prepca

Deprecated in 2009.

Remove NavigationToolbar2QTAgg

Added no functionality over the base NavigationToolbar2Qt

Remove the module matplotlib.mpl. Deprecated in 1.3 by PR #1670 and commit 78ce67d161625833cacff23cfe5d74920248c5b2

Changes in 1.4.x

Code changes

  • A major refactoring of the axes module was made. The axes module has been split into smaller modules:

    • the _base module, which contains a new private _AxesBase class. This class contains all methods except plotting and labelling methods.
    • the axes module, which contains the Axes class. This class inherits from _AxesBase, and contains all plotting and labelling methods.
    • the _subplot module, with all the classes concerning subplotting.

There are a couple of things that do not exists in the axes module’s namespace anymore. If you use them, you need to import them from their original location:

  • math -> import math
  • ma -> from numpy import ma
  • cbook -> from matplotlib import cbook
  • docstring -> from matplotlib import docstring
  • is_sequence_of_strings -> from matplotlib.cbook import is_sequence_of_strings
  • is_string_like -> from matplotlib.cbook import is_string_like
  • iterable -> from matplotlib.cbook import iterable
  • itertools -> import itertools
  • martist -> from matplotlib import artist as martist
  • matplotlib -> import matplotlib
  • mcoll -> from matplotlib import collections as mcoll
  • mcolors -> from matplotlib import colors as mcolors
  • mcontour -> from matplotlib import contour as mcontour
  • mpatches -> from matplotlib import patches as mpatches
  • mpath -> from matplotlib import path as mpath
  • mquiver -> from matplotlib import quiver as mquiver
  • mstack -> from matplotlib import stack as mstack
  • mstream -> from matplotlib import stream as mstream
  • mtable -> from matplotlib import table as mtable
  • As part of the refactoring to enable Qt5 support, the module matplotlib.backends.qt4_compat was renamed to matplotlib.qt_compat. qt4_compat is deprecated in 1.4 and will be removed in 1.5.

  • The errorbar() method has been changed such that the upper and lower limits (lolims, uplims, xlolims, xuplims) now point in the correct direction.

  • The fmt kwarg for plot() defaults.

  • A bug has been fixed in the path effects rendering of fonts, which now means that the font size is consistent with non-path effect fonts. See for more detail.

  • The Sphinx extensions ipython_directive and ipython_console_highlighting have been moved to the IPython project itself. While they remain in Matplotlib for this release, they have been deprecated. Update your extensions in to point to IPython.sphinxext.ipython_directive instead of matplotlib.sphinxext.ipython_directive.

  • In finance, almost all functions have been deprecated and replaced with a pair of functions name *_ochl and *_ohlc. The former is the ‘open-close-high-low’ order of quotes used previously in this module, and the latter is the ‘open-high-low-close’ order that is standard in finance.

  • For consistency the face_alpha keyword to matplotlib.patheffects.SimplePatchShadow has been deprecated in favour of the alpha keyword. Similarly, the keyword offset_xy is now named offset across all _Base` has been renamed to matplotlib.patheffects.AbstractPathEffect. matplotlib.patheffect.ProxyRenderer has been renamed to matplotlib.patheffects.PathEffectRenderer and is now a full RendererBase subclass.

  • The artist used to draw the outline of a colorbar has been changed from a matplotlib.lines.Line2D to matplotlib.patches.Polygon, thus colorbar.ColorbarBase.outline is now a matplotlib.patches.Polygon object.

  • The legend handler interface has changed from a callable, to any object which implements the legend_artists method (a deprecation phase will see this interface be maintained for v1.4). See Legend guide for further details. Further legend changes include:

    • matplotlib.axes.Axes._get_legend_handles() now returns a generator of handles, rather than a list.
    • The legend() function’s “loc” positional argument has been deprecated. Use the “loc” keyword instead.
  • The rcParams savefig.transparent has been added to control default transparency when saving figures.

  • Slightly refactored the Annotation family. The text location in Annotation is now handled entirely handled by the underlying Text object so set_position works as expected. The attributes xytext and textcoords have been deprecated in favor of xyann and anncoords so that Annotation and AnnotaionBbox can share a common sensibly named api for getting/setting the location of the text or box.

    • xyann -> set the location of the annotation
    • xy -> set where the arrow points to
    • anncoords -> set the units of the annotation location
    • xycoords -> set the units of the point location
    • set_position() -> Annotation only set location of annotation
  • matplotlib.mlab.specgram, matplotlib.mlab.psd, matplotlib.mlab.csd, matplotlib.mlab.cohere, matplotlib.mlab.cohere_pairs, matplotlib.pyplot.specgram, matplotlib.pyplot.psd, matplotlib.pyplot.csd, and matplotlib.pyplot.cohere now raise ValueError where they previously raised AssertionError.

  • For matplotlib.mlab.psd, matplotlib.mlab.csd, matplotlib.mlab.cohere, matplotlib.mlab.cohere_pairs, matplotlib.pyplot.specgram, matplotlib.pyplot.psd, matplotlib.pyplot.csd, and matplotlib.pyplot.cohere, in cases where a shape (n, 1) array is returned, this is now converted to a (n, ) array. Previously, (n, m) arrays were averaged to an (n, ) array, but (n, 1) arrays were returend unchanged. This change makes the dimensions consistent in both cases.

  • Added the rcParam axes.fromatter.useoffset to control the default value of useOffset in ticker.ScalarFormatter

  • Added Formatter sub-class StrMethodFormatter which does the exact same thing as FormatStrFormatter, but for new-style formatting strings.

  • Deprecated matplotlib.testing.image_util and the only function within, matplotlib.testing.image_util.autocontrast. These will be removed completely in v1.5.0.

  • The fmt argument of plot_date() has been changed from bo to just o, so color cycling can happen by default.

  • Removed the class FigureManagerQTAgg and deprecated NavigationToolbar2QTAgg which will be removed in 1.5.

  • Removed formerly public (non-prefixed) attributes rect and drawRect from FigureCanvasQTAgg; they were always an implementation detail of the (preserved) drawRectangle() function.

  • The function signatures of tight_bbox.adjust_bbox and tight_bbox.process_figure_for_rasterizing have been changed. A new fixed_dpi parameter allows for overriding the figure.dpi setting instead of trying to deduce the intended behaviour from the file format.

  • Added support for horizontal/vertical axes padding to mpl_toolkits.axes_grid1.ImageGrid — argument axes_pad can now be tuple-like if separate axis padding is required. The original behavior is preserved.

  • Added support for skewed transforms to matplotlib.transforms.Affine2D, which can be created using the skew and skew_deg methods.

  • Added clockwise parameter to control sectors direction in axes.pie

  • In matplotlib.lines.Line2D the markevery functionality has been extended. Previously an integer start-index and stride-length could be specified using either a two-element-list or a two-element-tuple. Now this can only be done using a two-element-tuple. If a two-element-list is used then it will be treated as numpy fancy indexing and only the two markers corresponding to the given indexes will be shown.

  • removed prop kwarg from mpl_toolkits.axes_grid1.anchored_artists.AnchoredSizeBar call. It was passed through to the base-class __init__ and is only used for setting padding. Now fontproperties (which is what is really used to set the font properties of AnchoredSizeBar) is passed through in place of prop. If fontpropreties is not passed in, but prop is, then prop is used inplace of fontpropreties. If both are passed in, prop is silently ignored.

  • The use of the index 0 in plt.subplot and related commands is deprecated. Due to a lack of validation calling plt.subplots(2, 2, 0) does not raise an exception, but puts an axes in the _last_ position. This is due to the indexing in subplot being 1-based (to mirror MATLAB) so before indexing into the GridSpec object used to determine where the axes should go, 1 is subtracted off. Passing in 0 results in passing -1 to GridSpec which results in getting the last position back. Even though this behavior is clearly wrong and not intended, we are going through a deprecation cycle in an abundance of caution that any users are exploiting this ‘feature’. The use of 0 as an index will raise a warning in 1.4 and an exception in 1.5.

  • Clipping is now off by default on offset boxes.

  • Matplotlib now uses a less-aggressive call to gc.collect(1) when closing figures to avoid major delays with large numbers of user objects in memory.

  • The default clip value of all pie artists now defaults to False.

