Variational inference (VI) in Turing.jl

In this post we’ll have a look at what’s know as variational inference (VI), a family of approximate Bayesian inference methods, and how to use it in Turing.jl as an alternative to other approaches such as MCMC. In particular, we will focus on one of the more standard VI methods called Automatic Differentation Variational Inference (ADVI).

Here we will focus on how to use VI in Turing and not much on the theory underlying VI. If you are interested in understanding the mathematics you can checkout our write-up or any other resource online (there a lot of great ones).

Using VI in Turing.jl is very straight forward. If model denotes a definition of a Turing.Model, performing VI is as simple as

m = model(data...) # instantiate model on the data
q = vi(m, vi_alg)  # perform VI on `m` using the VI method `vi_alg`, which returns a `VariationalPosterior`

Thus it’s no more work than standard MCMC sampling in Turing.

To get a bit more into what we can do with vi, we’ll first have a look at a simple example and then we’ll reproduce the tutorial on Bayesian linear regression using VI instead of MCMC. Finally we’ll look at some of the different parameters of vi and how you for example can use your own custom variational family.

We first import the packages to be used:

using Random
using Turing
using Turing: Variational
using StatsPlots, Measures

Random.seed!(42);

Simple example: Normal-Gamma conjugate model

The Normal-(Inverse)Gamma conjugate model is defined by the following generative process

\[\begin{align} s &\sim \mathrm{InverseGamma}(2, 3) \\ m &\sim \mathcal{N}(0, s) \\ x_i &\overset{\text{i.i.d.}}{=} \mathcal{N}(m, s), \quad i = 1, \dots, n \end{align}\]

Recall that conjugate refers to the fact that we can obtain a closed-form expression for the posterior. Of course one wouldn’t use something like variational inference for a conjugate model, but it’s useful as a simple demonstration as we can compare the result to the true posterior.

First we generate some synthetic data, define the Turing.Model and instantiate the model on the data:

# generate data
x = randn(2000);
@model function model(x)
    s ~ InverseGamma(2, 3)
    m ~ Normal(0.0, sqrt(s))
    for i in 1:length(x)
        x[i] ~ Normal(m, sqrt(s))
    end
end;
# Instantiate model
m = model(x);

Now we’ll produce some samples from the posterior using a MCMC method, which in constrast to VI is guaranteed to converge to the exact posterior (as the number of samples go to infinity).

We’ll produce 10 000 samples with 200 steps used for adaptation and a target acceptance rate of 0.65

If you don’t understand what “adaptation” or “target acceptance rate” refers to, all you really need to know is that NUTS is known to be one of the most accurate and efficient samplers (when applicable) while requiring little to no hand-tuning to work well.

setprogress!(false)
samples_nuts = sample(m, NUTS(), 10_000);
┌ Info: Found initial step size
└   ϵ = 0.025

Now let’s try VI. The most important function you need to now about to do VI in Turing is vi:

@doc(Variational.vi)
vi(model, alg::VariationalInference)
vi(model, alg::VariationalInference, q::VariationalPosterior)
vi(model, alg::VariationalInference, getq::Function, θ::AbstractArray)

Constructs the variational posterior from the model and performs the optimization following the configuration of the given VariationalInference instance.

Arguments

  • model: Turing.Model or Function z ↦ log p(x, z) where x denotes the observations

  • alg: the VI algorithm used

  • q: a VariationalPosterior for which it is assumed a specialized implementation of the variational objective used exists.

  • getq: function taking parameters θ as input and returns a VariationalPosterior

  • θ: only required if getq is used, in which case it is the initial parameters for the variational posterior

Additionally, you can pass

  • an initial variational posterior q, for which we assume there exists a implementation of update(::typeof(q), θ::AbstractVector) returning an updated posterior q with parameters θ.
  • a function mapping \(\theta \mapsto q_{\theta}\) (denoted above getq) together with initial parameters θ. This provides more flexibility in the types of variational families that we can use, and can sometimes be slightly more convenient for quick and rough work.

By default, i.e. when calling vi(m, advi), Turing use a mean-field approximation with a multivariate normal as the base-distribution. Mean-field refers to the fact that we assume all the latent variables to be independent. This the “standard” ADVI approach; see Automatic Differentiation Variational Inference (2016) for more. In Turing, one can obtain such a mean-field approximation by calling Variational.meanfield(model) for which there exists an internal implementation for update:

@doc(Variational.meanfield)
meanfield([rng, ]model::Model)

Creates a mean-field approximation with multivariate normal as underlying distribution.

