split_axon_dendrite not generating errors and just quiting about navis

[schlegelp]

The APL is a monster of a neuron. When I ran this code it effectively stopped at print(navis.split_axon_dendrite(on)) and started consuming obscene amounts of memory. There must be a bottleneck in the code that simply doesn't scale well. Will have a proper look and let you know. FWIW: smaller neurons should work fine.

[clbarnes]

How big is the neuron when serialised? Might be a useful stress test case to keep around.

[schlegelp]

This APL is ~12Mb pickled.

I've also found the culprit for the @dadanhan's problem: an inefficient path query in navis.flow_centrality (which is called by split_axon_dendrite. Should be able to make some improvements on that end soon.

[clbarnes]

One day I'll fix my rust implementation of Arbor.js... I've been playing around with some linestring smoothing and simplification stuff too.

[schlegelp]

One day I'll fix my rust implementation of Arbor.js... I've been playing around with some linestring smoothing and simplification stuff too.

Oh cool! Slightly parallel effort with what I started here: https://github.com/navis-org/fastcore

[schlegelp]

@dadanhan I pushed a couple changes to the main branch that speed up the flow metrics functions. If you install navis from Github your above code should go through albeit still not super fast: navis.split_axon_dendrite(on, metric='flow_centrality') takes ~7min on my laptop... I would avoid calling it multiple times.

A note on how split_axon_dendrite works: you can play around with the flow_thresh parameter to improve the split. If you do that you can avoid having to re-calculate the flow every time you run the function by calculate it beforehand:

>>> # From your above example
>>> on.reroot(on.soma, inplace=True)
>>> navis.flow_centrality(on)
>>> # `on` now has a "flow_centrality" column in the node table (check e.g. `on.nodes.head()`)
>>> # Splitting now re-uses the column 
>>> splits = navis.split_axon_dendrite(on, metric='flow_centrality', flow_thresh=0.9)

That way the actual navis.split_axon_dendrite call only takes a minutes or so.

Also not sure how well this algorithm works on something like the APL with its many parallel fibres. Here's a screenshot coloured by flow centrality (navis.plot3d(on, color_by='flow_centrality', palette='viridis'). As you can see the neurite with the highest centrality runs the entire length of the peduncle + lobes and then circles around. No clue whether that makes sense for this neuron or whether the skeleton has a messed up topology.

[dadanhan]

Thank you Philipp!