Code removal

  • Removed mlab.levypdf. The code raised a numpy error (and has for a long time) and was not the standard form of the Levy distribution. scipy.stats.levy should be used instead

Changes in 1.3.x

Changes in 1.3.1

It is rare that we make an API change in a bugfix release, however, for 1.3.1 since 1.3.0 the following change was made:

  • text.Text.cached (used to cache font objects) has been made into a private variable. Among the obvious encapsulation benefit, this removes this confusing-looking member from the documentation.
  • The method hist() now always returns bin occupancies as an array of type float. Previously, it was sometimes an array of type int, depending on the call.

Code removal

  • The following items that were deprecated in version 1.2 or earlier have now been removed completely.

    • The Qt 3.x backends (qt and qtagg) have been removed in favor of the Qt 4.x backends (qt4 and qt4agg).

    • The FltkAgg and Emf backends have been removed.

    • The matplotlib.nxutils module has been removed. Use the functionality on matplotlib.path.Path.contains_point and friends instead.

    • Instead of axes.Axes.get_frame, use axes.Axes.patch.

    • The following kwargs to the legend function have been renamed:

      • pad -> borderpad
      • labelsep -> labelspacing
      • handlelen -> handlelength
      • handletextsep -> handletextpad
      • axespad -> borderaxespad

      Related to this, the following rcParams have been removed:

      • legend.pad, legend.labelsep, legend.handlelen, legend.handletextsep and legend.axespad
    • For the hist function, instead of width, use rwidth (relative width).

    • On patches.Circle, the resolution kwarg has been removed. For a circle made up of line segments, use patches.CirclePolygon.

    • The printing functions in the Wx backend have been removed due to the burden of keeping them up-to-date.

    • mlab.liaupunov has been removed.

    •, mlab.load, and pylab.load have been removed. We recommend using numpy.savetxt and numpy.loadtxt instead.

    • widgets.HorizontalSpanSelector has been removed. Use widgets.SpanSelector instead.

Code deprecation

  • The CocoaAgg backend has been deprecated, with the possibility for deletion or resurrection in a future release.

  • The top-level functions in matplotlib.path that are implemented in C++ were never meant to be public. Instead, users should use the Pythonic wrappers for them in the path.Path and collections.Collection classes. Use the following mapping to update your code:

    • point_in_path -> path.Path.contains_point
    • get_path_extents -> path.Path.get_extents
    • point_in_path_collection -> collection.Collection.contains
    • path_in_path -> path.Path.contains_path
    • path_intersects_path -> path.Path.intersects_path
    • convert_path_to_polygons -> path.Path.to_polygons
    • cleanup_path -> path.Path.cleaned
    • points_in_path -> path.Path.contains_points
    • clip_path_to_rect -> path.Path.clip_to_bbox
  • matplotlib.colors.normalize and matplotlib.colors.no_norm have been deprecated in favour of matplotlib.colors.Normalize and matplotlib.colors.NoNorm respectively.

  • The ScalarMappable class’ set_colorbar is now deprecated. Instead, the attribute should be used. In previous Matplotlib versions this attribute was an undocumented tuple of (colorbar_instance, colorbar_axes) but is now just colorbar_instance. To get the colorbar axes it is possible to just use the ax attribute on a colorbar instance.

  • The mpl module is now deprecated. Those who relied on this module should transition to simply using import matplotlib as mpl.

Code changes

  • Patch now fully supports using RGBA values for its facecolor and edgecolor attributes, which enables faces and edges to have different alpha values. If the Patch object’s alpha attribute is set to anything other than None, that value will override any alpha-channel value in both the face and edge colors. Previously, if Patch had alpha=None, the alpha component of edgecolor would be applied to both the edge and face.

  • The optional isRGB argument to set_foreground() (and the other GraphicsContext classes that descend from it) has been renamed to isRGBA, and should now only be set to True if the fg color argument is known to be an RGBA tuple.

  • For Patch, the capstyle used is now butt, to be consistent with the default for most other objects, and to avoid problems with non-solid linestyle appearing solid when using a large linewidth. Previously, Patch used capstyle='projecting'.

  • Path objects can now be marked as readonly by passing readonly=True to its constructor. The built-in path singletons, obtained through Path.unit* class methods return readonly paths. If you have code that modified these, you will need to make a deepcopy first, using either:

    import copy
    path = copy.deepcopy(Path.unit_circle())
    # or
    path = Path.unit_circle().deepcopy()

    Deep copying a Path always creates an editable (i.e. non-readonly) Path.

  • The list at Path.NUM_VERTICES was replaced by a dictionary mapping Path codes to the number of expected vertices at NUM_VERTICES_FOR_CODE.

  • To support XKCD style plots, the matplotlib.path.cleanup_path() method’s signature was updated to require a sketch argument. Users of matplotlib.path.cleanup_path() are encouraged to use the new cleaned() Path method.

  • Data limits on a plot now start from a state of having “null” limits, rather than limits in the range (0, 1). This has an effect on artists that only control limits in one direction, such as axvline and axhline, since their limits will not longer also include the range (0, 1). This fixes some problems where the computed limits would be dependent on the order in which artists were added to the axes.

  • Fixed a bug in setting the position for the right/top spine with data position type. Previously, it would draw the right or top spine at +1 data offset.

  • In FancyArrow, the default arrow head width, head_width, has been made larger to produce a visible arrow head. The new value of this kwarg is head_width = 20 * width.

  • It is now possible to provide number of levels + 1 colors in the case of extend='both' for contourf (or just number of levels colors for an extend value min or max) such that the resulting colormap’s set_under and set_over are defined appropriately. Any other number of colors will continue to behave as before (if more colors are provided than levels, the colors will be unused). A similar change has been applied to contour, where extend='both' would expect number of levels + 2 colors.

  • A new keyword extendrect in colorbar() and ColorbarBase allows one to control the shape of colorbar extensions.

  • The extension of MultiCursor to both vertical (default) and/or horizontal cursor implied that self.line is replaced by self.vline for vertical cursors lines and self.hline is added for the horizontal cursors lines.

  • On POSIX platforms, the report_memory() function raises NotImplementedError instead of OSError if the ps command cannot be run.

  • The matplotlib.cbook.check_output() function has been moved to matplotlib.compat.subprocess().

Configuration and rcParams

  • On Linux, the user-specific matplotlibrc configuration file is now located in config/matplotlib/matplotlibrc to conform to the XDG Base Directory Specification.
  • The font.* rcParams now affect only text objects created after the rcParam has been set, and will not retroactively affect already existing text objects. This brings their behavior in line with most other rcParams.
  • Removed call of grid() in plotfile(). To draw the axes grid, set the axes.grid rcParam to True, or explicitly call grid().

Changes in 1.2.x

  • The classic option of the rc parameter toolbar is deprecated and will be removed in the next release.

  • The isvector() method has been removed since it is no longer functional.

  • The rasterization_zorder property on Axes a zorder below which artists are rasterized. This has defaulted to -30000.0, but it now defaults to None, meaning no artists will be rasterized. In order to rasterize artists below a given zorder value, set_rasterization_zorder must be explicitly called.

  • In scatter(), and scatter, when specifying a marker using a tuple, the angle is now specified in degrees, not radians.

  • Using twinx() or twiny() no longer overrides the current locaters and formatters on the axes.

  • In contourf(), the handling of the extend kwarg has changed. Formerly, the extended ranges were mapped after to 0, 1 after being normed, so that they always corresponded to the extreme values of the colormap. Now they are mapped outside this range so that they correspond to the special colormap values determined by the set_under() and set_over() methods, which default to the colormap end points.

  • The new rc parameter savefig.format replaces cairo.format and savefig.extension, and sets the default file format used by matplotlib.figure.Figure.savefig().

  • In pie() and pie(), one can now set the radius of the pie; setting the radius to ‘None’ (the default value), will result in a pie with a radius of 1 as before.

  • Use of projection_factory() is now deprecated in favour of axes class identification using process_projection_requirements() followed by direct axes class invocation (at the time of writing, functions which do this are: add_axes(), add_subplot() and gca()). Therefore:

    key = figure._make_key(*args, **kwargs)
    ispolar = kwargs.pop('polar', False)
    projection = kwargs.pop('projection', None)
    if ispolar:
        if projection is not None and projection != 'polar':
            raise ValueError('polar and projection args are inconsistent')
        projection = 'polar'
    ax = projection_factory(projection, self, rect, **kwargs)
    key = self._make_key(*args, **kwargs)
    # is now
    projection_class, kwargs, key = \
                       process_projection_requirements(self, *args, **kwargs)
    ax = projection_class(self, rect, **kwargs)

    This change means that third party objects can expose themselves as Matplotlib axes by providing a _as_mpl_axes method. See Developer’s guide for creating scales and transformations for more detail.

