Calibrate :: (Visibilties, SourceModel, GainModel) -> GainModel

[Joshuaalbert]

Direction-dependent calibration on visibilities.

How:

• We want to use an optimisation algorithm on the DFT predict model.
• Since the data contains faint sources as well as bright sources, we need to pre-transform the data so that DFT is the best approximation.
• We start by subtracting FFTPredict of faint sky from data, given the best estimate of the gain model (ionosphere, dish defects, perfect beam).
• Then this residual should contain bright source data, with artifacts from faint subtract. These will not be significant.
• We then do distributed LM or LBFGS calibration on these residuals shared by data locality (freq,time over devices)

[Joshuaalbert]

Need to force sharding of gradients to remain over frequnecy until an all gather for total gradients is needed.

[Joshuaalbert]

Loaded LWA data, but need to extend our data columns to handle multipl columns too. Quickly