Merge 593d305 into bfcf29c

docs/src/index.md

@@ -1,6 +1,92 @@
 # SSMProblems
 
-### API
+### Installation
+In the `julia` REPL:
+```julia
+]add SSMProblems
+```
+
+### Documentation
+
+`SSMProblems` defines a generic interface for State Space Problems (SSM). The main objective is to provide a consistent
+interface to work with SSMs and their logdensities.
+
+Consider a markovian model from[^Murray]:
+![state space model](./docs/images/state_space_model.png)
+
+[^Murray]:
+ > Murray, Lawrence & Lee, Anthony & Jacob, Pierre. (2013). Rethinking resampling in the particle filter on graphics processing units. 
+
+The model is fully specified by the following densities:
+- __Initialisation__: ``f_0(x)``
+- __Transition__: ``f(x)``
+- __Emission__: ``g(x)``
+
+And the dynamics of the model reduces to:
+```math
+x_t | x_{t-1} \sim f(x_t | x_{t-1})
+```
+```math
+y_t | x_t \sim g(y_t | x_{t})
+```
+assuming ``x_0 \sim f_0(x)``. 
+
+The joint law follows:
+
+```math
+p(x_{0:T}, y_{0:T}) = f_0(x_0) \prod_t g(y_t | x_t) f(x_t | x_{t-1})
+```
+
+Model users can define their `SSM` using the following interface:
+```julia
+
+struct Model <: AbstractStateSpaceModel end
+
+# Define the structure of the latent space
+particleof(::Model) = Float64
+dimension(::Model) = 2
+
+function transition!!(
+ rng::Random.AbstractRNG, 
+ step, 
+ model::Model, 
+ particle::AbstractParticl{<:AbstractStateSpaceModel}
+) 
+ if step == 1
+ ... # Sample from the initial density
+ end
+ ... # Sample from the transition density
+end
+
+function emission_logdensity(step, model::Model, particle::AbstractParticle) 
+ ... # Return log density of the model at *time* `step`
+end
+
+isdone(step, model::Model, particle::AbstractParticle) = ... # Stops the state machine
+
+# Optionally, if the transition density is known, the model can also specify it
+function transition_logdensity(step, prev_particle::AbstractParticle, next_particle::AbstractParticle)
+ ... # Scores the forward transition at `x`
+end
+```
+
+Package users can then consume the model `logdensity` through calls to `emission_logdensity`. 
+
+For example, a bootstrap filter targeting the filtering distribution ``p(x_t | y_{0:t})`` using `N` particles would roughly follow:
+```julia
+struct Particle{T<:AbstractStateSpaceModel} <: AbstractParticle{T} end 
+
+while !all(map(particle -> isdone(t, model, particles), particles)):
+ ancestors = resample(rng, logweigths)
+ particles = particles[ancestors]
+ for i in 1:N
+ particles[i] = transition!!(rng, t, model, particles[i])
+ logweights[i] += emission_logdensity(t, model, particles[i])
+ end
+end
+```
+
+### Interface
 ```@autodocs
 Modules = [SSMProblems]
 Order = [:type, :function]

examples/smc.jl

@@ -4,15 +4,23 @@ using Distributions
 using Plots
 using StatsFuns
 
-# Particle Filter implementation
-struct Particle{T} # Here we just need a tree
- parent::Union{Particle,Nothing}
+# Particle Filter 
+abstract type Node{T} end
+
+struct Root{T} <: Node{T} end
+Root(T) = Root{T}()
+Root() = Root(Any)
+
+struct Particle{T} <: Node{T}
+ parent::Node{T}
  state::T
 end
 
-Particle(state::T) where {T} = Particle(nothing, state)
-Particle() = Particle(nothing, nothing)
-Base.show(io::IO, p::Particle) = print(io, "Particle($(p.state))")
+Particle(state::T) where {T} = Particle(Root(T), state)
+
+Base.show(io::IO, p::Particle{T}) where {T} = print(io, "Particle{$T}($(p.state))")
+
+const ParticleContainer{T} = AbstractVector{<:Particle{T}}
 
 """
  linearize(particle)
@@ -21,16 +29,14 @@ Return the trace of a particle, i.e. the sequence of states from the root to the
 """
 function linearize(particle::Particle{T}) where {T}
  trace = T[]
- parent = particle.parent
- while !isnothing(parent)
- push!(trace, parent.state)
- parent = parent.parent
+ current = particle
+ while !isa(current, Root)
+ push!(trace, current.state)
+ current = current.parent
  end
  return trace
 end
 
-ParticleContainer = AbstractVector{<:Particle}
-
 ess(weights) = inv(sum(abs2, weights))
 get_weights(logweights::T) where {T<:AbstractVector{<:Real}} = StatsFuns.softmax(logweights)
 
@@ -40,79 +46,105 @@ function systematic_resampling(
  return rand(rng, Distributions.Categorical(weights), n)
 end
 
-function sweep!(rng::AbstractRNG, particles::ParticleContainer, resampling, threshold=0.5)
- t = 1
+function sweep!(
+ rng::AbstractRNG,
+ model::AbstractStateSpaceModel,
+ particles::ParticleContainer,
+ observations::AbstractArray,
+ resampling=systematic_resampling,
+ threshold=0.5,
+)
  N = length(particles)
- logweights = zeros(length(particles))
- while !isdone(t, particles[1].state)
+ logweights = zeros(N)
 
+ for (timestep, observation) in enumerate(observations)