  • A new keyword extendfrac in colorbar() and ColorbarBase allows one to control the size of the triangular minimum and maximum extensions on colorbars.

  • A new keyword capthick in errorbar() has been added as an intuitive alias to the markeredgewidth and mew keyword arguments, which indirectly controlled the thickness of the caps on the errorbars. For backwards compatibility, specifying either of the original keyword arguments will override any value provided by capthick.

  • Transform subclassing behaviour is now subtly changed. If your transform implements a non-affine transformation, then it should override the transform_non_affine method, rather than the generic transform method. Previously transforms would define transform and then copy the method into transform_non_affine:

    class MyTransform(mtrans.Transform):
        def transform(self, xy):
        transform_non_affine = transform

    This approach will no longer function correctly and should be changed to:

    class MyTransform(mtrans.Transform):
        def transform_non_affine(self, xy):
  • Artists no longer have x_isdata or y_isdata attributes; instead any artist’s transform can be interrogated with artist_instance.get_transform().contains_branch(ax.transData)

  • Lines added to an axes now take into account their transform when updating the data and view limits. This means transforms can now be used as a pre-transform. For instance:

    >>> import matplotlib.pyplot as plt
    >>> import matplotlib.transforms as mtrans
    >>> ax = plt.axes()
    >>> ax.plot(range(10), transform=mtrans.Affine2D().scale(10) + ax.transData)
    >>> print(ax.viewLim)
    Bbox('array([[  0.,   0.],\n       [ 90.,  90.]])')
  • One can now easily get a transform which goes from one transform’s coordinate system to another, in an optimized way, using the new subtract method on a transform. For instance, to go from data coordinates to axes coordinates:

    >>> import matplotlib.pyplot as plt
    >>> ax = plt.axes()
    >>> data2ax = ax.transData - ax.transAxes
    >>> print(ax.transData.depth, ax.transAxes.depth)
    3, 1
    >>> print(data2ax.depth)

    for versions before 1.2 this could only be achieved in a sub-optimal way, using ax.transData + ax.transAxes.inverted() (depth is a new concept, but had it existed it would return 4 for this example).

  • twinx and twiny now returns an instance of SubplotBase if parent axes is an instance of SubplotBase.

  • All Qt3-based backends are now deprecated due to the lack of py3k bindings. Qt and QtAgg backends will continue to work in v1.2.x for py2.6 and py2.7. It is anticipated that the Qt3 support will be completely removed for the next release.

  • ColorConverter, Colormap and Normalize now subclasses object

  • ContourSet instances no longer have a transform attribute. Instead, access the transform with the get_transform method.

Changes in 1.1.x

  • Added new matplotlib.sankey.Sankey for generating Sankey diagrams.
  • In imshow(), setting interpolation to ‘nearest’ will now always mean that the nearest-neighbor interpolation is performed. If you want the no-op interpolation to be performed, choose ‘none’.
  • There were errors in how the tri-functions were handling input parameters that had to be fixed. If your tri-plots are not working correctly anymore, or you were working around apparent mistakes, please see issue #203 in the github tracker. When in doubt, use kwargs.
  • The ‘symlog’ scale had some bad behavior in previous versions. This has now been fixed and users should now be able to use it without frustrations. The fixes did result in some minor changes in appearance for some users who may have been depending on the bad behavior.
  • There is now a common set of markers for all plotting functions. Previously, some markers existed only for scatter() or just for plot(). This is now no longer the case. This merge did result in a conflict. The string ‘d’ now means “thin diamond” while ‘D’ will mean “regular diamond”.

Changes beyond 0.99.x

  • The default behavior of matplotlib.axes.Axes.set_xlim(), matplotlib.axes.Axes.set_ylim(), and matplotlib.axes.Axes.axis(), and their corresponding pyplot functions, has been changed: when view limits are set explicitly with one of these methods, autoscaling is turned off for the matching axis. A new auto kwarg is available to control this behavior. The limit kwargs have been renamed to left and right instead of xmin and xmax, and bottom and top instead of ymin and ymax. The old names may still be used, however.
  • There are five new Axes methods with corresponding pyplot functions to facilitate autoscaling, tick location, and tick label formatting, and the general appearance of ticks and tick labels:
  • The method accepts a error_kw kwarg; it is a dictionary of kwargs to be passed to the errorbar function.
  • The matplotlib.axes.Axes.hist() color kwarg now accepts a sequence of color specs to match a sequence of datasets.
  • The EllipseCollection has been changed in two ways:
    • There is a new units option, ‘xy’, that scales the ellipse with the data units. This matches the :class:’~matplotlib.patches.Ellipse` scaling.
    • The height and width kwargs have been changed to specify the height and width, again for consistency with Ellipse, and to better match their names; previously they specified the half-height and half-width.
  • There is a new rc parameter axes.color_cycle, and the color cycle is now independent of the rc parameter lines.color. matplotlib.Axes.set_default_color_cycle() is deprecated.
  • You can now print several figures to one pdf file and modify the document information dictionary of a pdf file. See the docstrings of the class matplotlib.backends.backend_pdf.PdfPages for more information.
  • Removed configobj and enthought.traits packages, which are only required by the experimental traited config and are somewhat out of date. If needed, install them independently.
  • The new rc parameter savefig.extension sets the filename extension that is used by matplotlib.figure.Figure.savefig() if its fname argument lacks an extension.

  • In an effort to simplify the backend API, all clipping rectangles and paths are now passed in using GraphicsContext objects, even on collections and images. Therefore:

    draw_path_collection(self, master_transform, cliprect, clippath,
                         clippath_trans, paths, all_transforms, offsets,
                         offsetTrans, facecolors, edgecolors, linewidths,
                         linestyles, antialiaseds, urls)
    # is now
    draw_path_collection(self, gc, master_transform, paths, all_transforms,
                         offsets, offsetTrans, facecolors, edgecolors,
                         linewidths, linestyles, antialiaseds, urls)
    draw_quad_mesh(self, master_transform, cliprect, clippath,
                   clippath_trans, meshWidth, meshHeight, coordinates,
                   offsets, offsetTrans, facecolors, antialiased,
    # is now
    draw_quad_mesh(self, gc, master_transform, meshWidth, meshHeight,
                   coordinates, offsets, offsetTrans, facecolors,
                   antialiased, showedges)
    draw_image(self, x, y, im, bbox, clippath=None, clippath_trans=None)
    # is now
    draw_image(self, gc, x, y, im)
  • There are four new Axes methods with corresponding pyplot functions that deal with unstructured triangular grids:

Changes in 0.99

  • pylab no longer provides a load and save function. These are available in matplotlib.mlab, or you can use numpy.loadtxt and numpy.savetxt for text files, or and np.load for binary numpy arrays.
  • User-generated colormaps can now be added to the set recognized by Colormaps can be made the default and applied to the current image using matplotlib.pyplot.set_cmap().
  • changed use_mrecords default to False in mlab.csv2rec since this is partially broken
  • Axes instances no longer have a “frame” attribute. Instead, use the new “spines” attribute. Spines is a dictionary where the keys are the names of the spines (e.g., ‘left’,’right’ and so on) and the values are the artists that draw the spines. For normal (rectilinear) axes, these artists are Line2D instances. For other axes (such as polar axes), these artists may be Patch instances.
  • Polar plots no longer accept a resolution kwarg. Instead, each Path must specify its own number of interpolation steps. This is unlikely to be a user-visible change – if interpolation of data is required, that should be done before passing it to Matplotlib.

Changes for 0.98.x

  • psd(), csd(), and cohere() will now automatically wrap negative frequency components to the beginning of the returned arrays. This is much more sensible behavior and makes them consistent with specgram(). The previous behavior was more of an oversight than a design decision.

  • Added new keyword parameters nonposx, nonposy to matplotlib.axes.Axes methods that set log scale parameters. The default is still to mask out non-positive values, but the kwargs accept ‘clip’, which causes non-positive values to be replaced with a very small positive value.

