Skip to content

[Bug]: Rendering inconsistency between matplotlib-inline plt.show() and fig.savefig(...) #24281

Closed as not planned

Description

@posita

Bug summary

This originally came up in jupyterlite/jupyterlite#850, but I'm pretty sure the root cause is here. I'll reproduce the most helpful comment from that thread here. The following is a pretty good demonstration of how savefig clips (e.g.) the title of a pie chart.

Possibly related issues:

Code for reproduction

Run the following in Jupyter Lab to reproduce the attached images.

import matplotlib.pyplot
import base64, io
from enum import Enum
from ipywidgets import widgets
from fractions import Fraction
from itertools import chain

data = {3: Fraction(1, 216), 4: Fraction(1, 72), 5: Fraction(1, 36), 6: Fraction(5, 108), 7: Fraction(5, 72), 8: Fraction(7, 72), 9: Fraction(25, 216), 10: Fraction(1, 8), 11: Fraction(1, 8), 12: Fraction(25, 216), 13: Fraction(7, 72), 14: Fraction(5, 72), 15: Fraction(5, 108), 16: Fraction(1, 36), 17: Fraction(1, 72), 18: Fraction(1, 216)}

class Format(str, Enum):
    NATIVE = "native"
    PNG = "png"
    SVG = "svg"

class PngImage:
    def __init__(self, data):
        self.data = data

    def _repr_png_(self):
        return base64.b64encode(self.data).decode()

class SvgImage:
    def __init__(self, data):
        self.data = data

    def _repr_svg_(self):
        return self.data.decode()

@widgets.interact
def show(output_to_format = Format, use_subplots: bool = True):
    with matplotlib.style.context("bmh"):
        if use_subplots:
            _, ax = matplotlib.pyplot.subplots()
        else:
            fig = matplotlib.pyplot.figure()
            xmin, ymin, dx, dy = 0.1, 0.1, 0.9, 0.9
            ax = fig.add_axes((xmin, ymin, dx, dy))
        unique_outcomes = sorted(data.keys())
        ax.set_title(
            "Super duper duper long title that will probably exceed the width of the pie chart",
            fontdict={"fontweight": "bold",},
        )
        outcomes, values = zip(*((outcome, float(value)) for outcome, value in data.items()))
        ax.pie(outcomes, values)
        if output_to_format is Format.NATIVE:
            matplotlib.pyplot.show()
        else:
            buf = io.BytesIO()
            matplotlib.pyplot.savefig(buf, format=output_to_format)
            if output_to_format is Format.PNG:
                display(PngImage(buf.getvalue()))
            elif output_to_format is Format.SVG:
                display(SvgImage(buf.getvalue()))
            else:
                assert False, "should never be here"
            matplotlib.pyplot.clf()

print(matplotlib.__version__)

Actual outcome

Output method Using subplots Using add_axes
show() native-subplots native-manual-axes
savefig(..., format="png") png-subplots png-manual-axes
savefig(..., format="svg") svg-subplots svg-manual-axes

Expected outcome

I expected all backends to produce at least similar results (i.e., not ones where titles were clipped or missing entirely).

Additional information

No response

Operating system

No response

Matplotlib Version

3.5.2, 3.6.1

Matplotlib Backend

module://matplotlib_inline.backend_inline

Python version

3.10.6

Jupyter version

Lab 3.4.8

Installation

No response

Activity

  1. jklymak commented on Oct 26, 2022

    @jklymak
    Member

    It looks like the widget is doing the equivalent of matplotlib.pyplot.savefig(buf, format=output_to_format, bbox_inches='tight') which will clip the figure to whatever artists are in it. That is the choice the inline backend makes as well. I don't think this is a bug, or actionable on our part.

  2. posita commented on Oct 26, 2022

    @posita
    Author

    It looks like the widget is doing the equivalent of matplotlib.pyplot.savefig(buf, format=output_to_format, bbox_inches='tight') which will clip the figure to whatever artists are in it. That is the choice the inline backend makes as well. I don't think this is a bug, or actionable on our part.

    Why would that yield a different result from, say matplotlib.pyplot.tight_layout() ; matplotlib.pyplot.show()?

  3. posita commented on Oct 26, 2022

    @posita
    Author

    @jklymak, how would one go about getting similar results for both matplotlib.pyplot.show() and matplotlib.pyplot.savefig(...) then? Or are you saying this is an ipywidgets issue and should be filed against that project?

  4. jklymak commented on Oct 26, 2022

    @jklymak
    Member

    It looks like the widget is doing the equivalent of matplotlib.pyplot.savefig(buf, format=output_to_format, bbox_inches='tight') which will clip the figure to whatever artists are in it. That is the choice the inline backend makes as well. I don't think this is a bug, or actionable on our part.

    Why would that yield a different result from, say matplotlib.pyplot.tight_layout() ; matplotlib.pyplot.show()?

