Fix failing EnIS test

src/ensembles/importance_sampling.jl

@@ -123,15 +123,18 @@ function importance_weights(obs::AbstractVector, pred::AbstractMatrix, R::Diagon
     n_obs, ensemble_size = size(pred)
     @assert n_obs == length(obs)
     residual = obs .- pred
-    loglik = dropdims(-0.5*(1/diag(R))*residual.^2, dims=1)
+    loglik = sum(-0.5*inv(R)*residual.^2, dims=1)[1,:]
 
     # Log of normalizing constant
     log_z = logsumexp(loglik)
 
     # Weights
     logw = loglik .- log_z
     weights = exp.(logw)
-    @assert length(weights) == ensemble_size && round(sum(weights), digits=10) == 1.0 "particle weights do not sum to unity!"
+
+    # sanity checks
+    @assert length(weights) == ensemble_size "weight vector has incorrect dimensions: $(size(weights))"
+    @assert abs(sum(weights) - 1.0) < sqrt(eps()) "particle weights do not sum to unity! ∑w=$(sum(weights))"
 
     Neff = round(1/sum(weights.^2))
 

src/forward_problem.jl

@@ -11,7 +11,7 @@ struct SimulatorForwardProblem{probType,obsType,configType,names} <: SciMLBase.A
     config::configType
 end
 
-const SimulatorSciMLForwardProblem = SimulatorForwardProblem{<:SciMLBase.AbstractSciMLProblem}
+const SimulatorSciMLForwardProblem{probType} = SimulatorForwardProblem{probType} where {probType<:SciMLBase.AbstractSciMLProblem}
 
 """
     SimulatorForwardProblem(prob::SciMLBase.AbstractSciMLProblem, observables::SimulatorObservable...)
@@ -42,25 +42,32 @@ parameters `p0` and a default transient observable.
 SimulatorForwardProblem(f, p0::AbstractVector) = SimulatorForwardProblem(f, p0, SimulatorObservable(:y, state -> state.u))
 
 """
-    SciMLBase.remaker_of(forward_prob::SimulatorForwardProblem)
-
-Returns a function which will rebuild a `SimulatorForwardProblem` from its arguments.
-The remaker function additionally provides a keyword argument `copy_observables` which,
-if `true`, will `deepcopy` the observables to ensure independence. The default setting is `true`.
-"""
-function SciMLBase.remaker_of(forward_prob::SimulatorForwardProblem)
-    function remake_forward_prob(;
+    SciMLBase.remake(
+        forward_prob::SimulatorForwardProblem;
         prob=forward_prob.prob,
         observables=forward_prob.observables,
         config=forward_prob.config,
         copy_observables=true,
         kwargs...
     )
-        new_observables = copy_observables ? deepcopy(observables) : observables
-        return SimulatorForwardProblem(remake(prob; kwargs...), new_observables, config)
-    end
+
+Rebuilds a `SimulatorForwardProblem` from its individual components. If `copy_observables=true`,
+then `remake` will `deepcopy` the observables to ensure independence. The default setting is `true`.
+"""
+function SciMLBase.remake(
+    forward_prob::SimulatorForwardProblem;
+    prob=forward_prob.prob,
+    observables=forward_prob.observables,
+    config=forward_prob.config,
+    copy_observables=true,
+    kwargs...
+)
+    new_observables = copy_observables ? deepcopy(observables) : observables
+    return SimulatorForwardProblem(remake(prob; kwargs...), new_observables, config)
 end
 
+SciMLBase.remaker_of(forward_prob::SimulatorForwardProblem) = (;kwargs...) -> remake(forward_prob; kwargs...)
+
 # DiffEqBase dispatches to make solve/init interface work correctly
 DiffEqBase.check_prob_alg_pairing(prob::SimulatorSciMLForwardProblem, alg) = DiffEqBase.check_prob_alg_pairing(prob.prob, alg)
 DiffEqBase.isinplace(prob::SimulatorSciMLForwardProblem) = DiffEqBase.isinplace(prob.prob)

test/ensembles/enis_tests.jl

@@ -59,13 +59,13 @@ end
 
 @testset "EnIS: evensen_scalar_nonlinear" begin
     x_true = 1.0
-    b_true = 0.2
+    b_true = 0.1
     σ_y = 0.1
     ensemble_size = 1024
     x_prior = Normal(0,1)
     # log-normal with mean 0.1 and stddev 0.2
     # b_prior = autoprior(0.1, 0.2, lower=0.0, upper=Inf)
-    b_prior = Bijectors.TransformedDistribution(Normal(log(0.2), 1), Base.Fix1(broadcast, exp))
+    b_prior = Bijectors.TransformedDistribution(Normal(0.0, 1.0), Base.Fix1(broadcast, exp))
     rng = Random.MersenneTwister(1234)
     testprob = evensen_scalar_nonlinear(x_true, b_true; n_obs=100, rng, x_prior, b_prior)
     transform = bijector(testprob.prior.model)