  • Added new matplotlib.pyplot.fignum_exists() and matplotlib.pyplot.get_fignums(); they merely expose information that had been hidden in matplotlib._pylab_helpers.

  • Deprecated numerix package.

  • Added new matplotlib.image.imsave() and exposed it to the matplotlib.pyplot interface.

  • Remove support for pyExcelerator in exceltools – use xlwt instead

  • Changed the defaults of acorr and xcorr to use usevlines=True, maxlags=10 and normed=True since these are the best defaults

  • Following keyword parameters for matplotlib.label.Label are now deprecated and new set of parameters are introduced. The new parameters are given as a fraction of the font-size. Also, scatteryoffsets, fancybox and columnspacing are added as keyword parameters.

    Deprecated New
    pad borderpad
    labelsep labelspacing
    handlelen handlelength
    handlestextsep handletextpad
    axespad borderaxespad
  • Removed the configobj and experimental traits rc support

  • Modified matplotlib.mlab.psd(), matplotlib.mlab.csd(), matplotlib.mlab.cohere(), and matplotlib.mlab.specgram() to scale one-sided densities by a factor of 2. Also, optionally scale the densities by the sampling frequency, which gives true values of densities that can be integrated by the returned frequency values. This also gives better MATLAB compatibility. The corresponding matplotlib.axes.Axes methods and matplotlib.pyplot functions were updated as well.

  • Font lookup now uses a nearest-neighbor approach rather than an exact match. Some fonts may be different in plots, but should be closer to what was requested.

  • matplotlib.axes.Axes.set_xlim(), matplotlib.axes.Axes.set_ylim() now return a copy of the viewlim array to avoid modify-in-place surprises.

  • matplotlib.afm.AFM.get_fullname() and matplotlib.afm.AFM.get_familyname() no longer raise an exception if the AFM file does not specify these optional attributes, but returns a guess based on the required FontName attribute.

  • Changed precision kwarg in matplotlib.pyplot.spy(); default is 0, and the string value ‘present’ is used for sparse arrays only to show filled locations.

  • matplotlib.collections.EllipseCollection added.

  • Added angles kwarg to matplotlib.pyplot.quiver() for more flexible specification of the arrow angles.

  • Deprecated (raise NotImplementedError) all the mlab2 functions from matplotlib.mlab out of concern that some of them were not clean room implementations.

  • Methods matplotlib.collections.Collection.get_offsets() and matplotlib.collections.Collection.set_offsets() added to Collection base class.

  • matplotlib.figure.Figure.figurePatch renamed matplotlib.figure.Figure.patch; matplotlib.axes.Axes.axesPatch renamed matplotlib.axes.Axes.patch; matplotlib.axes.Axes.axesFrame renamed matplotlib.axes.Axes.frame. matplotlib.axes.Axes.get_frame(), which returns matplotlib.axes.Axes.patch, is deprecated.

  • Changes in the matplotlib.contour.ContourLabeler attributes (matplotlib.pyplot.clabel() function) so that they all have a form like .labelAttribute. The three attributes that are most likely to be used by end users, .cl, .cl_xy and .cl_cvalues have been maintained for the moment (in addition to their renamed versions), but they are deprecated and will eventually be removed.

  • Moved several functions in matplotlib.mlab and matplotlib.cbook into a separate module matplotlib.numerical_methods because they were unrelated to the initial purpose of mlab or cbook and appeared more coherent elsewhere.

Changes for 0.98.1

  • Removed broken matplotlib.axes3d support and replaced it with a non-implemented error pointing to 0.91.x

Changes for 0.98.0

  • matplotlib.image.imread() now no longer always returns RGBA data—if the image is luminance or RGB, it will return a MxN or MxNx3 array if possible. Also uint8 is no longer always forced to float.
  • Rewrote the callback infrastructure to use matplotlib.cbook.CallbackRegistry rather than custom callback handling. Any users of of the ScalarMappable should use the CallbackRegistry instead.
  • New axes function and Axes method provide control over the plot color cycle: matplotlib.axes.set_default_color_cycle() and matplotlib.axes.Axes.set_color_cycle().
  • Matplotlib now requires Python 2.4, so matplotlib.cbook will no longer provide set, enumerate(), reversed() or izip() compatibility functions.
  • In Numpy 1.0, bins are specified by the left edges only. The axes method matplotlib.axes.Axes.hist() now uses future Numpy 1.3 semantics for histograms. Providing binedges, the last value gives the upper-right edge now, which was implicitly set to +infinity in Numpy 1.0. This also means that the last bin doesn’t contain upper outliers any more by default.
  • New axes method and pyplot function, hexbin(), is an alternative to scatter() for large datasets. It makes something like a pcolor() of a 2-D histogram, but uses hexagonal bins.
  • New kwarg, symmetric, in matplotlib.ticker.MaxNLocator allows one require an axis to be centered around zero.
  • Toolkits must now be imported from mpl_toolkits (not matplotlib.toolkits)

Notes about the transforms refactoring

A major new feature of the 0.98 series is a more flexible and extensible transformation infrastructure, written in Python/Numpy rather than a custom C extension.

The primary goal of this refactoring was to make it easier to extend matplotlib to support new kinds of projections. This is mostly an internal improvement, and the possible user-visible changes it allows are yet to come.

See matplotlib.transforms for a description of the design of the new transformation framework.

For efficiency, many of these functions return views into Numpy arrays. This means that if you hold on to a reference to them, their contents may change. If you want to store a snapshot of their current values, use the Numpy array method copy().

The view intervals are now stored only in one place – in the matplotlib.axes.Axes instance, not in the locator instances as well. This means locators must get their limits from their matplotlib.axis.Axis, which in turn looks up its limits from the Axes. If a locator is used temporarily and not assigned to an Axis or Axes, (e.g., in matplotlib.contour), a dummy axis must be created to store its bounds. Call matplotlib.ticker.Locator.create_dummy_axis() to do so.

The functionality of Pbox has been merged with Bbox. Its methods now all return copies rather than modifying in place.

The following lists many of the simple changes necessary to update code from the old transformation framework to the new one. In particular, methods that return a copy are named with a verb in the past tense, whereas methods that alter an object in place are named with a verb in the present tense.


Old method New method
Bbox.get_bounds() transforms.Bbox.bounds
Bbox.width() transforms.Bbox.width
Bbox.height() transforms.Bbox.height
Bbox.intervalx().get_bounds() transforms.Bbox.intervalx
Bbox.intervalx().set_bounds() [Bbox.intervalx is now a property.]
Bbox.intervaly().get_bounds() transforms.Bbox.intervaly
Bbox.intervaly().set_bounds() [Bbox.intervaly is now a property.]
Bbox.xmin() transforms.Bbox.x0 or transforms.Bbox.xmin [1]
Bbox.ymin() transforms.Bbox.y0 or transforms.Bbox.ymin [1]
Bbox.xmax() transforms.Bbox.x1 or transforms.Bbox.xmax [1]
Bbox.ymax() transforms.Bbox.y1 or transforms.Bbox.ymax [1]
Bbox.overlaps(bboxes) Bbox.count_overlaps(bboxes)
bbox_all(bboxes) Bbox.union(bboxes) [transforms.Bbox.union() is a staticmethod.]
lbwh_to_bbox(l, b, w, h) Bbox.from_bounds(x0, y0, w, h) [transforms.Bbox.from_bounds() is a staticmethod.]
inverse_transform_bbox(trans, bbox) Bbox.inverse_transformed(trans)
Interval.contains_open(v) interval_contains_open(tuple, v)
Interval.contains(v) interval_contains(tuple, v)
identity_transform() matplotlib.transforms.IdentityTransform
blend_xy_sep_transform(xtrans, ytrans) blended_transform_factory(xtrans, ytrans)
scale_transform(xs, ys) Affine2D().scale(xs[, ys])
get_bbox_transform(boxin, boxout) BboxTransform(boxin, boxout) or BboxTransformFrom(boxin) or BboxTransformTo(boxout)
Transform.seq_xy_tup(points) Transform.transform(points)
Transform.inverse_xy_tup(points) Transform.inverted().transform(points)
[1](1, 2, 3, 4) The Bbox is bound by the points (x0, y0) to (x1, y1) and there is no defined order to these points, that is, x0 is not necessarily the left edge of the box. To get the left edge of the Bbox, use the read-only property xmin.