  # Resample step
  weights = get_weights(logweights)
  if ess(weights) <= threshold * N
  idx = resampling(rng, weights)
  particles = particles[idx]
- logweights = zeros(length(particles))
+ fill!(logweights, 0)
  end
 
  # Mutation step
  for i in eachindex(particles)
- parent = particles[i]
- mutated = transition!!(rng, t, parent.state)
- particles[i] = Particle(parent, mutated)
- logweights[i] += emission_logdensity(t, particles[i].state)
+ latent_state = transition!!(rng, model, timestep, particles[i].state)
+ particles[i] = Particle(particles[i], latent_state)
+ logweights[i] += emission_logdensity(
+ model, timestep, particles[i].state, observation
+ )
  end
-
- t += 1
  end
 
  # Return unweighted set
  idx = resampling(rng, get_weights(logweights))
  return particles[idx]
 end
 
-function sweep!(rng::AbstractRNG, n::Int, resampling, threshold=0.5)
- particles = [Particle(0.0) for _ in 1:n]
- return sweep!(rng, particles, resampling, threshold)
+# Turing style sample method
+function sample(
+ rng::AbstractRNG,
+ model::AbstractStateSpaceModel,
+ n::Int,
+ observations::AbstractVector;
+ resampling=systematic_resampling,
+ threshold=0.5,
+)
+ particles = map(1:N) do i
+ state = transition!!(rng, model)
+ Particle(state)
+ end
+ samples = sweep!(rng, model, particles, observations, resampling, threshold)
+ return samples
 end
 
 # Inference code
-Base.@kwdef struct Parameters
+Base.@kwdef struct LinearSSM <: AbstractStateSpaceModel
  v::Float64 = 0.2 # Transition noise stdev
  u::Float64 = 0.7 # Observation noise stdev
 end
 
 # Simulation
 T = 250
 seed = 1
-N = 1000
+N = 1_000
 rng = MersenneTwister(seed)
-params = Parameters(; v=0.2, u=0.7)
-
-function transition!!(rng::AbstractRNG, t::Int, state=nothing)
- if isnothing(state)
- return rand(rng, Normal(0, 1))
- end
- return rand(rng, Normal(state, params.v))
-end
 
-function emission_logdensity(t, state)
- return logpdf(Normal(state, params.u), observations[t])
-end
+model = LinearSSM(0.2, 0.7)
 
-isdone(t, state) = t > T
-isdone(t, ::Nothing) = false
+f0() = Normal(0, 1)
+f(t::Int, x::Float64) = Normal(x, model.v)
+g(t::Int, y::Float64) = Normal(y, model.u)
 
+# Generate synthtetic data
 x, observations = zeros(T), zeros(T)
-x[1] = rand(rng, Normal(0, 1))
+x[1] = rand(rng, f0())
 for t in 1:T
- observations[t] = rand(rng, Normal(x[t], params.u))
+ observations[t] = rand(rng, g(t, x[t]))
  if t < T
- x[t + 1] = rand(rng, Normal(x[t], params.v))
+ x[t + 1] = rand(rng, f(t, x[t]))
  end
 end
 
-samples = sweep!(rng, fill(Particle(x[1]), N), systematic_resampling)
+# Model dynamics
+function transition!!(rng::AbstractRNG, model::LinearSSM)
+ return rand(rng, f0())
+end
+
+function transition!!(rng::AbstractRNG, model::LinearSSM, timestep::Int, state::Float64)
+ return rand(rng, f(timestep, state))
+end
+
+function emission_logdensity(
+ model::LinearSSM, timestep::Int, state::Float64, observation::Float64
+)
+ return logpdf(g(timestep, state), observation)
+end
+
+# Sample latent state trajectories
+samples = sample(rng, LinearSSM(), N, observations)
 traces = reverse(hcat(map(linearize, samples)...))
 
 scatter(traces; color=:black, opacity=0.3, label=false)
 plot!(x; label="True state")
+plot!(mean(traces; dims=2); label="Posterior mean")

src/SSMProblems.jl

@@ -4,38 +4,32 @@ A unified interface to define State Space Models interfaces in the context of Pa
 module SSMProblems
 
 """
+ AbstractStateSpaceModel
 """
-abstract type AbstractParticle end
+abstract type AbstractStateSpaceModel end
 abstract type AbstractParticleCache end
 
 """
- transition!!(rng, step, particle[, cache])
+ transition!!(rng, model[, timestep, state, cache])
 
 Simulate the particle for the next time step from the forward dynamics.
 """
 function transition!! end
 
 """
- transition_logdensity(step, particle, x[, cache])
+ transition_logdensity(model, timestep, prev_state, next_state[, cache])
 
 (Optional) Computes the log-density of the forward transition if the density is available.
 """
 function transition_logdensity end
 
 """
- emission_logdensity(step, particle[, cache])
+ emission_logdensity(model, timestep, state, observation[, cache])
 
-Compute the log potential of current particle. This effectively "reweight" each particle.
+Compute the log potential of the current particle. This effectively "reweight" each particle.
 """
 function emission_logdensity end
 
-"""
- isdone(step, particle[, cache])
-
-Determine whether we have reached the last time step of the Markov process. Return `true` if yes, otherwise return `false`.
-"""
-function isdone end
-
-export transition!!, transition_logdensity, emission_logdensity, isdone, AbstractParticle
+export AbstractStateSpaceModel
 
 end