Errors when nesting Overlays and DynamicMaps

The stacktraces below were from master (as of last night), but I first ran into all these problems on the latest stable.

Following up on some questions I asked on gitter. I start with a analysis function that dynamically generates holoviews objects, possibly Overlays/Layouts, that include large Images and/or Curves. My goal is to regrid all Images (and datashade all Curves). For the following minimal test case, I precompute these holoviews plots and store them in dicts (dat1 and dat2); in my production code, these may be cached in a dict or dynamically generated (so z needs to be a kdim of a DynamicMap, not a HoloMap).

Set up the plots:

import numpy as np
import holoviews as hv
from holoviews.operation.datashader import datashade, regrid
hv.notebook_extension('bokeh')
dat1 = {}
dat2 = {}
for z in range(10):
    dat1[z] = hv.Image(np.random.random((1000,1000)), bounds=(0,0,10,10)) * hv.HLine(z)
    dat2[z] = hv.Curve((np.random.random(1000) - 0.5).cumsum()) * hv.VLine(z*100)

With no datashader, this works as expected:

def cb(z):
    return dat1[z]
dmap = hv.DynamicMap(cb, kdims=['z']).redim.values(z=range(10))
dmap

It shows a slider for z, and I can easily browse around.

It would be great to be able to write dmap.map(regrid, hv.Image) and have that work. Maybe one day? 😃 I assumed I needed to call .map() before I combine all the z-slices into a DynamicMap, like this:

def cb(z):
    return dat1[z].map(regrid, hv.Image)
dmap = hv.DynamicMap(cb, kdims=['z']).redim.values(z=range(10))
dmap

This fails. Stacktrace: https://gist.github.com/shenker/16b66fda15093544e9d86d69a60a12af

Heeding the warning message about .collate(), I try:

def cb(z):
    return dat1[z].map(regrid, hv.Image).collate()
dmap = hv.DynamicMap(cb, kdims=['z']).redim.values(z=range(10))
dmap

This also fails, with the same error. With either of these, if instead of displaying dmap, I display, e.g., dmap[4], I get an individual image and the dynamic regridding works just fine (obviously, no scroll bar for z).

I was wondering if this issue was specific to hv.Image, so I also tried the above substituting dat1 with dat2 and regrid with datashade, but this failed in similar ways.

Returning to the original working example without datashader,

def cb(z):
    return dat1[z]
dmap = hv.DynamicMap(cb, kdims=['z']).redim.values(z=range(10))

If I try simply dmap.collate(), I get this stacktrace when I use dat1 (with Images) and this stacktrace when I use dat2 (with Curves).

[philippjfr]

If you’re working with a small set of elements and declaring them ahead of time I’d recommend just using a HoloMap btw. Here’s how I’d write your example:

dat1, hlines = {}, {}
for z in range(10):
    dat1[z] = hv.Image(np.random.random((1000,1000)), bounds=(0,0,10,10))
    hlines[z] = hv.HLine(z)
    
regrid(hv.HoloMap(dat1, 'z')) * hv.HoloMap(hlines, 'z')

[jbednar]

The main issue here is that the regrid operation is dynamic by default which allows it to respond to zoom events on the axes, but you cannot nest DynamicMaps.

Would it be possible to detect nested DynamicMaps and raise an error with your suggestions above? It sounds like the current behavior is very confusing, i.e. that it gives a blank plot rather than some more direct indication that something that seems reasonable is not supported.