Old method New method
Axes.get_position() matplotlib.axes.Axes.get_position() [2]
Axes.set_position() matplotlib.axes.Axes.set_position() [3]
Axes.toggle_log_lineary() matplotlib.axes.Axes.set_yscale() [4]
Subplot class removed.

The Polar class has moved to matplotlib.projections.polar.

[2]matplotlib.axes.Axes.get_position() used to return a list of points, now it returns a matplotlib.transforms.Bbox instance.
[3]matplotlib.axes.Axes.set_position() now accepts either four scalars or a matplotlib.transforms.Bbox instance.
[4]Since the recfactoring allows for more than two scale types (‘log’ or ‘linear’), it no longer makes sense to have a toggle. Axes.toggle_log_lineary() has been removed.


Old method New method
Artist.set_clip_path(path) Artist.set_clip_path(path, transform) [5]
[5]matplotlib.artist.Artist.set_clip_path() now accepts a matplotlib.path.Path instance and a matplotlib.transforms.Transform that will be applied to the path immediately before clipping.


Old method New method
linestyle linestyles [6]
[6]Linestyles are now treated like all other collection attributes, i.e. a single value or multiple values may be provided.


Old method New method
ColorConvertor.to_rgba_list(c) ColorConvertor.to_rgba_array(c) [matplotlib.colors.ColorConvertor.to_rgba_array() returns an Nx4 Numpy array of RGBA color quadruples.]


Old method New method
Contour._segments matplotlib.contour.Contour.get_paths`() [Returns a list of matplotlib.path.Path instances.]


Old method New method
Figure.dpi.get() / Figure.dpi.set() matplotlib.figure.Figure.dpi (a property)


Old method New method
Patch.get_verts() matplotlib.patches.Patch.get_path() [Returns a matplotlib.path.Path instance]


Old method New method
GraphicsContext.set_clip_rectangle(tuple) GraphicsContext.set_clip_rectangle(bbox)
GraphicsContext.get_clip_path() GraphicsContext.get_clip_path() [7]
GraphicsContext.set_clip_path() GraphicsContext.set_clip_path() [8]


New methods:

Changed methods:

Removed methods:

  • draw_arc
  • draw_line_collection
  • draw_line
  • draw_lines
  • draw_point
  • draw_quad_mesh
  • draw_poly_collection
  • draw_polygon
  • draw_rectangle
  • draw_regpoly_collection
[7]matplotlib.backend_bases.GraphicsContext.get_clip_path() returns a tuple of the form (path, affine_transform), where path is a matplotlib.path.Path instance and affine_transform is a matplotlib.transforms.Affine2D instance.
[8]matplotlib.backend_bases.GraphicsContext.set_clip_path() now only accepts a matplotlib.transforms.TransformedPath instance.

Changes for 0.91.2

  • For csv2rec(), checkrows=0 is the new default indicating all rows will be checked for type inference
  • A warning is issued when an image is drawn on log-scaled axes, since it will not log-scale the image data.
  • Moved rec2gtk() to matplotlib.toolkits.gtktools
  • Moved rec2excel() to matplotlib.toolkits.exceltools
  • Removed, dead/experimental ExampleInfo, Namespace and Importer code from matplotlib.__init__

Changes for 0.91.1

Changes for 0.91.0

  • Changed cbook.is_file_like() to cbook.is_writable_file_like() and corrected behavior.
  • Added ax kwarg to pyplot.colorbar() and Figure.colorbar() so that one can specify the axes object from which space for the colorbar is to be taken, if one does not want to make the colorbar axes manually.
  • Changed cbook.reversed() so it yields a tuple rather than a (index, tuple). This agrees with the python reversed builtin, and cbook only defines reversed if python doesn’t provide the builtin.
  • Made skiprows=1 the default on csv2rec()
  • The gd and paint backends have been deleted.
  • The errorbar method and function now accept additional kwargs so that upper and lower limits can be indicated by capping the bar with a caret instead of a straight line segment.
  • The matplotlib.dviread file now has a parser for files like and, to map TeX font names to external files.
  • The file matplotlib.type1font contains a new class for Type 1 fonts. Currently it simply reads pfa and pfb format files and stores the data in a way that is suitable for embedding in pdf files. In the future the class might actually parse the font to allow e.g., subsetting.
  • matplotlib.FT2Font now supports FT_Attach_File(). In practice this can be used to read an afm file in addition to a pfa/pfb file, to get metrics and kerning information for a Type 1 font.
  • The AFM class now supports querying CapHeight and stem widths. The get_name_char method now has an isord kwarg like get_width_char.
  • Changed pcolor() default to shading=’flat’; but as noted now in the docstring, it is preferable to simply use the edgecolor kwarg.
  • The mathtext font commands (\cal, \rm, \it, \tt) now behave as TeX does: they are in effect until the next font change command or the end of the grouping. Therefore uses of $\cal{R}$ should be changed to ${\cal R}$. Alternatively, you may use the new LaTeX-style font commands (\mathcal, \mathrm, \mathit, \mathtt) which do affect the following group, e.g., $\mathcal{R}$.
  • Text creation commands have a new default linespacing and a new linespacing kwarg, which is a multiple of the maximum vertical extent of a line of ordinary text. The default is 1.2; linespacing=2 would be like ordinary double spacing, for example.
  • Changed default kwarg in matplotlib.colors.Normalize.__init__`() to clip=False; clipping silently defeats the purpose of the special over, under, and bad values in the colormap, thereby leading to unexpected behavior. The new default should reduce such surprises.
  • Made the emit property of set_xlim() and set_ylim() True by default; removed the Axes custom callback handling into a ‘callbacks’ attribute which is a CallbackRegistry instance. This now supports the ‘xlim_changed’ and ‘ylim_changed’ Axes events.

Changes for 0.90.1

The file has a (very limited and fragile) dvi reader
for usetex support. The API might change in the future so don't
depend on it yet.

Removed deprecated support for a float value as a gray-scale;
now it must be a string, like '0.5'.  Added alpha kwarg to

New method set_bounds(vmin, vmax) for formatters, locators sets
the viewInterval and dataInterval from floats.

Removed deprecated colorbar_classic.

Line2D.get_xdata and get_ydata valid_only=False kwarg is replaced
by orig=True.  When True, it returns the original data, otherwise
the processed data (masked, converted)

Some modifications to the units interface.
units.ConversionInterface.tickers renamed to
units.ConversionInterface.axisinfo and it now returns a
units.AxisInfo object rather than a tuple.  This will make it
easier to add axis info functionality (e.g., I added a default label
on this iteration) w/o having to change the tuple length and hence
the API of the client code every time new functionality is added.
Also, units.ConversionInterface.convert_to_value is now simply
named units.ConversionInterface.convert.

Axes.errorbar uses Axes.vlines and Axes.hlines to draw its error
limits int he vertical and horizontal direction.  As you'll see
in the changes below, these functions now return a LineCollection
rather than a list of lines.  The new return signature for
errorbar is  ylins, caplines, errorcollections where
errorcollections is a xerrcollection, yerrcollection

Axes.vlines and Axes.hlines now create and returns a LineCollection, not a list
of lines.  This is much faster.  The kwarg signature has changed,
so consult the docs

MaxNLocator accepts a new Boolean kwarg ('integer') to force
ticks to integer locations.

Commands that pass an argument to the Text constructor or to
Text.set_text() now accept any object that can be converted
with '%s'.  This affects xlabel(), title(), etc.

Barh now takes a **kwargs dict instead of most of the old
arguments. This helps ensure that bar and barh are kept in sync,
but as a side effect you can no longer pass e.g., color as a
positional argument.

ft2font.get_charmap() now returns a dict that maps character codes
to glyph indices (until now it was reversed)

Moved data files into lib/matplotlib so that setuptools' develop
mode works. Re-organized the mpl-data layout so that this source
structure is maintained in the installation. (i.e., the 'fonts' and
'images' sub-directories are maintained in site-packages.).
Suggest removing site-packages/matplotlib/mpl-data and
~/.matplotlib/ttffont.cache before installing

Changes for 0.90.0

All artists now implement a "pick" method which users should not
call.  Rather, set the "picker" property of any artist you want to
pick on (the epsilon distance in points for a hit test) and
register with the "pick_event" callback.  See
examples/ for details

Bar, barh, and hist have "log" binary kwarg: log=True
sets the ordinate to a log scale.

