|
1 | 1 | using AdvancedHMC, AbstractMCMC, Random |
2 | 2 | include("common.jl") |
3 | 3 |
|
4 | | -# Initalize samplers |
5 | | -nuts = NUTS(0.8) |
6 | | -nuts_32 = NUTS(0.8f0) |
7 | | -hmc = HMC(0.1, 25) |
8 | | -hmcda = HMCDA(0.8, 1.0) |
9 | | -hmcda_32 = HMCDA(0.8f0, 1.0) |
10 | | - |
11 | | -integrator = Leapfrog(1e-3) |
12 | | -kernel = HMCKernel(Trajectory{MultinomialTS}(integrator, GeneralisedNoUTurn())) |
13 | | -metric = DiagEuclideanMetric(2) |
14 | | -adaptor = AdvancedHMC.make_adaptor(nuts, metric, integrator) |
15 | | -custom = HMCSampler(kernel, metric, adaptor) |
16 | | - |
17 | | -# Check that everything is initalized correctly |
18 | 4 | @testset "Constructors" begin |
19 | | - # Types |
20 | | - @test typeof(nuts) == NUTS{Float64} |
21 | | - @test typeof(nuts_32) == NUTS{Float32} |
22 | | - @test typeof(hmc) == HMC{Float64} |
23 | | - @test typeof(hmcda) == HMCDA{Float64} |
24 | | - @test typeof(nuts) <: AdvancedHMC.AbstractHMCSampler |
25 | | - @test typeof(nuts) <: AbstractMCMC.AbstractSampler |
26 | | - |
27 | | - # NUTS |
28 | | - @test nuts.δ == 0.8 |
29 | | - @test nuts.max_depth == 10 |
30 | | - @test nuts.Δ_max == 1000.0 |
31 | | - @test nuts.init_ϵ == 0.0 |
32 | | - @test nuts.integrator == :leapfrog |
33 | | - @test nuts.metric == :diagonal |
34 | | - |
35 | | - # NUTS Float32 |
36 | | - @test nuts_32.δ == 0.8f0 |
37 | | - @test nuts_32.max_depth == 10 |
38 | | - @test nuts_32.Δ_max == 1000.0f0 |
39 | | - @test nuts_32.init_ϵ == 0.0f0 |
40 | | - |
41 | | - # HMC |
42 | | - @test hmc.n_leapfrog == 25 |
43 | | - @test hmc.init_ϵ == 0.1 |
44 | | - @test hmc.integrator == :leapfrog |
45 | | - @test hmc.metric == :diagonal |
46 | | - |
47 | | - # HMCDA |
48 | | - @test hmcda.δ == 0.8 |
49 | | - @test hmcda.λ == 1.0 |
50 | | - @test hmcda.init_ϵ == 0.0 |
51 | | - @test hmcda.integrator == :leapfrog |
52 | | - @test hmcda.metric == :diagonal |
53 | | - |
54 | | - # HMCDA Float32 |
55 | | - @test hmcda_32.δ == 0.8f0 |
56 | | - @test hmcda_32.λ == 1.0f0 |
57 | | - @test hmcda_32.init_ϵ == 0.0f0 |
58 | | -end |
59 | | - |
60 | | -@testset "First step" begin |
61 | | - rng = MersenneTwister(0) |
62 | | - θ_init = randn(rng, 2) |
63 | | - logdensitymodel = AbstractMCMC.LogDensityModel(ℓπ_gdemo) |
64 | | - _, nuts_state = |
65 | | - AbstractMCMC.step(rng, logdensitymodel, nuts; n_adapts = 0, init_params = θ_init) |
66 | | - _, hmc_state = |
67 | | - AbstractMCMC.step(rng, logdensitymodel, hmc; n_adapts = 0, init_params = θ_init) |
68 | | - _, nuts_32_state = |
69 | | - AbstractMCMC.step(rng, logdensitymodel, nuts_32; n_adapts = 0, init_params = θ_init) |
70 | | - _, custom_state = |
71 | | - AbstractMCMC.step(rng, logdensitymodel, custom; n_adapts = 0, init_params = θ_init) |
| 5 | + θ_init = randn(2) |
| 6 | + model = AbstractMCMC.LogDensityModel(ℓπ_gdemo) |
72 | 7 |
|
73 | | - # Metric |
74 | | - @test typeof(nuts_state.metric) == DiagEuclideanMetric{Float64,Vector{Float64}} |
75 | | - @test typeof(nuts_32_state.metric) == DiagEuclideanMetric{Float32,Vector{Float32}} |
