Matplotlib 3.8.4 update: restore WASM backend and partially restore HTML5 Canvas backend

[agriyakhetarpal]

This PR has been put together towards a resolution for pyodide/pyodide#4510, which aims to update the Matplotlib version for Pyodide in-tree from 3.5.2 to 3.8.4, and is the next step towards releasing matplotlib-pyodide version 0.2.3 – please see #60 for tracking the release. It restores the WASM backend to work with Matplotlib 3.8.4 and partially restores the HTML5 Canvas backend—an attempt was made to get it to work, but I did not see through it entirely—by internally redirecting it to use the WASM backend for rendering.

A complete restoration of the HTML5 backend or a reorganisation of the backends to a maintained or working state can be looked at later with a new matplotlib-pyodide release and a corresponding Pyodide release, and contributions towards this are very much welcome. Please see #65 for further discussions and additional context.

Specifically, the changes here are as follows:

• build-backend related files (setuptools-specific) have been git-ignored
• the event listener for the mouse movements has been silenced for now – it creates on the order of several hundred warnings in the console otherwise, which isn't good for UX and for those trying to debug things when working with Matplotlib
• the WASM backend works with Matplotlib 3.8.4
• there has been an attempt to update the HTML5 backend to use a path-based MathTextParser, since the bitmap-based one was deprecated in Matplotlib 3.4 and removed around 3.6. It is able to create and parse glyphs and draw them on the screen, but fonts are currently missing when drawn.
• noting the above and noting the HTML5 Canvas backend's popularity, we have redirected the HTML5 backend to also use the WASM backend internally, so that users don't need to change anything for their existing apps or lecture material.

[ryanking13]

Thanks!

[ryanking13]

Looks like matplotlib is complaining about the type of capstyle. Using enums may work.

from matplotlib._enums import CapStyle

self._capstyle = CapStyle(...)

[ryanking13]

Looks good. Thanks!

[agriyakhetarpal]

Still not fully fixed yet!

[agriyakhetarpal]

Bumping to Chrome 131 here did not fix the previous issues, so I guess we'll tackle it later, it's not strictly necessary. I'll patch Matplotlib once again with these changes in pyodide/pyodide#4510 and see if everything is working – if yes, we can merge this safely.

[agriyakhetarpal]

Thanks for the review, @ryanking13! I'd like to keep the HTML5 backend updates I made here and not completely remove them so that either I or someone else can return to them at some point and fill in the missing piece that is required to make them work. Since you've already approved these changes in the PR, I shall merge it now – but if you feel otherwise about anything, please let me know!

[hoodmane]

Thanks so much for working on this @agriyakhetarpal!

[jsclass42]

First thank you all very much for such an amazing project!

One thing I would like to bring to your attention is the change introduced at 0.27.0 of backend approach really breaks backend like FigureCanvasSVG. This is what we get when we try: "UserWarning: FigureCanvasSVG is non-interactive, and thus cannot be shown". Prior to this update, to implement/extend a backend was simple and versatile, which basically override FigureManager show() method, say, to simply print out some information.

Can we at least have a mechanism to make this non-interactive optional, or allow a custom backend implementation to return a simple flag to comply with whatever requirement (in this case, pretend to be interactive, thus show() will be called upon).

Currently, we are unable to upgrade to 0.27.0 because of this.

Again, thanks!

[ryanking13]

Thanks for the report @jsclass42, sadly, we are lacking human resources to manage two different backends and we are trying to find a way to support matplotlib-pyodide in a long term (#65). Meanwhile, you can try to patch matplotlib-pyodide, make your own backend and accomplish what you need.

[agriyakhetarpal]

Yes, agree with Gyeongjae – I would be grateful if you could fully restore the previous behaviour and find out what's missing for the backend(s) to work as they did earlier. If we're able to, we can easily get a new release here and include it with Pyodide 0.27.1/0.27.2.

[jsclass42]

Thank you @ryanking13 for your reply and suggestion! I decided to look into the codes, including the matplotlib/backend_bases.py, and it turns out the break on my end had nothing to do with pyodide or matplotlib-pyodide.

If it helps anyone: if you are running into problem like '... in non-interactive, and thus cannot be shown', the warning comes from the base FigureManagerBase .show() method, if you extend the class and not override the show() method NonGuiException( f"{type(self.canvas).__name__} is non-interactive, and thus cannot be " f"shown") , or, when you think you are overriding the method in your custom backend but, do not actually register your backend.

Again, thanks!