Boxplot can handle a list of vectors instead of just
an array, so vectors can have different lengths.

Plot can handle 2-D x and/or y; it plots the columns.

Added linewidth kwarg to bar and barh.

Made the default Artist._transform None (rather than invoking
identity_transform for each artist only to have it overridden
later).  Use artist.get_transform() rather than artist._transform,
even in derived classes, so that the default transform will be
created lazily as needed

New LogNorm subclass of Normalize added to
All Normalize subclasses have new inverse() method, and
the __call__() method has a new clip kwarg.

Changed class names in to match convention:
normalize -> Normalize, no_norm -> NoNorm.  Old names
are still available for now.

Removed obsolete pcolor_classic command and method.

Removed lineprops and markerprops from the Annotation code and
replaced them with an arrow configurable with kwarg arrowprops.
See examples/ - JDH

Changes for 0.87.7

Completely reworked the annotations API because I found the old
API cumbersome.  The new design is much more legible and easy to
read.  See matplotlib.text.Annotation and

markeredgecolor and markerfacecolor cannot be configured in
matplotlibrc any more. Instead, markers are generally colored
automatically based on the color of the line, unless marker colors
are explicitly set as kwargs - NN

Changed default comment character for load to '#' - JDH

math_parse_s_ft2font_svg from & now returns
width, height, svg_elements. svg_elements is an instance of Bunch ( and has the attributes svg_glyphs and svg_lines, which are both

Renderer.draw_arc now takes an additional parameter, rotation.
It specifies to draw the artist rotated in degrees anti-
clockwise.  It was added for rotated ellipses.

Renamed Figure.set_figsize_inches to Figure.set_size_inches to
better match the get method, Figure.get_size_inches.

Removed the copy_bbox_transform from; added
shallowcopy methods to all transforms.  All transforms already
had deepcopy methods.

FigureManager.resize(width, height): resize the window
specified in pixels

barh: x and y args have been renamed to width and bottom
respectively, and their order has been swapped to maintain
a (position, value) order.

bar and barh: now accept kwarg 'edgecolor'.

bar and barh: The left, height, width and bottom args can
now all be scalars or sequences; see docstring.

barh: now defaults to edge aligned instead of center
aligned bars

bar, barh and hist: Added a keyword arg 'align' that
controls between edge or center bar alignment.

Collections: PolyCollection and LineCollection now accept
vertices or segments either in the original form [(x,y),
(x,y), ...] or as a 2D numerix array, with X as the first column
and Y as the second. Contour and quiver output the numerix
form.  The transforms methods Bbox.update() and
Transformation.seq_xy_tups() now accept either form.

Collections: LineCollection is now a ScalarMappable like
PolyCollection, etc.

Specifying a grayscale color as a float is deprecated; use
a string instead, e.g., 0.75 -> '0.75'.

Collections: initializers now accept any mpl color arg, or
sequence of such args; previously only a sequence of rgba
tuples was accepted.

Colorbar: completely new version and api; see docstring.  The
original version is still accessible as colorbar_classic, but
is deprecated.

Contourf: "extend" kwarg replaces "clip_ends"; see docstring.
Masked array support added to pcolormesh.

Modified aspect-ratio handling:
    Removed aspect kwarg from imshow
    Axes methods:
        set_aspect(self, aspect, adjustable=None, anchor=None)
        set_adjustable(self, adjustable)
        set_anchor(self, anchor)
    Pylab interface:

 Backend developers: ft2font's load_char now takes a flags
 argument, which you can OR together from the LOAD_XXX

Changes for 0.86

Matplotlib data is installed into the matplotlib module.
This is similar to package_data.  This should get rid of
having to check for many possibilities in _get_data_path().
The MATPLOTLIBDATA env key is still checked first to allow
for flexibility.

1) Separated the color table data from out into
a new file,, to make it easier to find the actual
code in and to add new colormaps. Everything
from is imported by, so the split should be
2) Enabled automatic generation of a colormap from
a list of colors in contour; see modified
3) Support for imshow of a masked array, with the
ability to specify colors (or no color at all) for
masked regions, and for regions that are above or
below the normally mapped region.  See
4) In support of the above, added two new classes,
ListedColormap, and no_norm, to, and modified
the Colormap class to include common functionality. Added
a clip kwarg to the normalize class.

Changes for 0.85

Made xtick and ytick separate props in rc

made pos=None the default for tick formatters rather than 0 to
indicate "not supplied"

Removed "feature" of minor ticks which prevents them from
overlapping major ticks.  Often you want major and minor ticks at
the same place, and can offset the major ticks with the pad.  This
could be made configurable

Changed the internal structure of to a more OO style.
Calls to contour or contourf in or now return
a ContourSet object which contains references to the
LineCollections or PolyCollections created by the call,
as well as the configuration variables that were used.
The ContourSet object is a "mappable" if a colormap was used.

Added a clip_ends kwarg to contourf. From the docstring:
         * clip_ends = True
           If False, the limits for color scaling are set to the
           minimum and maximum contour levels.
           True (default) clips the scaling limits.  Example:
           if the contour boundaries are V = [-100, 2, 1, 0, 1, 2, 100],
           then the scaling limits will be [-100, 100] if clip_ends
           is False, and [-3, 3] if clip_ends is True.
Added kwargs linewidths, antialiased, and nchunk to contourf.  These
are experimental; see the docstring.

Changed Figure.colorbar():
    kw argument order changed;
    if mappable arg is a non-filled ContourSet, colorbar() shows
            lines instead hof polygons.
    if mappable arg is a filled ContourSet with clip_ends=True,
            the endpoints are not labelled, so as to give the
            correct impression of open-endedness.

Changed LineCollection.get_linewidths to get_linewidth, for

Changes for 0.84

Unified argument handling between hlines and vlines.  Both now
take optionally a fmt argument (as in plot) and a keyword args
that can be passed onto Line2D.

Removed all references to "data clipping" in rc and since
these were not used and not optimized.  I'm sure they'll be
resurrected later with a better implementation when needed.

'set' removed - no more deprecation warnings.  Use 'setp' instead.

Backend developers: Added flipud method to image and removed it
from to_str.  Removed origin kwarg from backend.draw_image.
origin is handled entirely by the frontend now.

Changes for 0.83

- Made HOME/.matplotlib the new config dir where the matplotlibrc
  file, the ttf.cache, and the tex.cache live.  The new default
  filenames in .matplotlib have no leading dot and are not hidden.
  e.g., the new names are matplotlibrc, tex.cache, and ttffont.cache.
  This is how ipython does it so it must be right.

  If old files are found, a warning is issued and they are moved to
  the new location.

- backends/ no longer imports new_figure_manager,
  draw_if_interactive and show from the default backend, but puts
  these imports into a call to pylab_setup.  Also, the Toolbar is no
  longer imported from WX/WXAgg.  New usage:

    from backends import pylab_setup
    new_figure_manager, draw_if_interactive, show = pylab_setup()

- Moved Figure.get_width_height() to FigureCanvasBase. It now
  returns int instead of float.

Changes for 0.82

- toolbar import change in GTKAgg, GTKCairo and WXAgg

- Added subplot config tool to GTK* backends -- note you must now
  import the NavigationToolbar2 from your backend of choice rather
  than from backend_gtk because it needs to know about the backend
  specific canvas -- see examples/  Ditto for
  wx backend -- see examples/

- hist bin change

    Sean Richards notes there was a problem in the way we created
    the binning for histogram, which made the last bin
    underrepresented.  From his post:

      I see that hist uses the linspace function to create the bins
      and then uses searchsorted to put the values in their correct
      bin. That's all good but I am confused over the use of linspace
      for the bin creation. I wouldn't have thought that it does
      what is needed, to quote the docstring it creates a "Linear
      spaced array from min to max". For it to work correctly
      shouldn't the values in the bins array be the same bound for
      each bin? (i.e. each value should be the lower bound of a
      bin). To provide the correct bins for hist would it not be
      something like

      def bins(xmin, xmax, N):
        if N==1: return xmax
        dx = (xmax-xmin)/N # instead of N-1
        return xmin + dx*arange(N)

     This suggestion is implemented in 0.81.  My test script with these
     changes does not reveal any bias in the binning

      from matplotlib.numerix.mlab import randn, rand, zeros, Float
      from matplotlib.mlab import hist, mean

