Add Custom Kernel

Project.toml

@@ -1,7 +1,7 @@
 name = "Baytes"
 uuid = "72ddfcfc-6e9d-43df-829b-7aed7c549d4f"
 authors = ["Patrick Aschermayr <[email protected]>"]
-version = "0.3.14"
+version = "0.3.15"
 
 [deps]
 ArgCheck = "dce04be8-c92d-5529-be00-80e4d2c0e197"
@@ -36,17 +36,18 @@ MCMCDiagnosticTools = "0.3"
 ModelWrappers = "0.5"
 PrettyTables = "2"
 ProgressMeter = "1.7"
-SimpleUnPack = "1"
 Random = "1.9"
+SimpleUnPack = "1"
 Statistics = "1.9"
 julia = "^1.9"
 
 [extras]
+BaytesDiff = "12a76ff9-393d-487f-8b39-e615b97e2f77"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
+ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 NLSolversBase = "d41bc354-129a-5804-8e4c-c37616107c6c"
 Optim = "429524aa-4258-5aef-a3af-852621145aeb"
-ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
-test = ["Test", "Distributions", "NLSolversBase", "Optim", "ForwardDiff"]
+test = ["Test", "BaytesDiff", "Distributions", "NLSolversBase", "Optim", "ForwardDiff"]

src/Baytes.jl

@@ -3,7 +3,21 @@ module Baytes
 
 ################################################################################
 #Import modules
-import BaytesCore: BaytesCore, update!, infer, results, init, init!, propose, propose!, propagate!
+import BaytesCore: 
+    BaytesCore, 
+    update!, 
+    infer, 
+    results, 
+    init, 
+    init!, 
+    propose, 
+    propose!,
+    propagate!,
+    generate, 
+    generate_showvalues,
+    get_result, 
+    result! 
+
 using BaytesCore:
     BaytesCore,
     AbstractAlgorithm,
@@ -25,7 +39,6 @@ using BaytesCore:
     Batch,
     SubSampled,
     adjust,
-    generate_showvalues,
     TemperingMethod,
     IterationTempering,
     JointTempering,
@@ -41,7 +54,8 @@ using BaytesCore:
 import ModelWrappers:
     ModelWrappers,
     sample,
-    sample!
+    sample!,
+    predict
 
 using ModelWrappers:
     ModelWrappers,
@@ -66,6 +80,12 @@ using ModelWrappers:
     UnflattenTypes,
     UnflattenStrict,
     UnflattenFlexible
+#=
+using BaytesDiff: 
+    BaytesDiff, 
+    ℓObjectiveResult,
+    ℓDensityResult
+=#
 
 using BaytesMCMC, BaytesFilters, BaytesPMCMC, BaytesSMC, BaytesOptim
 
@@ -202,7 +222,11 @@ export
     Optimizer, 
     OptimConstructor,
     OptimDefault, 
-
+    OptimLBFG,
+    CustomAlgorithmDefault,
+    CustomAlgorithm,
+    CustomAlgorithmConstructor,
+
     ## BaytesSMC
     SMC,
     SMCDefault,

test/runtests.jl

@@ -24,4 +24,5 @@ include("testhelper/TestHelper.jl")
 # Run Tests
 @testset "All tests" begin
     include("test-construction.jl")
+    include("test-custom.jl")
 end

test/test-construction.jl

@@ -373,6 +373,11 @@ using Optim, NLSolversBase
     end
 end
 
+############################################################################################
+# Check if Custom Sampler works
+
+
+
 ############################################################################################
 #Utility
 @testset "Utility, maxiterations" begin

test/test-custom.jl

@@ -0,0 +1,90 @@
+############################################################################################
+# Models to be used in construction
+objectives = [
+    Objective(ModelWrapper(MyBaseModel(), myparameter1, (;), FlattenDefault()), data_uv),
+    Objective(ModelWrapper(MyBaseModel(), myparameter1, (;), FlattenDefault(; output = Float32)), data_uv)
+]
+Nchains = 4
+tempermethods = [
+    IterationTempering(Float64, UpdateFalse(), 1.0, 1000),
+    IterationTempering(Float64, UpdateTrue(), 1.0, 1000),
+    JointTempering(Float64, UpdateFalse(), .5, Float64(Nchains), Nchains),
+    JointTempering(Float64, UpdateTrue(), .5, Float64(Nchains), Nchains)
+]
+
+#=
+iter = 2
+tempermethod = tempermethods[iter]
+=#
+
+## Add custom Step for propagate
+using BaytesDiff
+import BaytesOptim: BaytesOptim, propagate
+## Extend Custom Method
+function propagate(
+    _rng::Random.AbstractRNG, algorithm::CustomAlgorithm, objective::Objective{<:ModelWrapper{MyBaseModel}})
+    logobjective = BaytesDiff.ℓDensityResult(objective)
+    #logobjective.θᵤ[1] = 5
+    logobjective.θᵤ[1] = rand()
+    return logobjective
+end
+
+############################################################################################
+@testset "Sampling, type conversion" begin
+    for tempermethod in tempermethods
+        for iter in eachindex(objectives)
+            #println(tempermethod, " ", iter)
+                sampledefault = SampleDefault(;
+                    dataformat=Batch(),
+                    tempering=deepcopy(tempermethod), #IterationTempering(Float64, UpdateFalse(), 1.0, 1000),
+                    chains=4,
+                    iterations=100,
+                    burnin=max(1, Int64(floor(10/10))),
+                    thinning = 1,
+                    safeoutput=false,
+                    printoutput=false,
+                    printdefault=PrintDefault(),
+                    report=ProgressReport(;
+                        bar=false,
+                        log=SilentLog()
+                    ),
+                )
+                temperupdate = sampledefault.tempering.adaption
+                _obj = deepcopy(objectives[iter])
+                _flattentype = _obj.model.info.reconstruct.default.output
+
+                # Create Custom Algorithm
+                def = CustomAlgorithmDefault(; 
+                    generated=UpdateTrue()
+                )
+                opt = CustomAlgorithm(
+                    _rng,
+                    _obj,
+                    def,
+                )
+
+                ## Sample on its own
+                customconstruct = CustomAlgorithm(:μ) #CustomAlgorithm(keys(_obj.model.val))
+                trace, algorithms = sample(_rng, _obj.model, _obj.data, customconstruct ; default = deepcopy(sampledefault))
+                trace.val
+
+                ## Combine with MCMC
+                mcmc = MCMC(NUTS,(:σ,); stepsize = ConfigStepsize(;stepsizeadaption = UpdateFalse()))
+                trace, algorithms = sample(_rng, _obj.model, _obj.data, customconstruct, mcmc ; default = deepcopy(sampledefault))
+                trace.val
+
+                ## Use as Propagation Kernel in SMC
+                ibis = SMCConstructor(customconstruct, SMCDefault(jitterthreshold=0.99, resamplingthreshold=1.0))
+                trace, algorithms = sample(_rng, _obj.model, _obj.data, ibis; default = deepcopy(sampledefault))
+                trace.val
+                ## Always update Gradient Result if new data is added
+                    #!NOTE: But after first iteration, can capture results
+                    @test isa(trace.summary.info.captured, UpdateFalse)
+                ## Continue sampling
+                newdat = randn(_rng, length(_obj.data)+100)
+                trace2, algorithms2 = sample!(100, _rng, _obj.model, newdat, trace, algorithms)
+                    #!NOTE: But after first iteration, can capture results
+                @test isa(trace2.summary.info.captured, UpdateFalse)
+        end
+    end
+end