Setup time for neuronal models

[kirkw-1]

What is the expected setup time for neuronal models? Could we get a direction comparison say with LFPy setup for one of the preprint paper models? I'm building a model in jaxley which I also have in HOC code and setup time is far longer than LFPy, which I believe is to be expected, but would be good to have a benchmark to know if my setup code is efficient.

[kirkw-1]

(31 param, 196 branch single cell model setup+single simulation time 6mins45s in jaxley, versus 18s with"python/LFPy/NEURON"

[michaeldeistler]

Here is the morphology (SWC file) I am using:
morph_l5pc_with_axon.txt

[kirkw-1]

Great, thanks that's very helpful. Yes your example runs in similar time on my hardware.
Intel(R) Core(TM) i9-10940X CPU @ 3.30GHz, NVIDIA GeForce RTX 4070 Ti
Time to initialize the model: 34.42545199394226
Time to simulate: 14.764939308166504

There were 2 issues with my setup.
(1) (resolved) .set_ncomp() within a loop, we should set initialize=False then do a single initialise at end of loop. (took >200secs)
(2) (not resolved) .distance within a loop.
max_length = 0
for b in cell.apical.branches:
dist = b.loc(0.0).distance(cell.branch(0).loc(0.0))
if dist > max_length:
max_length = dist
I use this for setting channel param according to distance from soma. It takes 45secs.
This changelog entry [https://github.com//pull/603] refers to .distance with JIT so I experimented with @jit around the function but couldn't seem to get cell.apical.branches to work with JIT "because it's a generator" so I tried list(cell.apical.branches) etc. but no luck yet.

[michaeldeistler]

1. Great that this is resolved!
2. The newest Jaxley version (v0.10.0) changed how distance should be computed. See also here. The following should solve your issues (when using v0.10.0):

from jaxley.morphology import distance_pathwise, distance_direct
path_dists = distance_pathwise(cell.soma.branch(0).comp(0), cell)
cell.nodes["path_dist_from_soma"] = path_dists

Run this before the simulation, and then use cell.nodes["path_dist_from_soma"] to specify the parameters. I hope this makes sense, if not then please send me the full code snippet that uses max_length in your code.

Michael

[kirkw-1]

Thanks, yes it's much faster now. This code takes 6s and actual param setting 10s. My overall setup time is still high +65s on both CPU/GPU, but can't currently see any other candidates for bottlenecks so will come back to this later.

[michaeldeistler]

Alright, thanks for the update! Note that you can also save and load the cell, as explained here.