GeezerVison

Another size test:

$ git clone https://github.com/matplotlib/matplotlib.git mpl
$ cd mpl
$ du -hsc .git
461M	.git
461M	total

$ git fetch https://github.com/QuLogic/matplotlib.git ft213:ft213
$ du -hsc .git
477M	.git
477M	total

So once everything is repacked and compressed by git, the repo only gains 16M!

If we're going to regen the baselines, we should also take that opportunity to remove some of the backcompat knobs -- at least text.kerning_factor (which clearly only existed to avoid regen'ing the baselines), and possibly text.hinting_factor as well (at least I would switch the default to 1 and generate the images with that amount of hinting, even if we decide to keep this functionality).

Rather than regenerating the images at this point, I would start by pulling the tolerances used by Fedora upstream (so we can backport that to 3.10.x) and then in a second step regenerate the images on main and push the tolerances back to 0.

If we're going to regen the baselines, we should also take that opportunity to remove some of the backcompat knobs -- at least text.kerning_factor (which clearly only existed to avoid regen'ing the baselines), and possibly text.hinting_factor as well (at least I would switch the default to 1 and generate the images with that amount of hinting, even if we decide to keep this functionality).

Yes, good point; I will see what effect that has on things.

Rather than regenerating the images at this point, I would start by pulling the tolerances used by Fedora upstream (so we can backport that to 3.10.x) and then in a second step regenerate the images on main and push the tolerances back to 0.

Fedora doesn't have any other tolerances other than the 6 other files already changed (though in this PR I changed the results instead of the tolerance, we could do like in Fedora instead). For image tests, we overlay the correct images from my mpl-images repo and don't change any global tolerance.

Removing kerning_factor from _classic_test.patch causes an additional 100 test failures over this PR. Top 20 RMS are:

• lib/matplotlib/tests/test_backend_pdf.py::test_kerning[pdf] 23.532
• lib/matplotlib/tests/test_arrow_patches.py::test_boxarrow[png] 20.492
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout9[png] 17.849
• lib/matplotlib/tests/test_mathtext.py::test_mathtext_rendering_lightweight[png-mathtext1-dejavusans-8] 17.244
• lib/matplotlib/tests/test_table.py::test_auto_column[png] 16.872
• lib/matplotlib/tests/test_text.py::test_basic_wrap[png] 15.712
• lib/matplotlib/tests/test_axes.py::test_pie_nolabel_but_legend[png] 12.626
• lib/matplotlib/tests/test_tightlayout.py::test_tight_layout7[pdf] 12.406
• lib/matplotlib/tests/test_mathtext.py::test_mathtext_rendering[pdf-mathtext-dejavusans-46] 11.002
• lib/matplotlib/tests/test_axes.py::test_stairs_datetime[png] 10.796
• lib/matplotlib/tests/test_tightlayout.py::test_tight_layout7[png] 10.388
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout5[png] 10.191
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout2[png] 10.191
• lib/matplotlib/tests/test_tightlayout.py::test_tight_layout7[svg] 9.704
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout4[png] 9.014
• lib/matplotlib/tests/test_legend.py::test_multiple_keys[png] 8.322
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout3[png] 7.658
• lib/matplotlib/tests/test_mathtext.py::test_mathtext_rendering[pdf-mathtext-dejavuserif-60] 7.572
• lib/matplotlib/tests/test_text.py::test_font_styles[pdf] 7.355
• lib/matplotlib/tests/test_constrainedlayout.py::test_constrained_layout15[png] 7.259

I'll commit these all separately to make it easier to review, but we can squash these all together if it's all working fine. The majority of these are things like Te, te, x, er all kern a little bit closer together now.

On a related note, and I'm not sure which ones are relevant here, but I wonder if we should drop several PDF and SVG test images when they would be changed but don't effectively test anything different from the PNG.

For the repo size problem we could provide some guidance for the depth flag so most contributors don't need to download the whole thing. Eg --depth=500 gets you the last two years of commits and is x MB. Most folks don't need the full commit history (or even anything more than depth=1)

The first image comparison that shows up lib/matplotlib/tests/baseline_images/test_agg/agg_filter.png doesn't have any text in it at all so is presumably not an update due to freetype. Is this all images that have had tolerances ever adjusted in the past? If so, do you know if there would be any size difference with just text-based freetype updates versus all others?