Currently the only implementation of VariationalInference available is ADVI, which is very convenient and applicable as long as your Model is differentiable with respect to the variational parameters, that is, the parameters of your variational distribution, e.g. mean and variance in the mean-field approximation.

@doc(Variational.ADVI)
struct ADVI{AD} <: AdvancedVI.VariationalInference{AD}

Automatic Differentiation Variational Inference (ADVI) with automatic differentiation backend AD.

Fields

  • samples_per_step::Int64: Number of samples used to estimate the ELBO in each optimization step.

  • max_iters::Int64: Maximum number of gradient steps.

  • adtype::Any: AD backend used for automatic differentiation.

To perform VI on the model m using 10 samples for gradient estimation and taking 1000 gradient steps is then as simple as:

# ADVI
advi = ADVI(10, 1000)
q = vi(m, advi);
┌ Info: [ADVI] Should only be seen once: optimizer created for θ
└   objectid(θ) = 0xd283039726b15fc0

Unfortunately, for such a small problem Turing’s new NUTS sampler is so efficient now that it’s not that much more efficient to use ADVI. So, so very unfortunate…

With that being said, this is not the case in general. For very complex models we’ll later find that ADVI produces very reasonable results in a much shorter time than NUTS.

And one significant advantage of using vi is that we can sample from the resulting q with ease. In fact, the result of the vi call is a TransformedDistribution from Bijectors.jl, and it implements the Distributions.jl interface for a Distribution:

q isa MultivariateDistribution
true

This means that we can call rand to sample from the variational posterior q

histogram(rand(q, 1_000)[1, :])

and logpdf to compute the log-probability

logpdf(q, rand(q))
4.560747878869669

Let’s check the first and second moments of the data to see how our approximation compares to the point-estimates form the data:

var(x), mean(x)
(1.0133414191151338, 0.0030541280207364353)
(mean(rand(q, 1000); dims=2)...,)
(1.022471063520294, 0.0001596876848256648)

That’s pretty close! But we’re Bayesian so we’re not interested in just matching the mean. Let’s instead look the actual density q.

For that we need samples:

samples = rand(q, 10000);
size(samples)
(2, 10000)
p1 = histogram(
    samples[1, :]; bins=100, normed=true, alpha=0.2, color=:blue, label="", ylabel="density"
)
density!(samples[1, :]; label="s (ADVI)", color=:blue, linewidth=2)
density!(samples_nuts, :s; label="s (NUTS)", color=:green, linewidth=2)
vline!([var(x)]; label="s (data)", color=:black)
vline!([mean(samples[1, :])]; color=:blue, label="")

p2 = histogram(
    samples[2, :]; bins=100, normed=true, alpha=0.2, color=:blue, label="", ylabel="density"
)
density!(samples[2, :]; label="m (ADVI)", color=:blue, linewidth=2)
density!(samples_nuts, :m; label="m (NUTS)", color=:green, linewidth=2)
vline!([mean(x)]; color=:black, label="m (data)")
vline!([mean(samples[2, :])]; color=:blue, label="")

plot(p1, p2; layout=(2, 1), size=(900, 500), legend=true)

For this particular Model, we can in fact obtain the posterior of the latent variables in closed form. This allows us to compare both NUTS and ADVI to the true posterior \(p(s, m \mid x_1, \ldots, x_n)\).

The code below is just work to get the marginals \(p(s \mid x_1, \ldots, x_n)\) and \(p(m \mid x_1, \ldots, x_n)\). Feel free to skip it.

