Modify inference for predictions functionality

src/inference.jl

@@ -627,6 +627,7 @@ function inference(;
             end
         end
     else # In this case, the prediction functionality should only be performed if the data allows missings and actually contains missing values.
+        foreach((variable, value) -> isdata(value) && __inference_check_dataismissing(data[variable]) && !allows_missings(value) ? error("datavar $(variable) has missings inside but does not allow it. Add `where {allow_missing = true }`") : nothing, keys(vardict), values(vardict))
         predictvars = Dict(
             variable => KeepLast() for (variable, value) in pairs(vardict) if
             (isdata(value) && haskey(data, variable) && allows_missings(value) && __inference_check_dataismissing(data[variable]) && !isanonymous(value))
@@ -684,6 +685,10 @@ function inference(;
 
         is_free_energy, S, T = unwrap_free_energy_option(free_energy)
 
+        if !isempty(actors_pr) && is_free_energy
+            error("Cannot compute Bethe Free Energy for models with prediction variables. Please set `free_energy = false`.")
+        end
+
         if is_free_energy
             fe_actor        = ScoreActor(S, _iterations, 1)
             fe_objective    = BetheFreeEnergy(BetheFreeEnergyDefaultMarginalSkipStrategy, AsapScheduler(), free_energy_diagnostics)

test/test_inference.jl

@@ -722,11 +722,6 @@ end
 
 @testset "Predictions functionality" begin
 
-    # Given the current implementation of RxInfer ecosystem, this rule needs to be predifined 
-    @rule NormalMeanPrecision(:μ, Marginalisation) (m_out::PointMass, q_τ::PointMass) = begin
-        return missing
-    end
-
     # test #1 (array with missing + predictvars)
     data = (y = [1.0, -500.0, missing, 100.0],)
 
@@ -857,7 +852,7 @@ end
     @test haskey(result.predictions, :y)
     @test typeof(result.predictions[:y]) <: NormalDistributionsFamily
 
-    # test vmp model
+    # test #7 vmp model
     data = (y = [1.0, -10.0, 5.0],)
     @model function vmp_model(n)
         x = randomvar(n + 1)
@@ -894,6 +889,52 @@ end
 
     @test first(result.posteriors[:γ]) != last(result.posteriors[:γ])
     @test first(result.predictions[:o]) != last(result.predictions[:o])
+
+
+    # test #8 non gaussian likelihood (single datavar missing)
+    dataset = [1.0, 0.0, 1.0, missing,]
+    @model function coin_model(n)
+
+        y = datavar(Float64, n) where {allow_missing = true}
+
+        θ ~ Beta(4.0, 8.0)
+        for i in 1:n
+            y[i] ~ Bernoulli(θ)
+        end
+
+    end
+
+    result = inference(
+        model = coin_model(length(dataset)), 
+        data  = (y = dataset, )
+    )
+
+    @test typeof(last(result.predictions[:y])) <: Bernoulli
+
+    # test #9 allow_missing error handling
+    dataset = [1.0, 0.0, 1.0, missing,]
+    @model function coin_model(n)
+
+        y = datavar(Float64, n)
+
+        θ ~ Beta(4.0, 8.0)
+        for i in 1:n
+            y[i] ~ Bernoulli(θ)
+        end
+
+    end
+
+    @test_throws ErrorException inference(
+        model = coin_model(length(dataset)), 
+        data  = (y = dataset, )
+    )
+
+    #test #10 free_energy error handling
+    @test_throws ErrorException inference(
+        model = coin_model(length(dataset)), 
+        data  = (y = dataset, ),
+        free_energy = true
+    )
 end
 
 end