Possible to use decimate on hv.Spikes object?

HoloViews
1

Hi all!

I’m looking to optimize some code for plotting a very large set of neuronal spike times, and have found the hv.Spikes function to work very well. I want to try using hv.operation.decimate on top of the Spike object, but haven’t been able to yet.

Here is a simplified version that illustrates the issue:

data = np.random.randint(0, 100, 1000)
raster1 = hv.Spikes(data, kdims='Time')

x_r = (0, 100)
y_r = (0, 0.5)
decimate(raster1)

Which produces the error: WARNING:param.dynamic_operation: Callable raised "ValueError('not enough values to unpack (expected 2, got 1)')". Invoked as dynamic_operation(x_range=None, y_range=None)

Does anyone have an idea about how best to combine these two features, if possible? Thanks!

2

We should indeed improve decimate, but for now you can simply roll your own version like this:

data = np.random.randn(100000)
raster1 = hv.Spikes(data, kdims='Time')

x_range = hv.streams.RangeX(x_range=raster1.range(0))

def subsample(spikes, x_range, n=1000):
    sliced = spikes.select(Time=x_range)
    count = len(sliced)
    if count <= n:
        return sliced
    return sliced.iloc[np.random.randint(0, count, n)]

raster1.apply(subsample, streams=[x_range])
1 Like
3

This is very helpful, thank you!

One of the decimate functions that is really useful is that the plotting adjusts with zooming, so that if the resolution/downsampled points are gradually recovered with zooming in.

Would approach this by converting the raster plot into a HoloMap/DynamicMap, or is it simpler to add a while loop into the subsample function?

4

@barnold The function I supplied together with .apply and the RangeX stream should do exactly that too. When you zoom it’ll reveal more of spikes.

5

@philippjfr ah okay. in the toy code above, it’s difficult to tell how much the density changes. I implemented another toy version using your snippet and the NdOverlay objects to get closer to a final data version and found that the spike reveal feature isn’t present there.

So with a slightly more complicated initial raster:

raster = hv.NdOverlay({i: hv.Spikes(np.random.randint(0, 100, 1000), kdims='Time').opts(position=0.1*i)
                   for i in range(20)}).opts(yticks=[((i+1)*0.1-0.05, i) for i in range(20)]).opts(
    opts.Spikes(spike_length=0.1),
    opts.NdOverlay(show_legend=False))
raster

and an adjusted version of your examples:

def raster_subsample(raster, x_range, n=100):
    new_spikes_objects = []
    for spikes in raster:
        sliced = spikes.select(Time=x_range)
        count = len(sliced)
        if count <= n:
            new_spikes_objects.append(sliced)
            continue
        new_spikes_objects.append(sliced.iloc[np.random.randint(0, count, n)])
    
    new_overlay = hv.NdOverlay({j: new_spikes_objects[j] for j in     range(len(new_spikes_objects))}).opts(yticks=[((j+1)*0.1-0.05, j) for j in range(20)]).opts(
        opts.Spikes(spike_length=0.1),
        opts.NdOverlay(show_legend=False))

    return new_overlay

x_span = (0, 100)
x_range = hv.streams.RangeX(x_range=x_span) 

raster.apply(raster_subsample, streams=[x_range])

I manually redefine the x range here, otherwise only the index of each hv.Spike object is pulled.

image
image852×400 88 KB

In this case, the total number of spikes that are plotted remains constant with zooming (right raster, versus original on the left). Is there a more “holoviews” way of defining the x-range that might fix this?

Many thanks for your time and input!