Lasso select does not work on datashaded NdOverlay

[peterroelants]

Description of expected behavior and the observed behavior

Using the lasso selection tool on a NdOverlay that is datashade-ed does not work and results in the following message being printed:

linked_selection aborted because it could not display selection for all elements: ufunc 'invert' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''.

Complete, minimal, self-contained example code that reproduces the issue

This code

import numpy as np
import holoviews as hv
import datashader
from holoviews.operation.datashader import datashade, shade, dynspread, spread, rasterize

hv.extension('bokeh')

points_dct = {
    'p1': hv.Points(np.random.randn(10000, 2), kdims=['x', 'y'], label='p1'),
    'p2': hv.Points(np.random.randn(10000, 2) * 2 + 1, kdims=['x', 'y'], label='p2')
}
nd_overlay = hv.NdOverlay(points_dct, kdims=['cat'])
nd_overlay_shaded = spread(
    datashade(
        nd_overlay, 
        aggregator=datashader.by('cat', datashader.count()), cnorm="eq_hist")
    , px=1
)
hv.selection.link_selections(nd_overlay_shaded)

And using the lasso selection results in:

Please note that the box selection works, as well as the lasso-select on a datashaded plot that consists of just points (no NdOverlay.

ALL software version info

Python implementation: CPython
Python version       : 3.9.6
IPython version      : 7.25.0

Compiler    : GCC 9.3.0
OS          : Linux
Release     : 5.10.0-1038-oem
Machine     : x86_64
Processor   : x86_64
CPU cores   : 16
Architecture: 64bit

numpy     : 1.21.1
bokeh     : 2.3.3
datashader: 0.13.0
holoviews : 1.14.5

jupyter core     : 4.7.1
jupyter-notebook : 6.4.0
qtconsole        : not installed
ipython          : 7.25.0
ipykernel        : 6.0.3
jupyter client   : 6.1.12
jupyter lab      : 3.0.16
nbconvert        : 6.1.0
ipywidgets       : 7.6.3
nbformat         : 5.1.3
traitlets        : 5.0.5

[jlstevens]

Thanks for reporting this!

I can confirm the problem with your minimal example. I think the issue is not directly due to it being an NdOverlay but due to datashade being necessary which returns RGB images (unlike rasterize). Regardless, linked_selections should be able to support this case so I'll mark this as a bug.