These are chat archives for jeff-regier/Celeste.jl

16th
Feb 2016
Kyle Barbary
@kbarbary
Feb 16 2016 06:32
ModelInit.initialize_celeste() is a bottleneck for me right now. The PSF fit is taking a pretty long time (at least several minutes per tiled image). Also, it doesn't help that all the printing from GaussianMixtures makes it pretty hard to see what's going on. Is that about your experience? If so, any tips?
Naively, it seems like one should not have to fit the PSF so many times. We know it's smooth across the field.
Jeffrey Regier
@jeff-regier
Feb 16 2016 06:37
It varies smoothly, but it's different at every source. I suppose we could sacrifice some accuracy and just fit it, say, 32 times per frame, like on a 4 x 8 grid, and then use the closest one for each source.
Kyle Barbary
@kbarbary
Feb 16 2016 06:38
Yeah thats sort of what I was thinking. You could linearly interpolate the Gaussian parameters.
Jeffrey Regier
@jeff-regier
Feb 16 2016 06:39
That should be fine, and we may need to redo our psf model ultimately anyway
Maybe just use the nearest psf...interpolating gaussian mixture params doesn't seem so straightforward, because it's a mixture, not just one Gaussian
Kyle Barbary
@kbarbary
Feb 16 2016 06:44
hmmm... good point. I was thinking that the 3 gaussians always corresponded to the same components (so that each Gaussian's parameters would vary smoothly), but I guess that is not necessarily the case.
This is your experience though, that initialize_celeste() takes many (at least 10) minutes?
Jeffrey Regier
@jeff-regier
Feb 16 2016 06:47
We may need to replace our psf model enitrely anyway after seeing the results, so the nearest neighbor thing is probably good enough for now
Kyle Barbary
@kbarbary
Feb 16 2016 06:47
What's the expected runtime of the actual fitting step, per source, btw?
Jeffrey Regier
@jeff-regier
Feb 16 2016 06:49
Yes, I think that's about what Ryan was reporting. We only put up with that because until recent processing was even slower than that preprocessing
Kyle Barbary
@kbarbary
Feb 16 2016 07:06
an interim alternative would be to just move the PSF fit further down the pipeline: do it for each source when the source is fit. (It wouldn't save time overall, but it would be easier to iterate on just a few sources). Or is there some reason it has to be done in model initialization?
Jeffrey Regier
@jeff-regier
Feb 16 2016 07:24
That sounds like a good idea too