Currently minibatches require use of d3p.minibatch routines. The more canonical way in (num)pyro is to use the plate primitive to handle minibatch sampling - however, this is currently not possible in the DPSGD pipeline. Make it so.
This will require some way to pass d3p.random state through the inference into numpyro stacks and is hence not trivial.
Currently minibatches require use of d3p.minibatch routines. The more canonical way in (num)pyro is to use the
plateprimitive to handle minibatch sampling - however, this is currently not possible in the DPSGD pipeline. Make it so.This will require some way to pass d3p.random state through the inference into numpyro stacks and is hence not trivial.