width/height not taken into account in a `regrid` operation with the Bokeh backend

The regrid operation does not seem to take into account width and height when using the Bokeh backend. The example shows that it does work with the Matplotlib backend. Notice in the GIF how for a very short moment the Bokeh plot seems to have the right grid, and then updates to the original image resolution apparently.

Here’s the full code:

import numpy as np, holoviews as hv
from holoviews.operation.datashader import regrid
hv.extension('bokeh', 'matplotlib')

arr = np.random.rand(200, 200)
img = hv.Image(arr)
regrid(img, width=10, height=10)

hv.output(backend='matplotlib')
regrid(img, width=10, height=10)

[maximlt]

We have spent some time looking into that. The base ResamplingOperation - from which regrid, rasterize and others inherit - has parameters width and height and also sets up by default the RangeXY and PlotSize streams.

https://github.com/holoviz/holoviews/blob/e5224c6a278829ca6165ac04b36da6d5f55825cf/holoviews/operation/datashader.py#L73-L77

https://github.com/holoviz/holoviews/blob/e5224c6a278829ca6165ac04b36da6d5f55825cf/holoviews/operation/datashader.py#L105-L107

The parameters of the PlotSize streams are also called width and height.

https://github.com/holoviz/holoviews/blob/e5224c6a278829ca6165ac04b36da6d5f55825cf/holoviews/streams.py#L1424-L1431

Which means that when the PlotSize stream is triggered - after rendering - the regrid operation is called again this time with width and height obtained from the stream and as such different from the values provided to regrid by the user.

I’ll offer now solutions for different use cases and ask some questions for how to improve the current state:

1. you want regridding to be based on the number of pixels your plot output uses, let HoloViews compute it automatically with regrid(img)
2. you want to perform static/fixed regridding, i.e. on zoom/pan events no regridding should be performed again, then use dynamic=False: regrid(img, width=10, height=10, dynamic=False)
3. you want a mix of 1 and 2 with regridding being performed on zoom/pan events but with dimensions that you provide instead of dimensions being inferred by HoloViews (what this issue is about), override streams so that the operation doesn’t set the PlotSize stream: regrid(img, width=10, height=10, streams=[hv.streams.RangeXY])

3 in the list above feels more like a workaround for a bug rather than a clean solution. We played around with a potential bug fix that consists in removing the PlotSize stream if width or height is provided:

class ResamplingOperation(LinkableOperation):

    ...

    def __call__(self, *args, **kwargs):
        if 'width' in kwargs or 'height' in kwargs:
            print('warning: dropping PlotSize stream using explicit width/height')
            # What if values specified by kwargs are supposed to be the initial stream values?
            if PlotSize in self.streams and isinstance(self.streams, list):  # Also needs to handle dict
                streams = self.streams.copy()
                streams.remove(PlotSize)
                kwargs['streams'] = streams
        return super().__call__(*args, **kwargs)

    ...

We’d really appreciate to hear your thoughts on that @philippjfr:

• Did we identify the right problem? I.e. that the operation and stream share the same namespace which means they end up being linked? Is that by design?
• Are we heading in the right direction with our approach to fix this?

cc @jlstevens @Hoxbro @droumis

[philippjfr]

Seems reasonable.