While updating the BACKEND_SPECS in #5135 , I noticed that matplotlib 3.9 introduced a matplotlib.backends.backend_registry that could be used to replace BACKEND_SPECS entirely and offload the backend details to matplotlib.
e.g., could add the following to get the FigureCanvas and FigureManager classes in _get_canvas_classes
if matplotlib.__version_info__ >= (3, 9):
from matplotlib.backends import backend_registry
mod = backend_registry.load_backend_module(key)
FigureCanvas = getattr(mod, 'FigureCanvas')
FigureManager = getattr(mod, 'FigureManager')Only thing I'm a little uncertain of is the FigureManager -- from what I understand, the above code pulls out FigureCanvas and FigureManager classes that are aliased to the specific implementation for each backend and there should always be a FigureManager specified (which seems to have been the case for quite a while now... since matplotlib/matplotlib#8773). Any reason to preserve yt's behavior for the backends where None is specified for a FigureManager in BACKEND_SPECS? I think no, but wanted to get a second opinion.
here's the relevant bits of code i'm referring to:
|
BACKEND_SPECS = { |
|
"GTK": ["backend_gtk", "FigureCanvasGTK", "FigureManagerGTK"], |
|
"GTKAgg": ["backend_gtkagg", "FigureCanvasGTKAgg", None], |
|
"GTKCairo": ["backend_gtkcairo", "FigureCanvasGTKCairo", None], |
|
"MacOSX": ["backend_macosx", "FigureCanvasMac", "FigureManagerMac"], |
|
"Qt5Agg": ["backend_qt5agg", "FigureCanvasQTAgg", None], |
|
"QtAgg": ["backend_qtagg", "FigureCanvasQTAgg", None], |
|
"TkAgg": ["backend_tkagg", "FigureCanvasTkAgg", None], |
|
"WX": ["backend_wx", "FigureCanvasWx", None], |
|
"WXAgg": ["backend_wxagg", "FigureCanvasWxAgg", None], |
|
"GTK3Cairo": [ |
|
"backend_gtk3cairo", |
|
"FigureCanvasGTK3Cairo", |
|
"FigureManagerGTK3Cairo", |
|
], |
|
"GTK3Agg": ["backend_gtk3agg", "FigureCanvasGTK3Agg", "FigureManagerGTK3Agg"], |
|
"WebAgg": ["backend_webagg", "FigureCanvasWebAgg", None], |
|
"nbAgg": ["backend_nbagg", "FigureCanvasNbAgg", "FigureManagerNbAgg"], |
|
"agg": ["backend_agg", "FigureCanvasAgg", None], |
|
} |
and
|
def _get_canvas_classes(self): |
|
if self.interactivity: |
|
key = str(matplotlib.get_backend()) |
|
else: |
|
key = "agg" |
|
|
|
module, fig_canvas, fig_manager = BACKEND_SPECS[key] |
|
|
|
mod = __import__( |
|
"matplotlib.backends", |
|
globals(), |
|
locals(), |
|
[module], |
|
0, |
|
) |
|
submod = getattr(mod, module) |
|
FigureCanvas = getattr(submod, fig_canvas) |
|
if fig_manager is not None: |
|
FigureManager = getattr(submod, fig_manager) |
|
return FigureCanvas, FigureManager |
|
|
|
return FigureCanvas, None |
Also, I don't see a huge benefit to putting this change in now vs waiting til yt's min matplotlib version is 3.9, but wanted to get these thoughts down while they're fresh.
While updating the
BACKEND_SPECSin #5135 , I noticed that matplotlib 3.9 introduced amatplotlib.backends.backend_registrythat could be used to replaceBACKEND_SPECSentirely and offload the backend details to matplotlib.e.g., could add the following to get the
FigureCanvasandFigureManagerclasses in_get_canvas_classesOnly thing I'm a little uncertain of is the FigureManager -- from what I understand, the above code pulls out
FigureCanvasandFigureManagerclasses that are aliased to the specific implementation for each backend and there should always be aFigureManagerspecified (which seems to have been the case for quite a while now... since matplotlib/matplotlib#8773). Any reason to preserve yt's behavior for the backends whereNoneis specified for aFigureManagerinBACKEND_SPECS? I think no, but wanted to get a second opinion.here's the relevant bits of code i'm referring to:
yt/yt/visualization/base_plot_types.py
Lines 47 to 66 in 5ea7043
and
yt/yt/visualization/base_plot_types.py
Lines 152 to 173 in 5ea7043
Also, I don't see a huge benefit to putting this change in now vs waiting til yt's min matplotlib version is 3.9, but wanted to get these thoughts down while they're fresh.