    Because bbox_inches='tight' clips the figure to the artists as draw, plt.tight_layout adjusts the size of the axes so the artists fit in the figure, but doesn't change the size of the figure.

    @jklymak, how would one go about getting similar results for both matplotlib.pyplot.show() and matplotlib.pyplot.savefig(...) then?

    To get similar results, you pass bbox_inches='tight' to savefig. I think there is also a configuration option to change ipython to not do bbox_inches='tight', but forget what it is.

  5. changed the title [-][Bug]: Rendering inconsistency between `plt.show()` and `fig.savefig(...)`[/-] [+][Bug]: Rendering inconsistency between matplotlib-inline `plt.show()` and `fig.savefig(...)`[/+] on Oct 26, 2022
  6. QuLogic commented on Oct 26, 2022

    @QuLogic
    Member

    ipywidgets uses the inline backend to display plots. That defaults to a tight bbox.

    It can be disabled by overriding their save options.

    Going to close as not our bug.

  7. posita commented on Oct 26, 2022

    @posita
    Author

    Thanks for the insight! That actually led me to finding jupyter/notebook#2640, including jupyter/notebook#2640 (comment) , which offers a way to bring parity.

    UPDATE: Hah! Beat me to it, @QuLogic!

  8. tacaswell commented on Oct 26, 2022

    @tacaswell
    Member

    To be very clear here, there are a bunch of moving parts:

    • matplotlib
    • matplotlib.pyplot (the stateful MATLAB-like plotting application we provide)
    • malptolib-inline (the backend you are using that is externally maintained)
    • ipywidgets (which is involved in the example but not responsible for any of the problems)

    that have become a bit tangled. I just want to make sure we are not talking past each other.

    The inline backend (which used to be part of IPython and is now its own package) partially replicates the desktop interactive backends by dropping a static png into the output cells of the notebook when it would have normally poped up a GUI window. The png is embeded in the notebook as a data URL with the base64 encoded binary and the png data is rendered using (effectively) fig.savefig(buf, bbox_inches='tight'). The use of bbox_inches='tight' is the key detail as it clips the figure size to exactly fit the artists in the Figure. The somewhat surprising behavior of this clipping is that in addition to making the Figure smaller (trimming whitespace), it also can make the figure effectively bigger.

    A minimal example to reproduce this behavior:

    import matplotlib.pyplot
    from fractions import Fraction
    
    
    import matplotlib.pyplot as plt
    
    data = {
        3: Fraction(1, 216),
        4: Fraction(1, 72),
        5: Fraction(1, 36),
        6: Fraction(5, 108),
        7: Fraction(5, 72),
        8: Fraction(7, 72),
        9: Fraction(25, 216),
        10: Fraction(1, 8),
        11: Fraction(1, 8),
        12: Fraction(25, 216),
        13: Fraction(7, 72),
        14: Fraction(5, 72),
        15: Fraction(5, 108),
        16: Fraction(1, 36),
        17: Fraction(1, 72),
        18: Fraction(1, 216),
    }
    
    fig, ax = matplotlib.pyplot.subplots(figsize=(4, 4))
    unique_outcomes = sorted(data.keys())
    ax.set_title(
        "Super duper duper long title that will probably exceed the width of the pie chart",
        fontdict={
            "fontweight": "bold",
        },
    )
    outcomes, values = zip(*((outcome, float(value)) for outcome, value in data.items()))
    ax.pie(outcomes, values)
    # this will clip the title, but will be the size you asked for
    fig.savefig('test1.png')                       
    # this will include the full title, but it is bigger than you asked for and not square
    fig.savefig('test2.png', bbox_inches='tight')

    There is no chance that we will change the default behavior of savefig to do this resizing (targeting fixed-sized figures is a major part of what we do) and I think it is also far too late for matplotlib-inline to change their default (the wave of bug reports of people who implicitly depended on this would be huge!).

    I think you option are either to pass bbox_inches='tight' when you save or configure the inline backend to not pass it (I think

    %config InlineBackend.print_figure_kwargs = {'bbox_inches':None}
    

    will work interactively and there is a way to put this in your profile configuration files).

    https://stackoverflow.com/questions/37864735/matplotlib-and-ipython-notebook-displaying-exactly-the-figure-that-will-be-save


    I took the liberty of updating the title to include matplotlib-inline.

  9. posita commented on Oct 29, 2022

    @posita
    Author

    Thanks @tacaswell! That is a great breakdown and very helpful not only for me in understanding the layers, but hopefully for folks developing JupyterLite or anyone else who stumbles across this issue. I really appreciate you taking the time to lay it all out in such a way that even a stranger to these technologies (like me) can easily understand. It am especially grateful for the layout of possible options and explanations of why some were off the table. I feel much more aware of things now.

    Many, many thanks again! 🙇

  10. tacaswell commented on Oct 29, 2022

    @tacaswell
    Member

    @posita Your welcome!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions