Interactive Figures and Asynchronous Programming

Matplotlib supports rich interactive figures by embedding figures into a GUI window. The basic interactions of panning and zooming in an Axes to inspect your data is 'baked in' to Matplotlib. This is supported by a full mouse and keyboard event handling system that you can use to build sophisticated interactive graphs.

This guide is meant to be an introduction to the low-level details of how Matplotlib integration with a GUI event loop works. For a more practical introduction to the Matplotlib event API see event handling system, Interactive Tutorial, and Interactive Applications using Matplotlib.

Event Loops

Fundamentally, all user interaction (and networking) is implemented as an infinite loop waiting for events from the user (via the OS) and then doing something about it. For example, a minimal Read Evaluate Print Loop (REPL) is

exec_count = 0
while True:
    inp = input(f"[{exec_count}] > ")        # Read
    ret = eval(inp)                          # Evaluate
    print(ret)                               # Print
    exec_count += 1                          # Loop

This is missing many niceties (for example, it exits on the first exception!), but is representative of the event loops that underlie all terminals, GUIs, and servers [1]. In general the Read step is waiting on some sort of I/O -- be it user input or the network -- while the Evaluate and Print are responsible for interpreting the input and then doing something about it.

In practice we interact with a framework that provides a mechanism to register callbacks to be run in response to specific events rather than directly implement the I/O loop [2]. For example "when the user clicks on this button, please run this function" or "when the user hits the 'z' key, please run this other function". This allows users to write reactive, event-driven, programs without having to delve into the nitty-gritty [3] details of I/O. The core event loop is sometimes referred to as "the main loop" and is typically started, depending on the library, by methods with names like _exec, run, or start.

All GUI frameworks (Qt, Wx, Gtk, tk, OSX, or web) have some method of capturing user interactions and passing them back to the application (for example Signal / Slot framework in Qt) but the exact details depend on the toolkit. Matplotlib has a backend for each GUI toolkit we support which uses the toolkit API to bridge the toolkit UI events into Matplotlib's event handling system. You can then use FigureCanvasBase.mpl_connect to connect your function to Matplotlib's event handling system. This allows you to directly interact with your data and write GUI toolkit agnostic user interfaces.

Command Prompt Integration

So far, so good. We have the REPL (like the IPython terminal) that lets us interactively send code to the interpreter and get results back. We also have the GUI toolkit that runs an event loop waiting for user input and lets us register functions to be run when that happens. However, if we want to do both we have a problem: the prompt and the GUI event loop are both infinite loops that each think they are in charge! In order for both the prompt and the GUI windows to be responsive we need a method to allow the loops to 'timeshare' :

  1. let the GUI main loop block the python process when you want interactive windows
  2. let the CLI main loop block the python process and intermittently run the GUI loop
  3. fully embed python in the GUI (but this is basically writing a full application)

Blocking the Prompt

pyplot.show Display all open figures.
pyplot.pause Run the GUI event loop for interval seconds.
backend_bases.FigureCanvasBase.start_event_loop Start a blocking event loop.
backend_bases.FigureCanvasBase.stop_event_loop Stop the current blocking event loop.

The simplest "integration" is to start the GUI event loop in 'blocking' mode and take over the CLI. While the GUI event loop is running you can not enter new commands into the prompt (your terminal may echo the characters typed into the terminal, but they will not be sent to the Python interpreter because it is busy running the GUI event loop), but the figure windows will be responsive. Once the event loop is stopped (leaving any still open figure windows non-responsive) you will be able to use the prompt again. Re-starting the event loop will make any open figure responsive again (and will process any queued up user interaction).

To start the event loop until all open figures are closed use pyplot.show as

pyplot.show(block=True)

To start the event loop for a fixed amount of time (in seconds) use pyplot.pause.

If you are not using pyplot you can start and stop the event loops via FigureCanvasBase.start_event_loop and FigureCanvasBase.stop_event_loop. However, in most contexts where you would not be using pyplot you are embedding Matplotlib in a large GUI application and the GUI event loop should already be running for the application.

Away from the prompt, this technique can be very useful if you want to write a script that pauses for user interaction, or displays a figure between polling for additional data. See Scripts and functions for more details.

Input Hook integration

While running the GUI event loop in a blocking mode or explicitly handling UI events is useful, we can do better! We really want to be able to have a usable prompt and interactive figure windows.

