Include gridliner labels in tight bbox calculation

[lukelbd]

This PR replaces #1354.

Rationale

matplotlib.axes.Axes.get_tightbbox fails to include cartopy gridliner labels. This is because cartopy.mpl.gridliner.Gridliner._draw_gridliner is not called until Axes.draw is invoked, which comes after Figure.draw invokes Figure.tight_layout.

This change adds a get_tightbbox override to the GeoAxes class, which calls _draw_gridliner on all the GridLiner objects, so the tight layout algorithm will work. This mimics the approach used in the matplotlib API: Axes.draw and Axes.get_tightbbox contain many identical calls to various post-processing steps, like _update_title_position.

Implications

You can now use fig.tight_layout() on figures with cartopy axes containing gridliner labels.

Examples

I wasn't able to install a development version of cartopy, but I implemented a virtually identical change in a GeoAxes subclass in my proplot package (proplot calls fig.tight_layout() on all figures by default).

Below, the first example is without the get_tightbbox override (there are labels present, they are just off the edge of the figure). The second example is with the get_tightbbox override.

import proplot as plot
fig, ax = plot.subplots(proj='pcarree', axwidth=3)
ax.format(labels=True, lonlines=60, latlines=30, title='Plate Carrée')
fig.save('example.png')

A similar pyplot example could be generated by just calling fig.tight_layout() on any cartopy plot with gridlines.

[dopplershift]

In general, this looks good. I have one question, and then would you be willing to add your sample as a test?

[greglucas]

@lukelbd did you sign the CLA? Could you rebase if so to rerun CI on this.

[lukelbd]

@greglucas Sorry for the delay; I've signed the CLA, and can try to add a test by next weekend.

[greglucas]

Sounds good, @lukelbd! @dopplershift was there anything else you were requesting in addition to the test?

[dopplershift]

@greglucas I think at this point it just needs a test.

[greglucas]

ping @lukelbd.

[greglucas]

@lukelbd we are going to try and push another beta release out this weekend. Do you think you can add the test by then, or should we look to push this back to the next 0.19 release milestone?

[QuLogic]

Tagging the rc today, so I'm pushing this back to 0.19.

[greglucas]

I just ran into the case of saving a pdf and the side grid labels were partially cut off. I think this may solve that issue...? It would be good to get it into the code since it seems logical enough.

[lukelbd]

Sorry forgot about this PR, was really busy when I was pinged. Will take a look this weekend. I think this will also need to be updated for the 0.18 gridliner API

[lukelbd]

I added a test for the "latest everything" build only. Turns out matplotlib 1.5.1 does not use Figure.get_tightbbox() to calculate the bounding box -- it uses a get_figure_tight_layout() function in tight_layout.py instead. So it is impossible to override the tight layout algorithm:

https://github.com/matplotlib/matplotlib/blob/be91fac9fa2b7250080557e723af75124659da4e/lib/matplotlib/figure.py#L1720-L1756

Update 2020-07-14: The above reasoning was incorrect, but tight layout override does fail in matplotlib 1.5.0. After empirical testing, the tight layout override works only in matplotlib >=3.0. See #1575 for details.

My test is similar to test_grid_labels, but it calls Figure.tight_layout() at the end and ensures that every gridliner was drawn by get_tightbbox() by ensuring Gridliner._plotted == True.

Here is the baseline result:

And here is the result without the Figure.tight_layout() call:

[lukelbd]

Not sure why it's failing... I built the test image by running pytest as follows:

conda config --set always_yes yes --set changeps1 no --set show_channel_urls yes
conda config --add channels conda-forge
conda config --add channels conda-forge/label/testing
conda config --set restore_free_channel true

export CYTHON_COVERAGE=1
echo "backend: agg" > ~/.matplotlib/matplotlibrc  # macOS matplotlibrc

conda create -n cartopy-dev python=3.6 cython numpy matplotlib-base proj4 pykdtree scipy fiona pillow pytest pytest-xdist filelock pep8 pyshp shapely six requests pyepsg owslib pytest-cov coveralls
conda activate cartopy-dev

pip install --no-deps -e .
python -c "import cartopy; print('Version ', cartopy.__version__)" && python setup.py version
pytest

[lukelbd]

Nevermind -- looks like the latest push fixed it! I might have accidentally generated the baseline image without the backend: agg line or in the matplotlib 1.5.1 environment.

Anything else I need to do?

[greglucas]

Thanks for adding the image test! Overall I think this looks good. Just a few minor comments.

[lukelbd]

@greglucas Done!

[greglucas]

Looks good to me.

[lukelbd]

@dopplershift I think the remaining requested change is out of date