MouseLand
Run time and cluster detection changes in v4.0.2 vs v4.1.0
I recently upgraded KS4 to version v4.1.0 and was previously using version v.4.0.2. I compared the KS4 output on the same data between these two versions and could use some assistance in trying to understand some differences in the output.
I ran the same data file on version v4.1.0 (log name: kilosort4_v4.1.0_default) and v4.0.2 (log name: kilosort4_v4.0.2) using the default settings, and then a second time on v4.1.0 (log name: kilosort4_4.1.0_dschange) changing the cluster down sampling to 20 to match the settings from the v.4.0.2 run.
Differences I could use help understanding:
-
The run time is much longer when using v4.1.0 compared to v4.0.2 with default settings and increases dramatically in v4.1.0 when changing the cluster down sampling value to 20.
-
The number of spikes and clusters detected on v4.1.0 default settings is much higher but translates to noise rather than good units.
-
The difference in spatial features from both runs on v4.1.0 compared to the run on v.4.0.2.
Hello,
We're still looking at this. In the meantime, can you please try re-running the last version you posted (with cluster_downsampling = 20)? It's really strange that that would affect the amount of time spent on spike detection and drift correction so drastically, since that parameter doesn't interact with those steps in any way. I'd like to rule out a hardware fluke.
Thank you, I reran the last version I posted, and the run time improved drastically, so must have been a hardware fluke. See log here.
kilosort4_v.1.0_dschange_rerun.log
Just wanted to add that I've been playing with different versions since it's been a while since I've updated, and these extended run time and changes in spatial features/cluster detection persist through v4.0.37.