implement fitting cholesky factors (#5)

* implement fitting cholesky factors

including filling and extracting tridiagonal matrices on gpu and corresponding rrules for Zygote

* run Aqua only on version 1.11.2
* avoid testing CUDA code on machines without CUDA

Author: bgctw

Project.toml

@@ -4,8 +4,12 @@ authors = ["Thomas Wutzler <twutz@bgc-jena.mpg.de> and contributors"]
 version = "1.0.0-DEV"
 
 [deps]
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Combinatorics = "861a8166-3701-5b0c-9a16-15d98fcdc6aa"
 ComponentArrays = "b0b7db55-cfe3-40fc-9ded-d10e2dbeff66"
+GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
@@ -21,9 +25,13 @@ HybridVariationalInferenceLuxExt = "Lux"
 HybridVariationalInferenceSimpleChainsExt = "SimpleChains"
 
 [compat]
+ChainRulesCore = "1.25"
+CUDA = "5.5.2"
 Combinatorics = "1.0.2"
 ComponentArrays = "0.15.19"
 Flux = "v0.15.2"
+GPUArraysCore = "0.1, 0.2"
+LinearAlgebra = "1.10.0"
 Lux = "1.4.2"
 Random = "1.10.0"
 SimpleChains = "0.4"

README.md

@@ -9,7 +9,7 @@
 Extending Variational Inference (VI), an approximate bayesian inversion method,
 to hybrid models, i.e. models that combine mechanistic and machine-learning parts.
 
-The model inversion, inferes parametric approximations of posterior density
+The model inversion, infers parametric approximations of posterior density
 of model parameters, by comparing model outputs to uncertain observations. At
 the same time, a machine learning model is fit that predicts parameters of these
 approximations by covariates.

dev/Project.toml

@@ -1,5 +1,6 @@
 [deps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+ChainRulesTestUtils = "cdddcdb0-9152-4a09-a978-84456f9df70a"
 ComponentArrays = "b0b7db55-cfe3-40fc-9ded-d10e2dbeff66"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
 GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
@@ -11,5 +12,6 @@ SimpleChains = "de6bee2f-e2f4-4ec7-b6ed-219cc6f6e9e5"
 StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
+TransformVariables = "84d833dd-6860-57f9-a1a7-6da5db126cff"
 UnicodePlots = "b8865327-cd53-5732-bb35-84acbb429228"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"

dev/doubleMM.jl

@@ -65,28 +65,245 @@ loss_g(ϕg_opt1, xM, g)
 scatterplot(vec(θMs_true), vec(loss_g(ϕg_opt1, xM, g)[2]))
 @test cor(vec(θMs_true), vec(loss_g(ϕg_opt1, xM, g)[2])) > 0.9
 
-#----------- fit g and θP to y_o
-f = gen_hybridcase_PBmodel(case; scenario)
+tmpf = () -> begin
+    #----------- fit g and θP to y_o
+    # end2end inversion
+    f = gen_hybridcase_PBmodel(case; scenario)
 
-int_ϕθP = ComponentArrayInterpreter(CA.ComponentVector(
-    ϕg = 1:length(ϕg0), θP = par_templates.θP))
-p = p0 = vcat(ϕg0, par_templates.θP .* 0.9);  # slightly disturb θP_true
+    int_ϕθP = ComponentArrayInterpreter(CA.ComponentVector(
+        ϕg = 1:length(ϕg0), θP = par_templates.θP))
+    p = p0 = vcat(ϕg0, par_templates.θP .* 0.9);  # slightly disturb θP_true
 
-# Pass the site-data for the batches as separate vectors wrapped in a tuple
-train_loader = MLUtils.DataLoader((xM, xP, y_o), batchsize = n_batch)
+    # Pass the site-data for the batches as separate vectors wrapped in a tuple
+    train_loader = MLUtils.DataLoader((xM, xP, y_o), batchsize = n_batch)
 
