Added check for NDData.uncertainty being StdDevUncertainty array in construction of EPSFStar object #739
Conversation
Onoddil
force-pushed
the
nddata_uncertainty_description
branch
from
940f0a7
to
68b13fe
Compare
| # the weight. Before assigning the weights we must ignore zero | ||
| # variance data, however, setting the weights for these data points to | ||
| # zero by forcing their standard deviations to a large value. | ||
| elif data.uncertainty.uncertainty_type == 'std': |
There was a problem hiding this comment.
Note that in the upcoming astropy release (3.1) it looks to me like there will be an easier way to "normalize" other uncertainties into std. We'll want to do that for this option I'd think, but no point in doing that until the astropy release is available for testing.
There was a problem hiding this comment.
So is this a "accept this and we'll revisit it later" or a "hold off on this until after the release is available for testing"? I think it would be good in the future to expand this quite a fair amount to accept various kinds of uncertainty (perhaps parsing them through some convenience function to produce an internally accepted 'standard' type), and it sounds like this will help with that down the line.
There was a problem hiding this comment.
Astropy 3.1 is out, so I've assigned @mwcraig to this issue in case he can help with this :)
There was a problem hiding this comment.
Right now what you could is something like this:
uncertainty_conversion_to_std = {
'std': lambda x: x, # do nothing
'var': np.sqrt, # square root of the variance is the standard dev
'ivar': lambda x: 1 / np.sqrt(x) # invert and square root to get standard dev
}
uncert = data.uncertainty.array.copy()
std_dev = uncertainty_conversion_to_std[data.uncertainty.type](uncert)(note: I haven't tested this code but hopefully it conveys the idea).
… creating EPSFStar instance via extract_stars. Also added description comment to EPSFFitter.fit_star detailing where fitter weights originate.
Co-Authored-By: Onoddil <[email protected]>
Onoddil
force-pushed
the
nddata_uncertainty_description
branch
from
48f1964
to
19bbc6d
Compare
|
Have made the changes requested @eteq - I note however that astropy 3.1 has been released subsequently and therefore the new changes to uncertainties must be updated; I think it makes sense to merge this PR as a base to build the new functionality for std normalisation onto, but if that should be built into this PR before acceptance then let me know. |
Added the option for the weights in
EPSFStarobjects to be passed fromStdDevUncertaintyNDData arrays, extending from the originalweightsuncertainty_typeoption. Also added brief comment describing the origin ofweightsin EPSFFitterfitter. Closes #734