Update AdvancedPS to match SSMProblems interface #93
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
| f(θ::Parameters, state, t) = Normal(θ.a * state, θ.q) | ||
| g(θ::Parameters, state, t) = Normal(0, exp(0.5 * state)) | ||
| f₀(θ::Parameters) = Normal(0, θ.q) | ||
| f₀(model::StochasticVolatilityModel) = Normal(0, model.q) |
There was a problem hiding this comment.
This feels a bit redundant with transition!!. Consider merging these into the function transition!!.
| pg = AdvancedPS.PG(Nₚ, 1.0) | ||
| chains = sample(rng, model, pg, Nₛ; progress=false); | ||
| chains = sample(rng, model, pg, Nₛ; observations=y, progress=false); |
There was a problem hiding this comment.
The main issue here is that AbstractMCMC does not forward data explicitly, it just assumes the obervables are internal to the model
There was a problem hiding this comment.
I see. One way around this would be to define something like
struct ConditionedSSM
ssm::AbstractStateSpaceModel
observations::Vector{Vector{Float64}}
end
though that's one more bit of API bloat.
There was a problem hiding this comment.
I like the idea; it is also how the condition mechanism works for DynamicPPL models.
There was a problem hiding this comment.
Doesn't have to be nested, or another type, as long as these are defined by the user:
function emission_logensity(model::AbstractStateSpaceModel, state, step::Int)
return logpdf(Normal(state), observations[step]) # Global scope
endor
function emission_logensity(model::AbstractStateSpaceModel, state, step::Int)
return logpdf(Normal(state), model.observations[step])
endIt's a bit what model() | (; y=observations) would do in DynamicPPL
| p = particles[i] | ||
|
|
||
| # Obtain ``\\log p(yₜ | y₁, …, yₜ₋₁, x₁, …, xₜ, θ₁, …, θₜ)``, or `nothing` if the | ||
| # the execution of the model is finished. | ||
| # Here ``yᵢ`` are observations, ``xᵢ`` variables of the particle filter, and | ||
| # ``θᵢ`` are variables of other samplers. | ||
| isref = p === ref | ||
| score = advance!(p, isref) | ||
| score = advance!(p, y, isref) |
There was a problem hiding this comment.
That would make backward compatibility with Libtask quite difficult @yebai
There was a problem hiding this comment.
@THargreaves can you revert this change for now? It will take a bit more work on the Turing side before we can do it.
Draft work on updating AdvancedPS to match the SSMProblems interface.
To run this code you need to
devSSMProblems in the AdvancedPS environment anddevAdvancedPS in the examples/particle-gibbs environment.