-loss_gf = get_loss_gf(g, f, y_global_o, int_ϕθP)
-l1 = loss_gf(p0, train_loader.data...)[1]
+    loss_gf = get_loss_gf(g, f, y_global_o, int_ϕθP)
+    l1 = loss_gf(p0, train_loader.data...)[1]
 
-optf = Optimization.OptimizationFunction((ϕ, data) -> loss_gf(ϕ, data...)[1],
+    optf = Optimization.OptimizationFunction((ϕ, data) -> loss_gf(ϕ, data...)[1],
+        Optimization.AutoZygote())
+    optprob = OptimizationProblem(optf, p0, train_loader)
+
+    res = Optimization.solve(
+        optprob, Adam(0.02), callback = callback_loss(100), maxiters = 1000);
+
+    l1, y_pred_global, y_pred, θMs = loss_gf(res.u, train_loader.data...)
+    scatterplot(vec(θMs_true), vec(θMs))
+    scatterplot(log.(vec(θMs_true)), log.(vec(θMs)))
+    scatterplot(vec(y_pred), vec(y_o))
+    hcat(par_templates.θP, int_ϕθP(res.u).θP)
+end
+
+#---------- HADVI
+# TODO think about good general initializations
+coef_logσ2_logMs = [-5.769 -3.501; -0.01791 0.007951]
+logσ2_logP = CA.ComponentVector(r0=-8.997, K2=-5.893)
+mean_σ_o_MC = 0.006042
+
+# correlation matrices
+ρsP = zeros(sum(1:(n_θP-1)))
+ρsM = zeros(sum(1:(n_θM-1)))
+
+ϕunc = CA.ComponentVector(;
+    logσ2_logP=logσ2_logP,
+    coef_logσ2_logMs=coef_logσ2_logMs,
+    ρsP,
+    ρsM)
+int_unc = ComponentArrayInterpreter(ϕunc)
+
+# for a conservative uncertainty assume σ2=1e-10 and no relationship with magnitude
+ϕunc0 = CA.ComponentVector(;
+    logσ2_logP=fill(-10.0, n_θP),
+    coef_logσ2_logMs=reduce(hcat, ([-10.0, 0.0] for _ in 1:n_θM)),
+    ρsP,
+    ρsM)
+
+logσ2y = fill(2 * log(σ_o), size(y_o, 1))
+n_MC = 3
+
+
+#-------------- ADVI with g inside cost function
+using CUDA
+using TransformVariables
+
+transPMs_batch = as(
+    (P=as(Array, asℝ₊, n_θP),
+    Ms=as(Array, asℝ₊, n_θM, n_batch)))
+transPMs_all = as(
+    (P=as(Array, asℝ₊, n_θP),
+    Ms=as(Array, asℝ₊, n_θM, n_site)))
+    
+ϕ_true = θ = CA.ComponentVector(;
+    μP=θP_true,
+    ϕg=ϕg_opt,
+    unc=ϕunc);
+trans_gu = as(
+    (μP=as(Array, asℝ₊, n_θP),
+    ϕg=as(Array, n_ϕg),
+    unc=as(Array, length(ϕunc))))
+trans_g = as(
+    (μP=as(Array, asℝ₊, n_θP),
+    ϕg=as(Array, n_ϕg)))
+
+const int_PMs_batch = ComponentArrayInterpreter(CA.ComponentVector(; θP,
+    θMs=CA.ComponentMatrix(
+        zeros(n_θM, n_batch), first(CA.getaxes(θM)), CA.Axis(i=1:n_batch))))
+
+interpreters = interpreters_g = map(get_concrete,(; 
+    μP_ϕg_unc=ComponentArrayInterpreter(ϕ_true), 
+    PMs=int_PMs_batch,
+    unc=ComponentArrayInterpreter(ϕunc)
+    ))
+
+ϕg_true_vec = CA.ComponentVector(
+    TransformVariables.inverse(trans_gu, cv2NamedTuple(ϕ_true)))