      Nbins = 50
      Ntests = 200
      results = zeros((Ntests,Nbins), typecode=Float)
      for i in range(Ntests):
          print 'computing', i
          x = rand(10000)
          n, bins = hist(x, Nbins)
          results[i] = n
      print mean(results)

Changes for 0.81

- pylab and artist "set" functions renamed to setp to avoid clash
  with python2.4 built-in set.  Current version will issue a
  deprecation warning which will be removed in future versions

- imshow interpolation arguments changes for advanced interpolation
  schemes.  See help imshow, particularly the interpolation,
  filternorm and filterrad kwargs

- Support for masked arrays has been added to the plot command and
  to the Line2D object.  Only the valid points are plotted.  A
  "valid_only" kwarg was added to the get_xdata() and get_ydata()
  methods of Line2D; by default it is False, so that the original
  data arrays are returned. Setting it to True returns the plottable

- contour changes:

  Masked arrays: contour and contourf now accept masked arrays as
    the variable to be contoured.  Masking works correctly for
    contour, but a bug remains to be fixed before it will work for
    contourf.  The "badmask" kwarg has been removed from both

   Level argument changes:

     Old version: a list of levels as one of the positional
     arguments specified the lower bound of each filled region; the
     upper bound of the last region was taken as a very large
     number.  Hence, it was not possible to specify that z values
     between 0 and 1, for example, be filled, and that values
     outside that range remain unfilled.

     New version: a list of N levels is taken as specifying the
     boundaries of N-1 z ranges.  Now the user has more control over
     what is colored and what is not.  Repeated calls to contourf
     (with different colormaps or color specifications, for example)
     can be used to color different ranges of z.  Values of z
     outside an expected range are left uncolored.

       Old: contourf(z, [0, 1, 2]) would yield 3 regions: 0-1, 1-2, and >2.
       New: it would yield 2 regions: 0-1, 1-2.  If the same 3 regions were
       desired, the equivalent list of levels would be [0, 1, 2,

Changes for 0.80

- xlim/ylim/axis always return the new limits regardless of
  arguments.  They now take kwargs which allow you to selectively
  change the upper or lower limits while leaving unnamed limits
  unchanged.  See help(xlim) for example

Changes for 0.73

- Removed deprecated ColormapJet and friends

- Removed all error handling from the verbose object

- figure num of zero is now allowed

Changes for 0.72

- Line2D, Text, and Patch copy_properties renamed update_from and
  moved into artist base class

- LineCollecitons.color renamed to LineCollections.set_color for
  consistency with set/get introspection mechanism,

- pylab figure now defaults to num=None, which creates a new figure
  with a guaranteed unique number

- contour method syntax changed - now it is MATLAB compatible

    unchanged: contour(Z)
    old: contour(Z, x=Y, y=Y)
    new: contour(X, Y, Z)


 - Increased the default resolution for save command.

 - Renamed the base attribute of the ticker classes to _base to avoid conflict
   with the base method.  Sitt for subs

 - subs=none now does autosubbing in the tick locator.

 - New subplots that overlap old will delete the old axes.  If you
   do not want this behavior, use fig.add_subplot or the axes

Changes for 0.71

Significant numerix namespace changes, introduced to resolve
namespace clashes between python built-ins and mlab names.
Refactored numerix to maintain separate modules, rather than
folding all these names into a single namespace.  See the following
mailing list threads for more information and background

OLD usage

  from matplotlib.numerix import array, mean, fft

NEW usage

  from matplotlib.numerix import array
  from matplotlib.numerix.mlab import mean
  from matplotlib.numerix.fft import fft

numerix dir structure mirrors numarray (though it is an incomplete


but of course you can use 'numerix : Numeric' and still get the

pylab still imports most of the symbols from Numerix, MLab, fft,
etc, but is more cautious.  For names that clash with python names
(min, max, sum), pylab keeps the builtins and provides the numeric
versions with an a* prefix, e.g., (amin, amax, asum)

Changes for 0.70

MplEvent factored into a base class Event and derived classes
MouseEvent and KeyEvent

Removed definct set_measurement in wx toolbar

Changes for 0.65.1

removed add_axes and add_subplot from backend_bases.  Use
figure.add_axes and add_subplot instead.  The figure now manages the
current axes with gca and sca for get and set current axes.  If you
have code you are porting which called, e.g., figmanager.add_axes, you
can now simply do figmanager.canvas.figure.add_axes.

Changes for 0.65

mpl_connect and mpl_disconnect in the MATLAB interface renamed to
connect and disconnect

Did away with the text methods for angle since they were ambiguous.
fontangle could mean fontstyle (obligue, etc) or the rotation of the
text.  Use style and rotation instead.

Changes for 0.63

Dates are now represented internally as float days since 0001-01-01,

All date tickers and formatters are now in matplotlib.dates, rather
than matplotlib.tickers

converters have been abolished from all functions and classes.
num2date and date2num are now the converter functions for all date

Most of the date tick locators have a different meaning in their
constructors.  In the prior implementation, the first argument was a
base and multiples of the base were ticked.  e.g.,

  HourLocator(5)  # old: tick every 5 minutes

In the new implementation, the explicit points you want to tick are
provided as a number or sequence

   HourLocator(range(0,5,61))  # new: tick every 5 minutes

This gives much greater flexibility.  I have tried to make the
default constructors (no args) behave similarly, where possible.

Note that YearLocator still works under the base/multiple scheme.
The difference between the YearLocator and the other locators is
that years are not recurrent.

Financial functions:, date1, date2)

   date1, date2 are now datetime instances.  Return value is a list
   of quotes where the quote time is a float - days since gregorian
   start, as returned by date2num

   See examples/ for example usage of new API

Changes for 0.61

canvas.connect is now deprecated for event handling.  use
mpl_connect and mpl_disconnect instead.  The callback signature is
func(event) rather than func(widget, event)

Changes for 0.60

ColormapJet and Grayscale are deprecated.  For backwards
compatibility, they can be obtained either by doing

  from import ColormapJet


  from matplotlib.matlab import *

They are replaced by cm.jet and cm.grey

Changes for 0.54.3

removed the set_default_font / get_default_font scheme from the
font_manager to unify customization of font defaults with the rest of
the rc scheme.  See examples/ and help(rc) in

Changes for 0.54

MATLAB interface


Several of the backends used a PIXELS_PER_INCH hack that I added to try and make images render consistently across backends. This just complicated matters. So you may find that some font sizes and line widths appear different than before. Apologies for the inconvenience. You should set the dpi to an accurate value for your screen to get true sizes.

pcolor and scatter

There are two changes to the MATLAB interface API, both involving the patch drawing commands. For efficiency, pcolor and scatter have been rewritten to use polygon collections, which are a new set of objects from matplotlib.collections designed to enable efficient handling of large collections of objects. These new collections make it possible to build large scatter plots or pcolor plots with no loops at the python level, and are significantly faster than their predecessors. The original pcolor and scatter functions are retained as pcolor_classic and scatter_classic.

The return value from pcolor is a PolyCollection. Most of the propertes that are available on rectangles or other patches are also available on PolyCollections, e.g., you can say:

c = scatter(blah, blah)


c = scatter(blah, blah)
set(c, 'linewidth', 1.0, 'facecolor', 'r', 'alpha', 0.5)

Because the collection is a single object, you no longer need to loop over the return value of scatter or pcolor to set properties for the entire list.

If you want the different elements of a collection to vary on a property, e.g., to have different line widths, see matplotlib.collections for a discussion on how to set the properties as a sequence.