Is it beneficial/necessary to grab the images from that other repo you've been keeping, or would it make sense to blow away the images directory on CI and move all the resultant images into the directory to start fresh in a sense?

and possibly text.hinting_factor as well (at least I would switch the default to 1 and generate the images with that amount of hinting, even if we decide to keep this functionality).

Surprisingly, this doesn't appear to make any difference to test images. It only affects 9 tests, and only with regards to bboxes and tight layout that are explicitly compared. This doesn't matter whether I try with FreeType 2.6.1 or 2.13.3.

The first image comparison that shows up lib/matplotlib/tests/baseline_images/test_agg/agg_filter.png doesn't have any text in it at all so is presumably not an update due to freetype. Is this all images that have had tolerances ever adjusted in the past? If so, do you know if there would be any size difference with just text-based freetype updates versus all others?

Ah, you're right. I had not done a deep dive into the changes, but likely this was an image that had text that was removed, so it was regenerated from the FreeType changes, but effectively looked the same. I will go through these again and make sure we don't have any extraneous changes like this one.

Is it beneficial/necessary to grab the images from that other repo you've been keeping, or would it make sense to blow away the images directory on CI and move all the resultant images into the directory to start fresh in a sense?

The other repo was useful for tracking purposes and a quick changeover, but it might be best to start fresh and make sure we don't change too many extra images.

I haven't updated this PR for it yet, but going over the diffs, one surprising change is that while the mathtext tests with DejaVu Sans look better (more consistent baselines, tweaks to kerning), the same ones with Computer Modern look worse (mostly due to a wobbly baseline.) But on the other hand, a lot of those had the wobbly baseline before...

I've also opened #29827 to remove most of the redundant PDFs that would be changed here (except the ones from test_backend_pdf, plus a few others noted in that PR).

See also #14177 (comment) and #22459 re: wobbly baseline. I have some hopes that switching agg+usetex to rasterizing glyphs ourselves (see also #29807) would be one way to fix that.

Slightly unrelated, but if we decide to update this many PNGs, it should be a good time to add an oxipng pre-commit hook. https://github.com/shssoichiro/oxipng I would expect that the sizes are halved or so on average. (It would be useful independent of this, but I have not gotten around to propose it and it would be unfortunate to change all those images without compression.)

See also #14177 (comment) and #22459 re: wobbly baseline. I have some hopes that switching agg+usetex to rasterizing glyphs ourselves (see also #29807) would be one way to fix that.

But these are not usetex, just plain mathtext. However, thanks to those links, I've come across #5414 which claimed to fix the wobbly baseline. It turns out that that isn't due to changing to external FreeType, but to a change in MathtextBackendAgg.render_glyph. Copying that out fixes all the wobbly baselines, both before and after updating FreeType.

Unfortunately, updating FreeType still causes changes in those test images due to different kerning and metrics (the whole line may move up/down a pixel, but individual glyphs no longer do.) So I'm going to roll that change in here for the changed images to be squashed together.

I've now rebased this on top of #29827, and gone through and removed the extra image changes. Now it's about 60% the number it was before.

As I expect this will be squashed, there are a few things here:

1. Fixing the wobbly baseline in mathtext (e.g., for the Latin alphabet, the t is no longer floating, while for the Greek alphabet, everything shifts down to match theta).
2. Updating to FreeType 2.13.3.
3. Removing the kerning_factor setting from the tests (the majority of these are things like Te, te, x, er all kerning a little bit closer.)
4. Removing any compatibility styling flags on images that were already regenerated.
5. Compressing everything that was already changed with oxipng; this compresses from about 12M to 7M.

Are the new tests also using hinting_factor=1?

No, as I mentioned earlier, that barely changed anything. I'm trying to see if that's true or whether I didn't set it up right.

I removed the test images here, but I realized that I can pre-populate the text-overhaul-figures branch with the results so that CI on this PR remains "green". I also realized that I never tested cibuildwheel, so I'm pushing another commit to fix that up, hopefully.