# closed form computation of the Normal-inverse-gamma posterior
# based on "Conjugate Bayesian analysis of the Gaussian distribution" by Murphy
function posterior(μ₀::Real, κ₀::Real, α₀::Real, β₀::Real, x::AbstractVector{<:Real})
    # Compute summary statistics
    n = length(x)
    x̄ = mean(x)
    sum_of_squares = sum(xi -> (xi - x̄)^2, x)

    # Compute parameters of the posterior
    κₙ = κ₀ + n
    μₙ = (κ₀ * μ₀ + n * x̄) / κₙ
    αₙ = α₀ + n / 2
    βₙ = β₀ + (sum_of_squares + n * κ₀ / κₙ * (x̄ - μ₀)^2) / 2

    return μₙ, κₙ, αₙ, βₙ
end
μₙ, κₙ, αₙ, βₙ = posterior(0.0, 1.0, 2.0, 3.0, x)

# marginal distribution of σ²
# cf. Eq. (90) in "Conjugate Bayesian analysis of the Gaussian distribution" by Murphy
p_σ² = InverseGamma(αₙ, βₙ)
p_σ²_pdf = z -> pdf(p_σ², z)

# marginal of μ
# Eq. (91) in "Conjugate Bayesian analysis of the Gaussian distribution" by Murphy
p_μ = μₙ + sqrt(βₙ / (αₙ * κₙ)) * TDist(2 * αₙ)
p_μ_pdf = z -> pdf(p_μ, z)

# posterior plots
p1 = plot()
histogram!(samples[1, :]; bins=100, normed=true, alpha=0.2, color=:blue, label="")
density!(samples[1, :]; label="s (ADVI)", color=:blue)
density!(samples_nuts, :s; label="s (NUTS)", color=:green)
vline!([mean(samples[1, :])]; linewidth=1.5, color=:blue, label="")
plot!(range(0.75, 1.35; length=1_001), p_σ²_pdf; label="s (posterior)", color=:red)
vline!([var(x)]; label="s (data)", linewidth=1.5, color=:black, alpha=0.7)
xlims!(0.75, 1.35)

p2 = plot()
histogram!(samples[2, :]; bins=100, normed=true, alpha=0.2, color=:blue, label="")
density!(samples[2, :]; label="m (ADVI)", color=:blue)
density!(samples_nuts, :m; label="m (NUTS)", color=:green)
vline!([mean(samples[2, :])]; linewidth=1.5, color=:blue, label="")
plot!(range(-0.25, 0.25; length=1_001), p_μ_pdf; label="m (posterior)", color=:red)
vline!([mean(x)]; label="m (data)", linewidth=1.5, color=:black, alpha=0.7)
xlims!(-0.25, 0.25)

plot(p1, p2; layout=(2, 1), size=(900, 500))

Bayesian linear regression example using ADVI

This is simply a duplication of the tutorial 5. Linear regression but now with the addition of an approximate posterior obtained using ADVI.

As we’ll see, there is really no additional work required to apply variational inference to a more complex Model.

This section is basically copy-pasting the code from the linear regression tutorial.

Random.seed!(1);
using FillArrays
using RDatasets

using LinearAlgebra
# Import the "Default" dataset.
data = RDatasets.dataset("datasets", "mtcars");

# Show the first six rows of the dataset.
first(data, 6)
6×12 DataFrame
Row Model MPG Cyl Disp HP DRat WT QSec VS AM Gear Carb
String31 Float64 Int64 Float64 Int64 Float64 Float64 Float64 Int64 Int64 Int64 Int64
1 Mazda RX4 21.0 6 160.0 110 3.9 2.62 16.46 0 1 4 4
2 Mazda RX4 Wag 21.0 6 160.0 110 3.9 2.875 17.02 0 1 4 4
3 Datsun 710 22.8 4 108.0 93 3.85 2.32 18.61 1 1 4 1
4 Hornet 4 Drive 21.4 6 258.0 110 3.08 3.215 19.44 1 0 3 1
5 Hornet Sportabout 18.7 8 360.0 175 3.15 3.44 17.02 0 0 3 2
6 Valiant 18.1 6 225.0 105 2.76 3.46 20.22 1 0 3 1 
# Function to split samples.
function split_data(df, at=0.70)
    r = size(df, 1)
    index = Int(round(r * at))
    train = df[1:index, :]
    test = df[(index + 1):end, :]
    return train, test
end

# A handy helper function to rescale our dataset.
function standardize(x)
    return (x .- mean(x; dims=1)) ./ std(x; dims=1)
end

function standardize(x, orig)
    return (x .- mean(orig; dims=1)) ./ std(orig; dims=1)
end

# Another helper function to unstandardize our datasets.
function unstandardize(x, orig)
    return x .* std(orig; dims=1) .+ mean(orig; dims=1)
end

function unstandardize(x, mean_train, std_train)
    return x .* std_train .+ mean_train
end
unstandardize (generic function with 2 methods)
# Remove the model column.
select!(data, Not(:Model))

