|
| 1 | +module DynamicPPLFastLDFTests |
| 2 | + |
| 3 | +using DynamicPPL |
| 4 | +using Distributions |
| 5 | +using DistributionsAD: filldist |
| 6 | +using ADTypes |
| 7 | +using DynamicPPL.Experimental: FastLDF |
| 8 | + |
| 9 | +@testset "Automatic differentiation" begin |
| 10 | + # Used as the ground truth that others are compared against. |
| 11 | + ref_adtype = AutoForwardDiff() |
| 12 | + |
| 13 | + test_adtypes = if MOONCAKE_SUPPORTED |
| 14 | + [ |
| 15 | + AutoReverseDiff(; compile=false), |
| 16 | + AutoReverseDiff(; compile=true), |
| 17 | + AutoMooncake(; config=nothing), |
| 18 | + ] |
| 19 | + else |
| 20 | + [AutoReverseDiff(; compile=false), AutoReverseDiff(; compile=true)] |
| 21 | + end |
| 22 | + |
| 23 | + @testset "Unsupported backends" begin |
| 24 | + @model demo() = x ~ Normal() |
| 25 | + @test_logs (:warn, r"not officially supported") FastLDF(demo(); adtype=AutoZygote()) |
| 26 | + end |
| 27 | + |
| 28 | + @testset "Correctness" begin |
| 29 | + @testset "$(m.f)" for m in DynamicPPL.TestUtils.DEMO_MODELS |
| 30 | + varinfo = VarInfo(m) |
| 31 | + linked_varinfo = DynamicPPL.link(varinfo, m) |
| 32 | + f = FastLDF(m, getlogjoint_internal, linked_varinfo) |
| 33 | + x = linked_varinfo[:] |
| 34 | + |
| 35 | + # Calculate reference logp + gradient of logp using ForwardDiff |
| 36 | + ref_ad_result = run_ad(m, ref_adtype; varinfo=linked_varinfo, test=NoTest()) |
| 37 | + ref_logp, ref_grad = ref_ad_result.value_actual, ref_ad_result.grad_actual |
| 38 | + |
| 39 | + @testset "$adtype" for adtype in test_adtypes |
| 40 | + @info "Testing AD on: $(m.f) - $(short_varinfo_name(linked_varinfo)) - $adtype" |
| 41 | + |
| 42 | + @test run_ad( |
| 43 | + m, |
| 44 | + adtype; |
| 45 | + varinfo=linked_varinfo, |
| 46 | + test=WithExpectedResult(ref_logp, ref_grad), |
| 47 | + ) isa Any |
| 48 | + end |
| 49 | + end |
| 50 | + end |
| 51 | + |
| 52 | + # Test that various different ways of specifying array types as arguments work with all |
| 53 | + # ADTypes. |
| 54 | + @testset "Array argument types" begin |
| 55 | + test_m = randn(2, 3) |
| 56 | + |
| 57 | + function eval_logp_and_grad(model, m, adtype) |
| 58 | + ldf = FastLDF(model(); adtype=adtype) |
| 59 | + return LogDensityProblems.logdensity_and_gradient(ldf, m[:]) |
| 60 | + end |
| 61 | + |
| 62 | + @model function scalar_matrix_model(::Type{T}=Float64) where {T<:Real} |
| 63 | + m = Matrix{T}(undef, 2, 3) |
| 64 | + return m ~ filldist(MvNormal(zeros(2), I), 3) |
| 65 | + end |
| 66 | + |
| 67 | + scalar_matrix_model_reference = eval_logp_and_grad( |
| 68 | + scalar_matrix_model, test_m, ref_adtype |
| 69 | + ) |
| 70 | + |
| 71 | + @model function matrix_model(::Type{T}=Matrix{Float64}) where {T} |
| 72 | + m = T(undef, 2, 3) |
| 73 | + return m ~ filldist(MvNormal(zeros(2), I), 3) |
| 74 | + end |
| 75 | + |
| 76 | + matrix_model_reference = eval_logp_and_grad(matrix_model, test_m, ref_adtype) |
| 77 | + |
| 78 | + @model function scalar_array_model(::Type{T}=Float64) where {T<:Real} |
| 79 | + m = Array{T}(undef, 2, 3) |
| 80 | + return m ~ filldist(MvNormal(zeros(2), I), 3) |
| 81 | + end |
| 82 | + |
| 83 | + scalar_array_model_reference = eval_logp_and_grad( |
| 84 | + scalar_array_model, test_m, ref_adtype |
| 85 | + ) |
| 86 | + |
| 87 | + @model function array_model(::Type{T}=Array{Float64}) where {T} |
| 88 | + m = T(undef, 2, 3) |
| 89 | + return m ~ filldist(MvNormal(zeros(2), I), 3) |
| 90 | + end |
| 91 | + |
| 92 | + array_model_reference = eval_logp_and_grad(array_model, test_m, ref_adtype) |
| 93 | + |
| 94 | + @testset "$adtype" for adtype in test_adtypes |
| 95 | + scalar_matrix_model_logp_and_grad = eval_logp_and_grad( |
| 96 | + scalar_matrix_model, test_m, adtype |
| 97 | + ) |
| 98 | + @test scalar_matrix_model_logp_and_grad[1] ≈ scalar_matrix_model_reference[1] |
| 99 | + @test scalar_matrix_model_logp_and_grad[2] ≈ scalar_matrix_model_reference[2] |
| 100 | + matrix_model_logp_and_grad = eval_logp_and_grad(matrix_model, test_m, adtype) |
| 101 | + @test matrix_model_logp_and_grad[1] ≈ matrix_model_reference[1] |
| 102 | + @test matrix_model_logp_and_grad[2] ≈ matrix_model_reference[2] |
| 103 | + scalar_array_model_logp_and_grad = eval_logp_and_grad( |
| 104 | + scalar_array_model, test_m, adtype |
| 105 | + ) |
| 106 | + @test scalar_array_model_logp_and_grad[1] ≈ scalar_array_model_reference[1] |
| 107 | + @test scalar_array_model_logp_and_grad[2] ≈ scalar_array_model_reference[2] |
| 108 | + array_model_logp_and_grad = eval_logp_and_grad(array_model, test_m, adtype) |
| 109 | + @test array_model_logp_and_grad[1] ≈ array_model_reference[1] |
| 110 | + @test array_model_logp_and_grad[2] ≈ array_model_reference[2] |
| 111 | + end |
| 112 | + end |
| 113 | +end |
| 114 | + |
| 115 | +end |
0 commit comments