Remove LevMarLSQFitter #1899
Remove LevMarLSQFitter #1899
Conversation
| @@ -105,11 +105,13 @@ def centroid_1dg(data, error=None, mask=None): | |||
|
|
|||
| xy_data = [np.ma.sum(data, axis=i).data for i in (0, 1)] | |||
|
|
|||
| # would need to use a different fitter if parameter bounds are used | |||
| fitter = LMLSQFitter() | |||
There was a problem hiding this comment.
@larrybradley - gaussians have bounds enabled by default (on stddev) so if you are fitting a Gaussian I think it would be best to use TRFLSQFitter or another one, not LMLSQFitter
There was a problem hiding this comment.
@astrofrog Are you sure about that? Where are the bounds defined? I don't see any bounds defined in functional_models? Do you mean constrained to be postive? I still get negative stddev's sometimes.....
There was a problem hiding this comment.
I've also noticed that TRFLSQFitter is about 4-5% slower than LevMarLSQFitter even without any bounds.
There was a problem hiding this comment.
@larrybradley yes the fact the stddev is required to be positive - it's formally treated as a bound.
There was a problem hiding this comment.
There was a problem hiding this comment.
Do you see negative stddev values with LMLSQ or TRFLSQ or both?
There was a problem hiding this comment.
I've also noticed that
TRFLSQFitteris about 4-5% slower thanLevMarLSQFittereven without any bounds.
Luckily the fitters will be significantly faster with astropy 7.0.0 so should be more than enough to compensate for this!
There was a problem hiding this comment.
Ah, I see that's only for Gaussian1D. I was seeing negative stddev when fitting Gaussian2D. Why is Gaussian1D stddev bounded, but not Gaussian2D? Should I open a PR?
There was a problem hiding this comment.
This PR changes the default Astropy fitter for
PSFPhotometry,IterativePSFPhotometry, andEPSFFitterfromLevMarLSQFittertoTRFLSQFitter.LevMarLSQFitteruses the Levenberg-Marquardt algorithm via the SciPy legacy functionscipy.optimize.leastsq, which is no longer recommended. This fitter supports parameter bounds using an unsophisticated min/max condition where parameters that are out of bounds are simply reset to the min or max of the bounds during each step. This can cause parameters to stick to one of the bounds during the fitting process if the parameter gets close to the bound. If needed, this fitter can still be used by explicitly setting the fitter in thePSFPhotometry,IterativePSFPhotometry, andEPSFFitterclasses.The fitter used in
RadialProfileto fit the profile with a Gaussian was also changed fromLevMarLSQFittertoTRFLSQFitter.The fitter used in
centroid_1dgandcentroid_2dgwas changed fromLevMarLSQFittertoLMLSQFitter.See astropy/astropy#16983 for more information.
Closes: #1895