Repository navigation
Windows are not shown with Tk + interactive mode, from matplotlib 3.4 #19935
Description
Bug report
Bug summary
The behavior of Tk backend in interactive mode has changed. My code depends on the old behavior and does not work with matplotlib 3.4.
Code for reproduction
import time
import numpy as np
import matplotlib as mpl
from matplotlib import pyplot as plt
def test_heatmap_draw():
mpl.use("TkAgg")
plt.ion()
DATA_SHAPE = 10, 10
fig = plt.figure()
ax = fig.add_subplot()
img = ax.imshow(np.zeros(DATA_SHAPE), vmin=-1, vmax=1)
for _ in range(10):
img.set_data(np.random.uniform(-1, 1, DATA_SHAPE))
fig.canvas.draw()
time.sleep(0.1)
if __name__ == "__main__":
test_heatmap_draw()Actual outcome
No windows are shown with matplotlib >= 3.4.
I guess that #17789 is the trigger. Inserting self._master.update_idletasks worked, like:
class FigureCanvasTkAgg(FigureCanvasAgg, FigureCanvasTk):
def draw(self):
super().draw()
self.blit()
self._master.update_idletasks()Expected outcome
The heatmap is interactively updated with matplotlib 3.3.4.
- Is this change intended? Then I want to know a proper way to write this kind of app. I know that
plt.pause(0.1)works, but actually don't want to pause in my actual use case.- Also, could I help enhancing the document around this? E.g., in https://matplotlib.org/stable/users/interactive.html .
- If this is not intended, how could I fix backends in a reasonable way?
BTW, thank you for encouraging me on twitter.
Activity
richardsheridan commentedon Apr 11, 2021
My opinion is that this is the intended behavior. It is harmful to call
update_idletasksorupdatein TCL library code for the reasons mentioned in that issue. However, client code should be free to use it. In this case the appropriate api, since you are driving your own event loop, isfigure.canvas.flush_eventswhich just callsupdatebut in a backend-agnostic way.Updating the docs would be very nice, I'll help where I can!
kngwyu commentedon Apr 12, 2021
@richardjgowers
Thank you for your help! That makes sense.
For documents, I think the Explicitly spinning the Event Loop is enough good, but I feel the more entry-level description of interactive mode is somewhat confusing.
It says that:
However, I guess that we need
flush_eventsor some blocking functions to render images. Are these descriptions are simply outdated? Or, do I miss something?Any help is greatly appricated.
richardsheridan commentedon Apr 12, 2021
A little farther down in that page there are some exceptions to those bullet points:
and
I think it is a subtle and possibly confusing point that interactive mode is not interactive unless you are working in a command prompt with event loop integration. That is, if you run a script with matplotlib code such as what you posted above, you will have a different result of interactivity compared with typing each command at the REPL.In my mind, the thing you need to do to get interactivity is pump GUI events, and
plt.ionjust configures some of the pyplot API to "work better" at the REPL.Perhaps bringing these facts forward in
doc/users/interactive.rstbefore the bullet points would help?QuLogic commentedon Apr 13, 2021
You should replace this with
plt.pause(0.1)to run the main loop.kngwyu commentedon Apr 13, 2021
Right, I didn't actually understand that point.
Thank you for the pointer. I would like also to mention that some backends behaved inconsistently on this point, but they are fixed.