# Split our dataset 70%/30% into training/test sets.
train, test = split_data(data, 0.7)
train_unstandardized = copy(train)

# Standardize both datasets.
std_train = standardize(Matrix(train))
std_test = standardize(Matrix(test), Matrix(train))

# Save dataframe versions of our dataset.
train_cut = DataFrame(std_train, names(data))
test_cut = DataFrame(std_test, names(data))

# Create our labels. These are the values we are trying to predict.
train_label = train_cut[:, :MPG]
test_label = test_cut[:, :MPG]

# Get the list of columns to keep.
remove_names = filter(x -> !in(x, ["MPG"]), names(data))

# Filter the test and train sets.
train = Matrix(train_cut[:, remove_names]);
test = Matrix(test_cut[:, remove_names]);
# Bayesian linear regression.
@model function linear_regression(x, y, n_obs, n_vars, ::Type{T}=Vector{Float64}) where {T}
    # Set variance prior.
    σ² ~ truncated(Normal(0, 100), 0, Inf)

    # Set intercept prior.
    intercept ~ Normal(0, 3)

    # Set the priors on our coefficients.
    coefficients ~ MvNormal(Zeros(n_vars), 10.0 * I)

    # Calculate all the mu terms.
    mu = intercept .+ x * coefficients
    return y ~ MvNormal(mu, σ² * I)
end;
n_obs, n_vars = size(train)
m = linear_regression(train, train_label, n_obs, n_vars);

Performing VI

First we define the initial variational distribution, or, equivalently, the family of distributions to consider. We’re going to use the same mean-field approximation as Turing will use by default when we call vi(m, advi), which we obtain by calling Variational.meanfield. This returns a TransformedDistribution with a TuringDiagMvNormal as the underlying distribution and the transformation mapping from the reals to the domain of the latent variables.

q0 = Variational.meanfield(m)
typeof(q0)
MultivariateTransformed{TuringDiagMvNormal{Vector{Float64}, Vector{Float64}}, Stacked{Vector{Any}, Vector{UnitRange{Int64}}}} (alias for Bijectors.TransformedDistribution{DistributionsAD.TuringDiagMvNormal{Array{Float64, 1}, Array{Float64, 1}}, Stacked{Array{Any, 1}, Array{UnitRange{Int64}, 1}}, ArrayLikeVariate{1}})
advi = ADVI(10, 10_000)
ADVI{AutoForwardDiff{nothing, Nothing}}(10, 10000, AutoForwardDiff{nothing, Nothing}(nothing))

Turing also provides a couple of different optimizers:

  • TruncatedADAGrad (default)
  • DecayedADAGrad as these are well-suited for problems with high-variance stochastic objectives, which is usually what the ELBO ends up being at different times in our optimization process.

With that being said, thanks to Requires.jl, if we add a using Flux prior to using Turing we can also make use of all the optimizers in Flux, e.g. ADAM, without any additional changes to your code! For example:

using Flux, Turing
using Turing.Variational

vi(m, advi; optimizer=Flux.ADAM())

just works.

For this problem we’ll use the DecayedADAGrad from Turing:

opt = Variational.DecayedADAGrad(1e-2, 1.1, 0.9)
AdvancedVI.DecayedADAGrad(0.01, 1.1, 0.9, IdDict{Any, Any}())
q = vi(m, advi, q0; optimizer=opt)
typeof(q)
MultivariateTransformed{TuringDiagMvNormal{Vector{Float64}, Vector{Float64}}, Stacked{Vector{Any}, Vector{UnitRange{Int64}}}} (alias for Bijectors.TransformedDistribution{DistributionsAD.TuringDiagMvNormal{Array{Float64, 1}, Array{Float64, 1}}, Stacked{Array{Any, 1}, Array{UnitRange{Int64}, 1}}, ArrayLikeVariate{1}})

Note: as mentioned before, we internally define a update(q::TransformedDistribution{<:TuringDiagMvNormal}, θ::AbstractVector) method which takes in the current variational approximation q together with new parameters z and returns the new variational approximation. This is required so that we can actually update the Distribution object after each optimization step.

