dereklhansen / smc.jl Goto Github PK

Similar to #9, I'd like to make the SMC code type-stable and also orient it around a model struct. That approach seems the most elegant.

Alternatively ,aybe the Kalman code should be in its own module?

This will likely entail a bit of refactoring, checking for type stability, etc. We also need to code up some stuff for the models. This is the last major code area that is not covered yet.

I'm more familiar with Github Actions now, so I'd like to get this going

One area for improvement is to not allocate memory when new particles are sampled. We currently don't do this to keep compatibility with autodiff.

The way to do this is actually at the model layer (e.g. LinearGaussian) rather than at the SMCModel layer

function rp_base(m::LinearGaussian, rng, Zs, t)
    μ, Σ = calc_prop(m, Zs, t)
    @assert isposdef(Σ)
    Zs_new = μ + rand(rng, MvNormal(Σ), size(Zs, 2))
    return Zs_new
end

turns into

function rp_base(m::LinearGaussian, rng, Zs, t)
    μ, Σ = calc_prop(m, Zs, t)
    @assert isposdef(Σ)
    eps = rand!(rng, MvNormal(Σ), m._eps_old)
    eps .+= μ
    m._eps_old = m._eps_new # swich the underlying storage
    m._eps_new = eps
    return Zs_new
end

Use JuliaFormatter automatically and test for it