Production

[Joshuaalbert]

These tasks need to be completed to bring this to production.

Releasing current functionality

• Speed up DDTEC inference with analytic hessian and jacobian, as well as faster HMM code.

• Add sanity checks for finite values at all stages

• Single h5parms per step isolated to own working directories

• diagnostic plots per step in working directory

• Better outlier detection

• Solving systematic inference from Jones

Enabling LBA

• Fast tomographic code

Improvements

• Full screen + systematic solve directly on Jones scalars (wrapping problem). Use independent solve as starting point.

• Compute covariance only when directions significantly change. Reuse same covariance on all timesteps where covariance is similar for better marginalisation. Then grid search hyper parameters.

• Inference directly on visibilities