Alternatively, we can instead provide the mapping \(\theta \mapsto q_{\theta}\) directly together with initial parameters using the signature vi(m, advi, getq, θ_init) as mentioned earlier. We’ll see an explicit example of this later on!

To compute statistics for our approximation we need samples:

z = rand(q, 10_000);

Now we can for example look at the average

avg = vec(mean(z; dims=2))
12-element Vector{Float64}:
  0.23280752308971772
  0.006427909049664767
  0.3665619112251093
 -0.11046793167996063
 -0.08101113817995316
  0.613745932095346
 -0.00022773341507825883
  0.07996409486126356
 -0.06803340755087516
  0.13443570902038943
  0.1929652767406172
 -0.6166685902366881

The vector has the same ordering as the model, e.g. in this case σ² has index 1, intercept has index 2 and coefficients has indices 3:12. If you forget or you might want to do something programmatically with the result, you can obtain the sym → indices mapping as follows:

_, sym2range = bijector(m, Val(true));
sym2range
(intercept = UnitRange{Int64}[2:2], σ² = UnitRange{Int64}[1:1], coefficients = UnitRange{Int64}[3:12])

For example, we can check the sample distribution and mean value of σ²:

histogram(z[1, :])
avg[union(sym2range[:σ²]...)]
1-element Vector{Float64}:
 0.23280752308971772
avg[union(sym2range[:intercept]...)]
1-element Vector{Float64}:
 0.006427909049664767
avg[union(sym2range[:coefficients]...)]
10-element Vector{Float64}:
  0.3665619112251093
 -0.11046793167996063
 -0.08101113817995316
  0.613745932095346
 -0.00022773341507825883
  0.07996409486126356
 -0.06803340755087516
  0.13443570902038943
  0.1929652767406172
 -0.6166685902366881

Note: as you can see, this is slightly awkward to work with at the moment. We’ll soon add a better way of dealing with this.

With a bit of work (this will be much easier in the future), we can also visualize the approximate marginals of the different variables, similar to plot(chain):

function plot_variational_marginals(z, sym2range)
    ps = []

    for (i, sym) in enumerate(keys(sym2range))
        indices = union(sym2range[sym]...)  # <= array of ranges
        if sum(length.(indices)) > 1
            offset = 1
            for r in indices
                p = density(
                    z[r, :];
                    title="$(sym)[$offset]",
                    titlefontsize=10,
                    label="",
                    ylabel="Density",
                    margin=1.5mm,
                )
                push!(ps, p)
                offset += 1
            end
        else
            p = density(
                z[first(indices), :];
                title="$(sym)",
                titlefontsize=10,
                label="",
                ylabel="Density",
                margin=1.5mm,
            )
            push!(ps, p)
        end
    end

    return plot(ps...; layout=(length(ps), 1), size=(500, 2000), margin=4.0mm)
end
plot_variational_marginals (generic function with 1 method)
plot_variational_marginals(z, sym2range)

And let’s compare this to using the NUTS sampler:

chain = sample(m, NUTS(), 10_000);
┌ Info: Found initial step size
└   ϵ = 0.025
plot(chain; margin=12.00mm)
vi_mean = vec(mean(z; dims=2))[[
    union(sym2range[:coefficients]...)...,
    union(sym2range[:intercept]...)...,
    union(sym2range[:σ²]...)...,
]]
12-element Vector{Float64}:
  0.3665619112251093
 -0.11046793167996063
 -0.08101113817995316
  0.613745932095346
 -0.00022773341507825883
  0.07996409486126356
 -0.06803340755087516
  0.13443570902038943
  0.1929652767406172
 -0.6166685902366881
  0.006427909049664767
  0.23280752308971772
mcmc_mean = mean(chain, names(chain, :parameters))[:, 2]
12-element Vector{Float64}:
  0.2469120515092212
 -0.001323451531470103
  0.372183837260125
 -0.09170702881605569
 -0.09123746017513568
  0.6077324455753209
 -0.013325130316349543
  0.08334232794374836
 -0.06670846406408032
  0.12512196208089507
  0.17994962183074095
 -0.6024674699016047
plot(mcmc_mean; xticks=1:1:length(mcmc_mean), linestyle=:dot, label="NUTS")
plot!(vi_mean; linestyle=:dot, label="VI")

One thing we can look at is simply the squared error between the means:

sum(abs2, mcmc_mean .- vi_mean)
2.367181677762164

That looks pretty good! But let’s see how the predictive distributions looks for the two.