We can do this using the 'input hook' feature of the interactive prompt. This hook is called by the prompt as it waits for the user to type (even for a fast typist the prompt is mostly waiting for the human to think and move their fingers). Although the details vary between prompts the logic is roughly

  1. start to wait for keyboard input
  2. start the GUI event loop
  3. as soon as the user hits a key, exit the GUI event loop and handle the key
  4. repeat

This gives us the illusion of simultaneously having interactive GUI windows and an interactive prompt. Most of the time the GUI event loop is running, but as soon as the user starts typing the prompt takes over again.

This time-share technique only allows the event loop to run while python is otherwise idle and waiting for user input. If you want the GUI to be responsive during long running code it is necessary to periodically flush the GUI event queue as described above. In this case it is your code, not the REPL, which is blocking the process so you need to handle the "time-share" manually. Conversely, a very slow figure draw will block the prompt until it finishes drawing.

Full embedding

It is also possible to go the other direction and fully embed figures (and a Python interpreter) in a rich native application. Matplotlib provides classes for each toolkit which can be directly embedded in GUI applications (this is how the built-in windows are implemented!). See Embedding Matplotlib in graphical user interfaces for more details.

Scripts and functions

backend_bases.FigureCanvasBase.flush_events Flush the GUI events for the figure.
backend_bases.FigureCanvasBase.draw_idle Request a widget redraw once control returns to the GUI event loop.
figure.Figure.ginput Blocking call to interact with a figure.
pyplot.ginput Blocking call to interact with a figure.
pyplot.show Display all open figures.
pyplot.pause Run the GUI event loop for interval seconds.

There are several use-cases for using interactive figures in scripts:

  • capture user input to steer the script
  • progress updates as a long running script progresses
  • streaming updates from a data source

Blocking functions

If you only need to collect points in an Axes you can use figure.Figure.ginput or more generally the tools from blocking_input the tools will take care of starting and stopping the event loop for you. However if you have written some custom event handling or are using widgets you will need to manually run the GUI event loop using the methods described above.

You can also use the methods described in Blocking the Prompt to suspend run the GUI event loop. Once the loop exits your code will resume. In general, any place you would use time.sleep you can use pyplot.pause instead with the added benefit of interactive figures.

For example, if you want to poll for data you could use something like

fig, ax = plt.subplots()
ln, = ax.plot([], [])

while True:
    x, y = get_new_data()
    ln.set_data(x, y)
    plt.pause(1)

which would poll for new data and update the figure at 1Hz.

Explicitly spinning the Event Loop

backend_bases.FigureCanvasBase.flush_events Flush the GUI events for the figure.
backend_bases.FigureCanvasBase.draw_idle Request a widget redraw once control returns to the GUI event loop.

If you have open windows that have pending UI events (mouse clicks, button presses, or draws) you can explicitly process those events by calling FigureCanvasBase.flush_events. This will run the GUI event loop until all UI events currently waiting have been processed. The exact behavior is backend-dependent but typically events on all figure are processed and only events waiting to be processed (not those added during processing) will be handled.

For example

import time
import matplotlib.pyplot as plt
import numpy as np
plt.ion()

fig, ax = plt.subplots()
th = np.linspace(0, 2*np.pi, 512)
ax.set_ylim(-1.5, 1.5)

ln, = ax.plot(th, np.sin(th))

def slow_loop(N, ln):
    for j in range(N):
        time.sleep(.1)  # to simulate some work
        ln.figure.canvas.flush_events()

slow_loop(100, ln)

While this will feel a bit laggy (as we are only processing user input every 100ms whereas 20-30ms is what feels "responsive") it will respond.

If you make changes to the plot and want it re-rendered you will need to call draw_idle to request that the canvas be re-drawn. This method can be thought of draw_soon in analogy to asyncio.loop.call_soon.

We can add this our example above as

def slow_loop(N, ln):
    for j in range(N):
        time.sleep(.1)  # to simulate some work
        if j % 10:
            ln.set_ydata(np.sin(((j // 10) % 5 * th)))
            ln.figure.canvas.draw_idle()

        ln.figure.canvas.flush_events()

slow_loop(100, ln)

The more frequently you call FigureCanvasBase.flush_events the more responsive your figure will feel but at the cost of spending more resources on the visualization and less on your computation.

