Initialising parameter values with compositional sampling in Turing.jl

Hi all,

Is there a way to set the initialisation location for samplers when using the compositional sampling method in Turing.jl? Setting the initial values when sampling from a continuous parameter space (with hamiltonian sampler only) works well by defining the “init_theta” argument in the sample() call. However, with the compositional sampling method (PG + Gibbs + HMC or NUTS) I always get an assertion error as the dimensions do not match. Is there some specific arrangement for the initial values that needs to be used with combined discrete and continuous parameters?