76 | | - @test custom_state.metric == metric |
| 8 | + @testset "$T" for T in [Float32, Float64] |
| 9 | + @testset "$(nameof(typeof(sampler)))" for (sampler, expected) in [ |
| 10 | + ( |
| 11 | + HMC(T(0.1), 25), |
| 12 | + ( |
| 13 | + adaptor_type = NoAdaptation, |
| 14 | + metric_type = DiagEuclideanMetric{T}, |
| 15 | + integrator_type = Leapfrog{T}, |
| 16 | + ), |
| 17 | + ), |
| 18 | + # This should peform the correct promotion for the 2nd argument. |
| 19 | + ( |
| 20 | + HMCDA(T(0.1), 1), |
| 21 | + ( |
| 22 | + adaptor_type = StanHMCAdaptor, |
| 23 | + metric_type = DiagEuclideanMetric{T}, |
| 24 | + integrator_type = Leapfrog{T}, |
| 25 | + ), |
| 26 | + ), |
| 27 | + ( |
| 28 | + NUTS(T(0.8)), |
| 29 | + ( |
| 30 | + adaptor_type = StanHMCAdaptor, |
| 31 | + metric_type = DiagEuclideanMetric{T}, |
| 32 | + integrator_type = Leapfrog{T}, |
| 33 | + ), |
| 34 | + ), |
| 35 | + ( |
| 36 | + NUTS(T(0.8); metric = :unit), |
| 37 | + ( |
| 38 | + adaptor_type = StanHMCAdaptor, |
| 39 | + metric_type = UnitEuclideanMetric{T}, |
| 40 | + integrator_type = Leapfrog{T}, |
| 41 | + ), |
| 42 | + ), |
| 43 | + ( |
| 44 | + NUTS(T(0.8); metric = :dense), |
| 45 | + ( |
| 46 | + adaptor_type = StanHMCAdaptor, |
| 47 | + metric_type = DenseEuclideanMetric{T}, |
| 48 | + integrator_type = Leapfrog{T}, |
| 49 | + ), |
| 50 | + ), |
| 51 | + ] |
| 52 | + # Make sure the sampler element type is preserved. |
| 53 | + @test AdvancedHMC.sampler_eltype(sampler) == T |
77 | 54 |
|
78 | | - # Integrator |
79 | | - @test typeof(nuts_state.κ.τ.integrator) == Leapfrog{Float64} |
80 | | - @test typeof(nuts_32_state.κ.τ.integrator) == Leapfrog{Float32} |
81 | | - @test custom_state.κ.τ.integrator == integrator |
| 55 | + # Step. |
| 56 | + rng = Random.default_rng() |
| 57 | + transition, state = |
| 58 | + AbstractMCMC.step(rng, model, sampler; n_adapts = 0, init_params = θ_init) |
82 | 59 |
|
83 | | - # Kernel |
84 | | - @test nuts_state.κ == AdvancedHMC.make_kernel(nuts, nuts_state.κ.τ.integrator) |
85 | | - @test custom_state.κ == kernel |
| 60 | + # Verify that the types are preserved in the transition. |
| 61 | + @test eltype(transition.z.θ) == T |
| 62 | + @test eltype(transition.z.r) == T |
| 63 | + @test eltype(transition.z.ℓπ.value) == T |
| 64 | + @test eltype(transition.z.ℓπ.gradient) == T |
| 65 | + @test eltype(transition.z.ℓκ.value) == T |
| 66 | + @test eltype(transition.z.ℓκ.gradient) == T |
86 | 67 |
|
87 | | - # Adaptor |
88 | | - @test typeof(nuts_state.adaptor) <: StanHMCAdaptor |
89 | | - @test hmc_state.adaptor == NoAdaptation() |
90 | | - @test custom_state.adaptor == adaptor |
| 68 | + # Verify that the state is what we expect. |
| 69 | + @test AdvancedHMC.getmetric(state) isa expected.metric_type |
| 70 | + @test AdvancedHMC.getintegrator(state) isa expected.integrator_type |
| 71 | + @test AdvancedHMC.getadaptor(state) isa expected.adaptor_type |
| 72 | + end |
| 73 | + end |
91 | 74 | end |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4&size=40){kind=avatar}
0 commit comments