Francois Lanusse

Small addition to the previous post, I have added inverse variance weights in the Observation and removed in the code the arbitrary scaling factor we had on the prior term....

@esheldon Great that you got it to work. Another important factor here is the step size used with the optimizer. At 1e-3 optim is unstable, at 1e-4 it works much...

Thanks for this @fred3m , So I think the pixel resolution is not going to make a lot of difference, there is a big overhead in running the tensorflow session,...

Almost done addressing the issue linked above. Demo notebook of training a pixel cnn from cosmos images at HSC resolution using distributed cloud TPU training :-) https://colab.research.google.com/drive/1p-DOHO6aU1ToPG4lca_pjHyHMt7sxs5l

So, after a weekend of experimentation on multiband images here is a list of things that are not yet perfect: - Relative normalization of SED and morphology: the default `morph_max`...

Yes yes yes, and I need you help @herjy :-) I *tried* to build the training set at the HSC pixel resolution, with a standardized effective, PSF, so that all...

@googlebot I signed it!

Pinging some of the usual suspects @tlmakinen @lavaux

Awesome :-) I thought that might have been the case, would you be up for merging any additional updates with this branch so that we can get this repo up...

ok.... well.... either it's black magic, or there is something I don't understand, but in any case my mind is blown.... ```python mesh_shape = (2,) # On 2 GPUs devices...