+ϕcg_true = interpreters.μP_ϕg_unc(ϕg_true_vec)
+ϕ_ini = ζ = vcat(ϕcg_true[[:μP, :ϕg]] .* 1.2, ϕcg_true[[:unc]]);
+ϕ_ini0 = ζ = vcat(ϕcg_true[:μP] .* 0.0, SimpleChains.init_params(g), ϕunc0);
+
+neg_elbo_transnorm_gf(rng, g, f, ϕcg_true, y_o[:, 1:n_batch], x_o[:, 1:n_batch],
+    transPMs_batch, map(get_concrete, interpreters);
+    n_MC=8, logσ2y)
+Zygote.gradient(ϕ -> neg_elbo_transnorm_gf(
+        rng, g, f, ϕ, y_o[:, 1:n_batch], x_o[:, 1:n_batch],
+        transPMs_batch, interpreters; n_MC=8, logσ2y), ϕcg_true)
+
+() -> begin
+    train_loader = MLUtils.DataLoader((x_o, y_o), batchsize = n_batch)
+
+    optf = Optimization.OptimizationFunction((ζg, data) -> begin
+            x_o, y_o = data
+            neg_elbo_transnorm_gf(
+                rng, g, f, ζg, y_o, x_o, transPMs_batch, map(get_concrete, interpreters_g); n_MC=5, logσ2y)
+        end,
+        Optimization.AutoZygote())
+    optprob = Optimization.OptimizationProblem(optf, CA.getdata(ϕ_ini), train_loader);
+    res = Optimization.solve(optprob, Optimisers.Adam(0.02), callback=callback_loss(50), maxiters=800);
+    #optprob = Optimization.OptimizationProblem(optf, ϕ_ini0);
+    #res = Optimization.solve(optprob, Adam(0.02), callback=callback_loss(50), maxiters=1_400);
+end
+        
+#using Lux
+ϕ = ϕcg_true |> gpu;
+x_o_gpu = x_o |> gpu;
+# y_o = y_o |> gpu
+# logσ2y = logσ2y |> gpu
+n_covar = size(x_o, 1)
+g_flux = Flux.Chain(
+    # dense layer with bias that maps to 8 outputs and applies `tanh` activation
+    Flux.Dense(n_covar => n_covar * 4, tanh),
+    Flux.Dense(n_covar * 4 => n_covar * 4, logistic),
+    # dense layer without bias that maps to n outputs and `identity` activation
+    Flux.Dense(n_covar * 4 => n_θM, identity, bias=false),
+)
+() -> begin
+    using Lux
+    g_lux = Lux.Chain(
+        # dense layer with bias that maps to 8 outputs and applies `tanh` activation
+        Lux.Dense(n_covar => n_covar * 4, tanh),
+        Lux.Dense(n_covar * 4 => n_covar * 4, logistic),
+        # dense layer without bias that maps to n outputs and `identity` activation
+        Lux.Dense(n_covar * 4 => n_θM, identity, use_bias=false),
+    )
+    ps, st = Lux.setup(Random.default_rng(), g_lux)
+    ps_ca = CA.ComponentArray(ps) |> gpu
+    st = st |> gpu
+    g_luxs = StatefulLuxLayer{true}(g_lux, nothing, st)
+    g_luxs(x_o_gpu[:, 1:n_batch], ps_ca)
+    ax_g = CA.getaxes(ps_ca)
+    g_luxs(x_o_gpu[:, 1:n_batch], CA.ComponentArray(ϕ.ϕg, ax_g))
+    interpreters = (interpreters..., ϕg = ComponentArrayInterpreter(ps_ca))