Stale Artists

Artists (as of Matplotlib 1.5) have a stale attribute which is True if the internal state of the artist has changed since the last time it was rendered. By default the stale state is propagated up to the Artists parents in the draw tree, e.g., if the color of a Line2D instance is changed, the axes.Axes and figure.Figure that contain it will also be marked as "stale". Thus, fig.stale will report if any artist in the figure has been modified and is out of sync with what is displayed on the screen. This is intended to be used to determine if draw_idle should be called to schedule a re-rendering of the figure.

Each artist has a Artist.stale_callback attribute which holds a callback with the signature

def callback(self: Artist, val: bool) -> None:
   ...

which by default is set to a function that forwards the stale state to the artist's parent. If you wish to suppress a given artist from propagating set this attribute to None.

figure.Figure instances do not have a containing artist and their default callback is None. If you call pyplot.ion and are not in IPython we will install a callback to invoke draw_idle whenever the figure.Figure becomes stale. In IPython we use the 'post_execute' hook to invoke draw_idle on any stale figures after having executed the user's input, but before returning the prompt to the user. If you are not using pyplot you can use the callback Figure.stale_callback attribute to be notified when a figure has become stale.

Draw Idle

backend_bases.FigureCanvasBase.draw Render the Figure.
backend_bases.FigureCanvasBase.draw_idle Request a widget redraw once control returns to the GUI event loop.
backend_bases.FigureCanvasBase.flush_events Flush the GUI events for the figure.

In almost all cases, we recommend using backend_bases.FigureCanvasBase.draw_idle over backend_bases.FigureCanvasBase.draw. draw forces a rendering of the figure whereas draw_idle schedules a rendering the next time the GUI window is going to re-paint the screen. This improves performance by only rendering pixels that will be shown on the screen. If you want to be sure that the screen is updated as soon as possible do

fig.canvas.draw_idle()
fig.canvas.flush_events()

Threading

Most GUI frameworks require that all updates to the screen, and hence their main event loop, run on the main thread. This makes pushing periodic updates of a plot to a background thread impossible. Although it seems backwards, it is typically easier to push your computations to a background thread and periodically update the figure on the main thread.

In general Matplotlib is not thread safe. If you are going to update Artist objects in one thread and draw from another you should make sure that you are locking in the critical sections.

Eventloop integration mechanism

CPython / readline

The Python C API provides a hook, PyOS_InputHook, to register a function to be run "The function will be called when Python's interpreter prompt is about to become idle and wait for user input from the terminal.". This hook can be used to integrate a second event loop (the GUI event loop) with the python input prompt loop. The hook functions typically exhaust all pending events on the GUI event queue, run the main loop for a short fixed amount of time, or run the event loop until a key is pressed on stdin.

Matplotlib does not currently do any management of PyOS_InputHook due to the wide range of ways that Matplotlib is used. This management is left to downstream libraries -- either user code or the shell. Interactive figures, even with matplotlib in 'interactive mode', may not work in the vanilla python repl if an appropriate PyOS_InputHook is not registered.

Input hooks, and helpers to install them, are usually included with the python bindings for GUI toolkits and may be registered on import. IPython also ships input hook functions for all of the GUI frameworks Matplotlib supports which can be installed via %matplotlib. This is the recommended method of integrating Matplotlib and a prompt.

IPython / prompt toolkit

With IPython >= 5.0 IPython has changed from using cpython's readline based prompt to a prompt_toolkit based prompt. prompt_toolkit has the same conceptual input hook, which is fed into prompt_toolkit via the IPython.terminal.interactiveshell.TerminalInteractiveShell.inputhook() method. The source for the prompt_toolkit input hooks lives at IPython.terminal.pt_inputhooks

Footnotes

[1]

A limitation of this design is that you can only wait for one input, if there is a need to multiplex between multiple sources then the loop would look something like

fds = [...]
while True:                    # Loop
    inp = select(fds).read()   # Read
    eval(inp)                  # Evaluate / Print
[2]Or you can write your own if you must.
[3]These examples are aggressively dropping many of the complexities that must be dealt with in the real world such as keyboard interrupts, timeouts, bad input, resource allocation and cleanup, etc.