Multiscale 3D rendering warning should happen once only

[jni]

When I switch to 3D view in a multiscale array, I get the following warning:

/Users/jni/projects/napari/napari/layers/image/_slice.py:113: UserWarning: Multiscale rendering is only supported in 2D. In 3D, only the lowest resolution scale is displayed!

This is fine, except the warning happens on every slice event:

...
/Users/jni/projects/napari/napari/layers/image/_slice.py:113: UserWarning: Multiscale rendering is only supported in 2D. In 3D, only the lowest resolution scale is displayed!
/Users/jni/projects/napari/napari/layers/image/_slice.py:113: UserWarning: Multiscale rendering is only supported in 2D. In 3D, only the lowest resolution scale is displayed!
/Users/jni/projects/napari/napari/layers/image/_slice.py:113: UserWarning: Multiscale rendering is only supported in 2D. In 3D, only the lowest resolution scale is displayed!
/Users/jni/projects/napari/napari/layers/image/_slice.py:113: UserWarning: Multiscale rendering is only supported in 2D. In 3D, only the lowest resolution scale is displayed!
...

That's a fraction of the console output, and the toast pop-up on the canvas just keeps popping as you interact with the viewer.

@kephale is making good progress to help us get rid of that warning altogether, but in the meantime, a quick fix would be to set a flag for whether we've warned already, and not warn if so.

[andy-sweet]

I was optimistic that this could be fixed with warnings.simplefilter('once') and maybe warnings.catch_warnings, but that doesn't seem to have effect here.

I then noticed filterwarnings in pyproject.toml, which attempts to ignore all warnings matching this message. I'm not sure if that would have an effect in editable installs, but I could also reproduce this issue on my non-editable v0.4.17 install, so it doesn't seem to have an effect at all - or am I misinterpreting standard warning filtering in Python? I also wonder if napari is handling warnings in a special way that means these get ignored?

We could of course fix this by storing a static boolean that gets set after the first warning, but that feels silly given there are built-in mechanisms for handling this stuff.

I'm going to unassign myself for now and hope someone with more background/knowledge steps in here.

[andy-sweet]

I then noticed filterwarnings in pyproject.toml, which attempts to ignore all warnings matching this message. I'm not sure if that would have an effect in editable installs, but I could also reproduce this issue on my non-editable v0.4.17 install, so it doesn't seem to have an effect at all - or am I misinterpreting standard warning filtering in Python?

From community meeting debugging, I am misinterpreting because that filterwarnings only applies to tests. So one mystery solved.

[andy-sweet]

I couldn't resist picking this one back and I think this is actually a Python bug: python/cpython#73858

If I understand correctly, the problem is that on exiting catch_warnings forgets about any previously seen warnings, which is why we get repeats even though the 'default' action is to show once. The tiny minimal reproducing example there shows this off:

import warnings
for _ in range(3):
    with warnings.catch_warnings():
        pass
    warnings.warn('warning')

In our case, the troublemaker is (of course...) the thumbnail code which uses catch_warnings twice:

• Image._update_thumbnail
• Image.thumbnail setter

To fix this specific issue (i.e. when going through the slicing path), we might be able to remove the first usage (as we require scipy >= 1.4) and we may also be able to remove the second one if we use the scikit-image utility and that doesn't issue a warning.

But the general problem is that any other use of catch_warnings that gets hit while using napari will cause it to forget, so the issue may get hit again (though likely not on every slice/refresh).

[andy-sweet]

I think this one is back. Serves me right for not writing an explicit test... Here's one that can reproduce the issue:

def test_3d_slice_twice_with_multiscale_image_warns_once():
    data = [np.zeros((4, 4, 4, 4)), np.zeros((2, 2, 2, 2))]
    layer = Image(data, multiscale=True, rgb=False)

    with pytest.warns(
        UserWarning, match="Multiscale rendering is only supported in 2D"
    ):
        layer._slice_dims(point=(0, 0, 0, 0), ndisplay=3)

    layer._slice_dims(point=(1, 0, 0, 0), ndisplay=3)

and another that covers more code:

def test_3d_slice_twice_with_multiscale_image_warns_once():
    # Start in 2D to avoid warning when adding the layer.
    viewer = ViewerModel(ndisplay=2)
    data = [np.zeros((4, 4, 4, 4)), np.zeros((2, 2, 2, 2))]
    viewer.add_image(data, multiscale=True, rgb=False)

    # Switch to 3D to trigger the warning.
    with pytest.warns(
        UserWarning, match="Multiscale rendering is only supported in 2D"
    ):
        viewer.dims.ndisplay = 3

    # Simulate moving the slider to slice again, without a warning.
    assert viewer.dims.point[0] != 0
    viewer.dims.point = (0, 0, 0, 0)

You'll also have to remove the related ignore line in pyproject.toml.

My assumption is that there was a catch_warnings added after #5603.

Unless the underlying cpython issue is fixed and backported, it seems like the only fix here is to forbid use of catch_warnings in that code path.

This is an issue on main and for v0.4.18 unfortunately.