+    ϕg = CA.ComponentArray(ϕ.ϕg, ax_g)
+    ϕgc = interpreters.ϕg(ϕ.ϕg)
+    g_gpu = g_luxs
+end
+g_gpu = g_flux
+
+#Zygote.gradient(ϕg -> sum(g_gpu(x_o_gpu[:, 1:n_batch],ϕg)), ϕgc)
+# Zygote.gradient(ϕg -> sum(compute_g(g_gpu, x_o_gpu[:, 1:n_batch], ϕg, interpreters)), ϕ.ϕg)
+# Zygote.gradient(ϕ -> sum(tmp_gen1(g_gpu, x_o_gpu[:, 1:n_batch], ϕ, interpreters)), ϕ.ϕg)
+# Zygote.gradient(ϕ -> sum(tmp_gen2(g_gpu, x_o_gpu[:, 1:n_batch], ϕ, interpreters)), CA.getdata(ϕ))
+# Zygote.gradient(ϕ -> sum(tmp_gen2(g_gpu, x_o_gpu[:, 1:n_batch], ϕ, interpreters)), ϕ) |> cpu
+# Zygote.gradient(ϕ -> sum(tmp_gen3(g_gpu, x_o_gpu[:, 1:n_batch], ϕ, interpreters)), ϕ) |> cpu
+# Zygote.gradient(ϕ -> sum(tmp_gen4(g_gpu, x_o_gpu[:, 1:n_batch], ϕ, interpreters)[1]), ϕ) |> cpu
+# generate_ζ(rng, g_gpu, f, ϕ, x_o_gpu[:, 1:n_batch], interpreters)
+# Zygote.gradient(ϕ -> sum(generate_ζ(rng, g_gpu, f, ϕ, x_o_gpu[:, 1:n_batch], interpreters)[1]), ϕ) |> cpu
+# include(joinpath(@__DIR__, "uncNN", "elbo.jl")) # callback_loss
+# neg_elbo_transnorm_gf(rng, g_gpu, f, ϕ, y_o[:, 1:n_batch], 
+#     x_o_gpu[:, 1:n_batch], transPMs_batch, interpreters; logσ2y)
+# Zygote.gradient(ϕ -> sum(neg_elbo_transnorm_gf(rng, g_gpu, f, ϕ, y_o[:, 1:n_batch], 
+# x_o_gpu[:, 1:n_batch], transPMs_batch, interpreters; logσ2y)[1]), ϕ) |> cpu
+
+
+fcost(ϕ) = neg_elbo_transnorm_gf(rng, g_gpu, f, ϕ, y_o[:, 1:n_batch], 
+    x_o_gpu[:, 1:n_batch], transPMs_batch, map(get_concrete, interpreters);
+    n_MC=8, logσ2y = logσ2y)
+fcost(ϕ)
+gr = Zygote.gradient(fcost, ϕ) |> cpu;
+Zygote.gradient(fcost, CA.getdata(ϕ))
+
+
+train_loader = MLUtils.DataLoader((x_o_gpu, y_o), batchsize = n_batch)
+
+optf = Optimization.OptimizationFunction((ζg, data) -> begin
+        x_o, y_o = data
+        neg_elbo_transnorm_gf(
+            rng, g_gpu, f, ζg, y_o, x_o, transPMs_batch, map(get_concrete, interpreters_g); n_MC=5, logσ2y)
+    end,
     Optimization.AutoZygote())
-optprob = OptimizationProblem(optf, p0, train_loader)
+optprob = Optimization.OptimizationProblem(optf, CA.getdata(ϕ_ini) |> gpu, train_loader);
+res = res_gpu = Optimization.solve(optprob, Optimisers.Adam(0.02), callback=callback_loss(50), maxiters=800);
+
+ζ_VIc = interpreters_g.μP_ϕg_unc(res.u |> cpu)
+ζMs_VI = g(x_o, ζ_VIc.ϕg)
+ϕunc_VI = int_unc(ζ_VIc.unc)
+
+hcat(θP_true, exp.(ζ_VIc.μP))
+plt = scatterplot(vec(θMs_true), vec(exp.(ζMs_VI)))
+#lineplot!(plt, 0.0, 1.1, identity)
+# 
+hcat(ϕunc, ϕunc_VI) # need to compare to MC sample
+# hard to estimate for original very small theta's but otherwise good
+
+# test predicting correct obs-uncertainty of predictive posterior
+n_sample_pred = 200
+intm_PMs_gen = ComponentArrayInterpreter(CA.ComponentVector(; θP,
+    θMs=CA.ComponentMatrix(
+        zeros(n_θM, n_site), first(CA.getaxes(θM)), CA.Axis(i=1:n_sample_pred))))
+
+include(joinpath(@__DIR__, "uncNN", "elbo.jl")) # callback_loss
+ζs, _ = generate_ζ(rng, g, f, res.u |> cpu, x_o, 
+    (;interpreters..., PMs = intm_PMs_gen); n_MC=n_sample_pred)
+# ζ = ζs[:,1]   
+θsc = stack(ζ -> CA.getdata(CA.ComponentVector(
+        TransformVariables.transform(transPMs_all, ζ))), eachcol(ζs));
+y_pred = stack(map(ζ -> first(predict_y(ζ, f, transPMs_all)), eachcol(ζs)));
+
+size(y_pred)
+σ_o_post = mapslices(std, y_pred; dims=3);
+#describe(σ_o_post)
+vcat(σ_o, mean_σ_o_MC, mean(σ_o_post), sqrt(mean(abs2, σ_o_post)))
+mean_y_pred = map(mean, eachslice(y_pred; dims=(1, 2)))
+#describe(mean_y_pred - y_o)
+histogram(vec(mean_y_pred - y_true)) # predictions centered around y_o (or y_true)
+
+# look at θP, θM1 of first site
+intm = ComponentArrayInterpreter(int_θdoubleMM(1:length(int_θdoubleMM)), (n_sample_pred,))
+ζs1c = intm(ζs[1:length(int_θdoubleMM), :])
+vcat(θP_true, θM_true)
+histogram(exp.(ζs1c[:r0, :]))
+histogram(exp.(ζs1c[:K2, :]))
+histogram(exp.(ζs1c[:r1, :]))
+histogram(exp.(ζs1c[:K1, :]))
+# all parameters estimated to high (true not in cf bounds)
+scatterplot(ζs1c[:r1, :], ζs1c[:K1, :])  # r1 and K1 strongly correlated (from θM)
+scatterplot(ζs1c[:r0, :], ζs1c[:K2, :])  # r0 and K also correlated (from θP)
+scatterplot(ζs1c[:r0, :], ζs1c[:K1, :])  # no correlation (modeled independent)
+
+# TODO compare distributions to MC sample
+
+
+
+
 
-res = Optimization.solve(
-    optprob, Adam(0.02), callback = callback_loss(100), maxiters = 1000);
 
-l1, y_pred_global, y_pred, θMs = loss_gf(res.u, train_loader.data...)
-scatterplot(vec(θMs_true), vec(θMs))
-scatterplot(log.(vec(θMs_true)), log.(vec(θMs)))
-scatterplot(vec(y_pred), vec(y_o))
-hcat(par_templates.θP, int_ϕθP(res.u).θP)

src/HybridVariationalInference.jl

@@ -4,6 +4,10 @@ using ComponentArrays: ComponentArrays as CA
 using Random
 using StatsBase # fit ZScoreTransform
 using Combinatorics # gen_hybridcase_synthetic/combinations
+using GPUArraysCore
+using LinearAlgebra
+using CUDA
+using ChainRulesCore
 
 export ComponentArrayInterpreter, flatten1, get_concrete
 include("ComponentArrayInterpreter.jl")
@@ -25,6 +29,9 @@ include("gencovar.jl")
 export callback_loss
 include("util_opt.jl")
 
+#export - all internal
+include("cholesky.jl")
+
 export DoubleMM
 include("DoubleMM/DoubleMM.jl")