Re: [yt-dev] Implementing Backend Independent Plots: Pull Request 2294

Hi Austin and Nathan, On Thu, Jul 28, 2016 at 10:51 AM, Nathan Goldbaum wrote:

Hi all,

I just want to voice my main issues with the PR as it's currently implemented. I've communicated this with Austin in private but want to make sure this is all publicly written down so others can offer advice.

Right now the PR makes it possible to do something like this:

import yt yt.toggle_interactivity()

slc = yt.SlicePlot(ds, 2, 'density') slc.show()

and then have a GUI window pop up showing the plot.

The toggle_interactivity() function queries matplotlib to get it's current backend, and then uses that setting to choose a FigureCanvas subclass to draw the plot onto.

Unfortunately, matplotlib's support for creating working plots outside of pyplot using a GUI backend is basically nil. For example, using Austin's pull request with the MacOSX backend, I can't interact with the plot window that pops up (mousing over it produces a beachball hang) and I need to exit the python interpreter to close the window. If matplotlib is configured to use the QT4 backend (e.g. most anaconda installs on linux), nothing shows up except for a white window. In the TK backend, an exception gets printed out. It *does* seem to work with the notebook backend though. I haven't tried any of the other GUI backends.

I didn't try manually testing Austin's pull request so the PR went through a pass or two of review under the assumption that the GUI backends were mostly working. I'd like to apologize to Austin for giving him comments and suggestions that led to code improvements even though right now I'm arguing that we should abandon the approach as implemented - I could definitely have handled reviewing that PR with more empathy.

Given that we have very little control over matplotlib's configuration and which backend it will select at runtime, I don't think the API that Austin has implemented will be very useful for most users. Most people will either see a non-interactive plot window pop up that they can't close, no plot at all, or get an error. At the matplotlib BOF at SciPy this year there was some discussion about making matplotlib's backends more uniform and exposing them better to the object oriented interface, but I don't think it's straightforward right now to use them as we'd like to without relying on pyplot somehow.

That said, I know that Austin wants to use this for his own QT application. I think the best thing to do there is to add an interface for users who want to supply their own canvas objects to draw to. That way anyone who wants to manage their own GUI event loop will still be able to draw to a GUI canvas using yt's plotting code. I'd be very happy to spend time getting this working.

Alternatively, if we do want to keep the API Austin has worked out, I think we'll need to manage our own GUI event loop, and write backend-dependent support code to handle that. I think that this approach will be difficult although certainly with a nice payoff - it's a very commonly requested feature.

I'm not sure I'm as pessimistic -- I've been experimenting with using the PR inside ipython with the matplotlib backend set, and it seems to work. I think that it's reasonable to mandate that if we're running in an IPython session (not a notebook) where the thread model is set via --matplotlib= , that we should allow for this functionality.

Sorry that was a bit of a novel - wanted to distill my thinking on this as clearly as possible.

-Nathan

On Thu, Jul 28, 2016 at 12:26 PM, Gilbert, Austin B

...

wrote:

...

Hello All,

I recently submitted a pull request that enables all yt plots to utilize the various interactive backends matplotlib(mpl) has to offer. The primary benefit of this change is that it enables a few more features from mpl in jupyter notebooks, but it also creates the possibility for user created applications using python gui packages (pyqt, tk, cairo, etc.). However, enabling this feature outside of a notebook requires the user to configure matplotlib for the backend they are using, as well as implementing their own event loop for the application. I would like to hear from others if they feel this is to complex a task to put on the user, and also if they would utilize the feature.

Austin Gilbert

_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org

_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org