Drastic decrease in performance for matplotlib plotting

I found the cause of the bug, shown in a couple of message down. Merging the fix PR and closing this issue.

I am experience a drastic decrease in performance when plotting shapes, observed on the Visium HD notebook.

The notebook from the docs (which is using either the latest release code but maybe also some earlier code), took ~10 seconds to make this plot, for ~90000 shapes:

With the latest main, using method='matplotlib' didn't finish in one hour. Here is the code

%%time
plt.figure(figsize=(10, 10))
ax = plt.gca()

gene_name = "AA986860"
sdata.pl.render_shapes("Visium_HD_Mouse_Small_Intestine_square_016um", color=gene_name, method="matplotlib").pl.show(
    coordinate_systems="global", ax=ax
)

I have therefore used datashader, which took ~40 seconds (still worse than the original code, but I reported this in a separate issue #296). Here I would put the focus on finding out why matplotlib based plotting got so slow.

A final note, I am setting up a codspeed GitHub Action so in the future benchmarks will be automatically run at each commit; I am working on this in a draft PR #295.

Here is the viz from a .pstat benchmark file collected by running the test (recorded while executing 'method'='matplotlib')

@pytest.mark.parametrize("method", ["matplotlib", "datashader"])
@pytest.mark.benchmark
def test_plot_can_render_large_polygons(method: str):
    sdata.pl.render_shapes(element="polygons", method=method).pl.show()

(test available from the PR #295).

It seems that the problem are the calls to add_collection() from spatialdata_plot.pl.render._render_shapes()

    elif method == "matplotlib":
        _cax = _get_collection_shape(
            shapes=shapes,
            s=render_params.scale,
            c=color_vector,
            render_params=render_params,
            rasterized=sc_settings._vector_friendly,
            cmap=render_params.cmap_params.cmap,
            norm=norm,
            fill_alpha=render_params.fill_alpha,
            outline_alpha=render_params.outline_alpha,
            zorder=render_params.zorder,
            # **kwargs,
        )
        cax = ax.add_collection(_cax)

        # Transform the paths in PatchCollection
        for path in _cax.get_paths():
            path.vertices = trans.transform(path.vertices)
            cax = ax.add_collection(_cax)

Any comment?

Removing the line cax = ax.add_collection(_cax) fixes the issue; it was introduced in this PR here: https://github.com/scverse/spatialdata-plot/pull/244/files#diff-eb3d4aa0a6246689774651947ba2b351d639ab0f0585b67de070a3c33f2e92f4.

It seems to be just a typo after a copy-paste.

Navigation Menu

Repository navigation

Drastic decrease in performance for matplotlib plotting #297

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Drastic decrease in performance for matplotlib plotting #297

Description

Activity

LucaMarconato commented on Jul 13, 2024

LucaMarconato commented on Jul 13, 2024

LucaMarconato commented on Jul 13, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions