Main PR (including GTK3): + https://github.com/matplotlib/matplotlib/pull/4143
Backend specific branch diffs: + https://github.com/OceanWolf/matplotlib/compare/backend-refactor...OceanWolf:backend-refactor-tkagg + https://github.com/OceanWolf/matplotlib/compare/backend-refactor...OceanWolf:backend-refactor-qt + https://github.com/OceanWolf/matplotlib/compare/backend-refactor...backend-refactor-wx
This MEP refactors the backends to give a more structured and consistent API, removing generic code and consolidate existing code. To do this we propose splitting:
FigureManagerBaseand its derived classes into the core functionality class
FigureManagerand a backend specific class
ShowBaseand its derived classes into
This MEP aims to consolidate the backends API into one single uniform
API, removing generic code out of the backend (which includes
Gcf), and push code to a more appropriate
level in matplotlib. With this we automatically remove
inconsistencies that appear in the backends, such as
FigureManagerBase.resize(w, h) which sometimes sets the canvas,
and other times set the entire window to the dimensions given,
depending on the backend.
Two main places for generic code appear in the classes derived from
FigureManagerBase has three jobs at the moment:
- The documentation describes it as a ``Helper class for pyplot mode, wraps everything up into a neat bundle’‘
- But it doesn’t just wrap the canvas and toolbar, it also does all of the windowing tasks itself. The conflation of these two tasks gets seen the best in the following line:
`python self.set_window_title("Figure %d" % num) `This combines backend specific code
self.set_window_title(title)with matplotlib generic code
title = "Figure %d" % num.
- Currently the backend specific subclass of
FigureManagerdecides when to end the mainloop. This also seems very wrong as the figure should have no control over the other figures.
ShowBase has two jobs:
- It has the job of going through all figure managers registered in
_pylab_helpers.Gcfand telling them to show themselves.
- And secondly it has the job of performing the backend specific
mainloopto block the main programme and thus keep the figures from dying.
The description of this MEP gives us most of the solution:
FigureManagerBaseletting it simply wrap this new class along with the other backend classes. Create a new
WindowBaseclass that can handle this functionality, with pass-through methods (:arrow_right:) to
WindowBase. Classes that subclass
WindowBaseshould also subclass the GUI specific window class to ensure backward compatibility (
manager.window == manager.window).
MainLoopBase, which encapsulates the end of the loop as well. We give an instance of
FigureManageras a key unlock the exit method (requiring all keys returned before the loop can die). Note this opens the possibility for multiple backends to run concurrently.
FigureManagerBasehas no backend specifics in it, to rename it to
FigureManager, and move to a new file
FigureManagerBaseclass and its dependencies intact.
NavigationBasehas morphed into a backend independent
|FigureManagerBase(canvas, num)||FigureManager(figure, num)||
|destroy||calls destroy on all components||destroy|
|show_popup||show_poup||Not used anywhere in mpl, and does nothing.|
|_get_toolbar||A common method to all subclasses of FigureManagerBase|
|end||Gets called automagically when no more instances of the subclass exist|
|__call__||Method moved to Gcf.show_all|
As eluded to above when discussing MEP 22, this refactor makes it easy
to add in new generic features. At the moment, MEP 22 has to make
ugly hacks to each class extending from
this code, this only needs to get made in the single
class. This also makes the later deprecation of
NavigationToolbar2 very straightforward, only needing to touch the
MEP 23 makes for another use case where this refactored code will come in very handy.
As we leave all backend code intact, only adding missing methods to
existing classes, this should work seamlessly for all use cases. The
only difference will lie for backends that used
FigureManager.resize to resize the canvas and not the window, due
to the standardisation of the API.
I would envision that the classes made obsolete by this refactor get
deprecated and removed on the same timetable as
NavigationToolbar2, also note that the change in call signature to
FigureCanvasWx constructor, while backward compatible, I think
the old (imho ugly style) signature should get deprecated and removed
in the same manner as everything else.
If there were any alternative solutions to solving the same problem, they should be discussed here, along with a justification for the chosen approach.
Mdehoon: Can you elaborate on how to run multiple backends concurrently?
OceanWolf: @mdehoon, as I say, not for this MEP, but I see this MEP
opens it up as a future possibility. Basically the
class acts a per backend Gcf, in this MEP it tracks the number of
figures open per backend, and manages the mainloops for those
backends. It closes the backend specific mainloop when it detects
that no figures remain open for that backend. Because of this I
imagine that with only a small amount of tweaking that we can do
full-multi-backend matplotlib. No idea yet why one would want to, but
I leave the possibility there in MainLoopBase. With all the
backend-code specifics refactored out of
FigureManager also aids
in this, one manager to rule them (the backends) all.
Mdehoon: @OceanWolf, OK, thanks for the explanation. Having a uniform API for the backends is very important for the maintainability of matplotlib. I think this MEP is a step in the right direction.