For scatter, the size argument is now in points^2 (the area of the symbol in points) as in MATLAB and is not in data coords as before. Using sizes in data coords caused several problems. So you will need to adjust your size arguments accordingly or use scatter_classic.

mathtext spacing

For reasons not clear to me (and which I’ll eventually fix) spacing no longer works in font groups. However, I added three new spacing commands which compensate for this ‘’ (regular space), ‘/’ (small space) and ‘hspace{frac}’ where frac is a fraction of fontsize in points. You will need to quote spaces in font strings, is:

title(r'$\rm{Histogram\ of\ IQ:}\ \mu=100,\ \sigma=15$')

Object interface - Application programmers


The x and y axis instances no longer have autoscale view. These are handled by axes.autoscale_view

Axes creation

You should not instantiate your own Axes any more using the OO API. Rather, create a Figure as before and in place of:

f = Figure(figsize=(5,4), dpi=100)
a = Subplot(f, 111)


f = Figure(figsize=(5,4), dpi=100)
a = f.add_subplot(111)

That is, add_axis no longer exists and is replaced by:

add_axes(rect, axisbg=defaultcolor, frameon=True)
add_subplot(num, axisbg=defaultcolor, frameon=True)

Artist methods

If you define your own Artists, you need to rename the _draw method to draw

Bounding boxes

matplotlib.transforms.Bound2D is replaced by matplotlib.transforms.Bbox. If you want to construct a bbox from left, bottom, width, height (the signature for Bound2D), use matplotlib.transforms.lbwh_to_bbox, as in

bbox = clickBBox = lbwh_to_bbox(left, bottom, width, height)

The Bbox has a different API than the Bound2D. e.g., if you want to get the width and height of the bbox

width = fig.bbox.x.interval() height = fig.bbox.y.interval()
width = fig.bbox.width() height = fig.bbox.height()

Object constructors

You no longer pass the bbox, dpi, or transforms to the various Artist constructors. The old way or creating lines and rectangles was cumbersome because you had to pass so many attributes to the Line2D and Rectangle classes not related directly to the geometry and properties of the object. Now default values are added to the object when you call axes.add_line or axes.add_patch, so they are hidden from the user.

If you want to define a custom transformation on these objects, call o.set_transform(trans) where trans is a Transformation instance.

In prior versions of you wanted to add a custom line in data coords, you would have to do

l = Line2D(dpi, bbox, x, y,
color = color, transx = transx, transy = transy, )

now all you need is

l = Line2D(x, y, color=color)

and the axes will set the transformation for you (unless you have set your own already, in which case it will eave it unchanged)


The entire transformation architecture has been rewritten. Previously the x and y transformations where stored in the xaxis and yaxis instances. The problem with this approach is it only allows for separable transforms (where the x and y transformations don’t depend on one another). But for cases like polar, they do. Now transformations operate on x,y together. There is a new base class matplotlib.transforms.Transformation and two concrete implementations, matplotlib.transforms.SeparableTransformation and matplotlib.transforms.Affine. The SeparableTransformation is constructed with the bounding box of the input (this determines the rectangular coordinate system of the input, i.e., the x and y view limits), the bounding box of the display, and possibly nonlinear transformations of x and y. The 2 most frequently used transformations, data coordinates -> display and axes coordinates -> display are available as ax.transData and ax.transAxes. See which uses axes coords.

Also, the transformations should be much faster now, for two reasons

  • they are written entirely in extension code

  • because they operate on x and y together, they can do the entire transformation in one loop. Earlier I did something along the lines of:

    xt = sx*func(x) + tx
    yt = sy*func(y) + ty

    Although this was done in numerix, it still involves 6 length(x) for-loops (the multiply, add, and function evaluation each for x and y). Now all of that is done in a single pass.

If you are using transformations and bounding boxes to get the cursor position in data coordinates, the method calls are a little different now. See the updated examples/ which shows you how to do this.

Likewise, if you are using the artist bounding boxes to pick items on the canvas with the GUI, the bbox methods are somewhat different. You will need to see the updated examples/

See unit/ for many examples using the new transformations.

Changes for 0.50

* refactored Figure class so it is no longer backend dependent.
  FigureCanvasBackend takes over the backend specific duties of the
  Figure.  matplotlib.backend_bases.FigureBase moved to

* backends must implement FigureCanvasBackend (the thing that
  controls the figure and handles the events if any) and
  FigureManagerBackend (wraps the canvas and the window for MATLAB
  interface).  FigureCanvasBase implements a backend switching

* Figure is now an Artist (like everything else in the figure) and
  is totally backend independent


* backend faceColor argument changed to rgbFace

* colormap stuff moved to

* arg_to_rgb in backend_bases moved to class ColorConverter in

* GD users must upgrade to gd-2.0.22 and gdmodule-0.52 since new gd
  features (clipping, antialiased lines) are now used.

* Renderer must implement points_to_pixels

Migrating code:

MATLAB interface:

  The only API change for those using the MATLAB interface is in how
  you call figure redraws for dynamically updating figures.  In the
  old API, you did


  In the new API, you do

    manager = get_current_fig_manager()

  See the examples,, and


  There is one important API change for application developers.
  Figure instances used subclass GUI widgets that enabled them to be
  placed directly into figures.  e.g., FigureGTK subclassed
  gtk.DrawingArea.  Now the Figure class is independent of the
  backend, and FigureCanvas takes over the functionality formerly
  handled by Figure.  In order to include figures into your apps,
  you now need to do, for example

    # gtk example
    fig = Figure(figsize=(5,4), dpi=100)
    canvas = FigureCanvasGTK(fig)  # a gtk.DrawingArea

  If you use the NavigationToolbar, this in now initialized with a
  FigureCanvas, not a Figure.  The examples,, and all reflect the new
  API so use these as a guide.

  All prior calls to

   figure.draw()  and

  should now be

   canvas.draw()  and

  Apologies for the inconvenience.  This refactorization brings
  significant more freedom in developing matplotlib and should bring
  better plotting capabilities, so I hope the inconvenience is worth

Changes for 0.42

* Refactoring AxisText to be backend independent.  Text drawing and
  get_window_extent functionality will be moved to the Renderer.

* backend_bases.AxisTextBase is now text.Text module

* All the erase and reset functionality removed from AxisText - not
  needed with double buffered drawing.  Ditto with state change.
  Text instances have a get_prop_tup method that returns a hashable
  tuple of text properties which you can use to see if text props
  have changed, e.g., by caching a font or layout instance in a dict
  with the prop tup as a key -- see RendererGTK.get_pango_layout in
  backend_gtk for an example.

* Text._get_xy_display renamed Text.get_xy_display

* Artist set_renderer and wash_brushes methods removed

* Moved Legend class from matplotlib.axes into matplotlib.legend

* Moved Tick, XTick, YTick, Axis, XAxis, YAxis from matplotlib.axes
  to matplotlib.axis

* moved process_text_args to matplotlib.text

* After getting Text handled in a backend independent fashion, the
  import process is much cleaner since there are no longer cyclic

* matplotlib.matlab._get_current_fig_manager renamed to
  matplotlib.matlab.get_current_fig_manager to allow user access to
  the GUI window attribute, e.g., figManager.window for GTK and
  figManager.frame for wx

Changes for 0.40

- Artist
    * __init__ takes a DPI instance and a Bound2D instance which is
      the bounding box of the artist in display coords
    * get_window_extent returns a Bound2D instance
    * set_size is removed; replaced by bbox and dpi
    * the clip_gc method is removed.  Artists now clip themselves with
      their box
    * added _clipOn boolean attribute.  If True, gc clip to bbox.

- AxisTextBase
    * Initialized with a transx, transy which are Transform instances
    * set_drawing_area removed
    * get_left_right and get_top_bottom are replaced by get_window_extent

- Line2D Patches now take transx, transy
    * Initialized with a transx, transy which are Transform instances

- Patches
   * Initialized with a transx, transy which are Transform instances

- FigureBase attributes dpi is a DPI instance rather than scalar and
  new attribute bbox is a Bound2D in display coords, and I got rid
  of the left, width, height, etc... attributes.  These are now
  accessible as, for example, bbox.x.min is left, bbox.x.interval()
  is width, bbox.y.max is top, etc...

- GcfBase attribute pagesize renamed to figsize

- Axes
    * removed figbg attribute
    * added fig instance to __init__
    * resizing is handled by figure call to resize.

- Subplot
    * added fig instance to __init__

- Renderer methods for patches now take gcEdge and gcFace instances.
  gcFace=None takes the place of filled=False

- True and False symbols provided by cbook in a python2.3 compatible

- new module transforms supplies Bound1D, Bound2D and Transform
  instances and more

- Changes to the MATLAB helpers API

  * _matlab_helpers.GcfBase is renamed by Gcf.  Backends no longer
    need to derive from this class.  Instead, they provide a factory
    function new_figure_manager(num, figsize, dpi).  The destroy
    method of the GcfDerived from the backends is moved to the derived

  * FigureManagerBase moved to backend_bases

  * Gcf.get_all_figwins renamed to Gcf.get_all_fig_managers


  Make sure to self._reset = False in AxisTextWX._set_font.  This was
  something missing in my backend code.