Peter Melchior

I'm surprised, but it's conceivable that the k-Means initialization isn't great in higher-D settings. Can you check if there are crazy values in the component parameters when you only initialize...

I remain confused. The original problem arose in 2D, the k-Mean problem in 7D. These are widely different settings. Are you saying that even in the 7D data set, the...

You can chose any existing (options are `'minmax','random','kmeans','none'`) or custom initialization. If the 7D histogram shows you the location and/or size of specific features, you could turn that into an...

Hello William. Are you asking about why that selection `U_k = np.flatnonzero(indices)` is done, or do you have trouble that is caused by this line?

I'm not sure how this can happen. The relevant part of the code is this: https://github.com/pmelchior/pygmmis/blob/87ad02dd607896205ccde3ca668971c6dcacd026/pygmmis.py#L993-L1001 The array `indices` is of type bool, and the first step of setting `U_k`...

Do you have an elementwise l1 penalty? If so, you can use `operators.prox_soft`. `adaprox` should work then.

Suggestion from Robert was to take care of (some) masked pixels on our end with the following procedure: * keep the weights of masked pixels at their nominal (nonzero) value...

In addition, we should canonize the use of deconvolution coadds for initialization of all sources bright enough to pull it off. For everything else, the new compact initialization from #204...

I think a modularization is sensible but not trivial. I'm also not sure who many people would use it. The problem I see is that initialization of multiple parameters are...

After review I think that modularizing the init methods is doable. I'll describe this in another issue. As for the detection coadd, here's my proposal.: * We compute the deconvolved...