search

SpikeInterface / spikeinterface

A Python-based module for creating flexible and robust spike sorting pipelines.
https://spikeinterface.readthedocs.io
MIT License
532 stars 187 forks source link

Smoothing using filter_gaussian #2394

Closed TomBugnon closed 9 months ago

TomBugnon commented 10 months ago

Hi @DradeAW @alejoe91

I would like to be able to smooth or low-pass filter a recording. Bandpass recording with near-0 freq_min doesn't work in my case because I don't want to normalize across distant epochs.

I understand low-pass filtering is not supported because it is incompatible with chunking, and there is unfortunately no SmoothRecording (or the like) allowing to apply a rolling mean or median.

However if I understand properly, the GaussianFilterRecording applies convolution with a difference of gaussians to act as a bandpass filter. Currently, freq_min is enforced to be strictly positive, but it seems like it could be easily modified to support simple gaussian smoothing with freq_min=None. Do you think that would work @DradeAW ? Best

DradeAW commented 10 months ago

Yes, it would require some changes to the code but we could make freq_min and freq_max optional!

Then on this line https://github.com/SpikeInterface/spikeinterface/blob/03126a3abbd95621bac0d6f76190326d9081dba1/src/spikeinterface/preprocessing/filter_gaussian.py#L80, we simply have to put only one of them if the other one is None :)

TomBugnon commented 10 months ago

great thanks, I'll give it a shot!

Also from what I'm seeing so far, get_traces seems much much slower than for the regular bandpassfilterrecording... is that really what it takes?

DradeAW commented 10 months ago

Yes Gaussian filtering is inherently slower than butterworth or bessel filtering.

My implementation is actually a lot faster than the scipy implementation from the testing I did.

zm711 commented 9 months ago

Since the GaussianFilter was merged I'll close this as completed by #2397.