Prediction

Similarily to the linear regression tutorial, we’re going to compare to multivariate ordinary linear regression using the GLM package:

# Import the GLM package.
using GLM

# Perform multivariate OLS.
ols = lm(
    @formula(MPG ~ Cyl + Disp + HP + DRat + WT + QSec + VS + AM + Gear + Carb), train_cut
)

# Store our predictions in the original dataframe.
train_cut.OLSPrediction = unstandardize(GLM.predict(ols), train_unstandardized.MPG)
test_cut.OLSPrediction = unstandardize(GLM.predict(ols, test_cut), train_unstandardized.MPG);
# Make a prediction given an input vector, using mean parameter values from a chain.
function prediction_chain(chain, x)
    p = get_params(chain)
    α = mean(p.intercept)
    β = collect(mean.(p.coefficients))
    return α .+ x * β
end
prediction_chain (generic function with 1 method)
# Make a prediction using samples from the variational posterior given an input vector.
function prediction(samples::AbstractVector, sym2ranges, x)
    α = mean(samples[union(sym2ranges[:intercept]...)])
    β = vec(mean(samples[union(sym2ranges[:coefficients]...)]; dims=2))
    return α .+ x * β
end

function prediction(samples::AbstractMatrix, sym2ranges, x)
    α = mean(samples[union(sym2ranges[:intercept]...), :])
    β = vec(mean(samples[union(sym2ranges[:coefficients]...), :]; dims=2))
    return α .+ x * β
end
prediction (generic function with 2 methods)
# Unstandardize the dependent variable.
train_cut.MPG = unstandardize(train_cut.MPG, train_unstandardized.MPG)
test_cut.MPG = unstandardize(test_cut.MPG, train_unstandardized.MPG);
# Show the first side rows of the modified dataframe.
first(test_cut, 6)
6×12 DataFrame
Row MPG Cyl Disp HP DRat WT QSec VS AM Gear Carb OLSPrediction
Float64 Float64 Float64 Float64 Float64 Float64 Float64 Float64 Float64 Float64 Float64 Float64
1 15.2 1.04746 0.565102 0.258882 -0.652405 0.0714991 -0.716725 -0.977008 -0.598293 -0.891883 -0.469126 19.8583
2 13.3 1.04746 0.929057 1.90345 0.380435 0.465717 -1.90403 -0.977008 -0.598293 -0.891883 1.11869 16.0462
3 19.2 1.04746 1.32466 0.691663 -0.777058 0.470584 -0.873777 -0.977008 -0.598293 -0.891883 -0.469126 18.5746
4 27.3 -1.25696 -1.21511 -1.19526 1.0037 -1.38857 0.288403 0.977008 1.59545 1.07026 -1.26303 29.3233
5 26.0 -1.25696 -0.888346 -0.762482 1.62697 -1.18903 -1.09365 -0.977008 1.59545 3.0324 -0.469126 30.7731
6 30.4 -1.25696 -1.08773 -0.381634 0.451665 -1.79933 -0.968007 0.977008 1.59545 3.0324 -0.469126 25.2892 
z = rand(q, 10_000);
# Calculate the predictions for the training and testing sets using the samples `z` from variational posterior
train_cut.VIPredictions = unstandardize(
    prediction(z, sym2range, train), train_unstandardized.MPG
)
test_cut.VIPredictions = unstandardize(
    prediction(z, sym2range, test), train_unstandardized.MPG
)

train_cut.BayesPredictions = unstandardize(
    prediction_chain(chain, train), train_unstandardized.MPG
)
test_cut.BayesPredictions = unstandardize(
    prediction_chain(chain, test), train_unstandardized.MPG
);
vi_loss1 = mean((train_cut.VIPredictions - train_cut.MPG) .^ 2)
bayes_loss1 = mean((train_cut.BayesPredictions - train_cut.MPG) .^ 2)
ols_loss1 = mean((train_cut.OLSPrediction - train_cut.MPG) .^ 2)

vi_loss2 = mean((test_cut.VIPredictions - test_cut.MPG) .^ 2)
bayes_loss2 = mean((test_cut.BayesPredictions - test_cut.MPG) .^ 2)
ols_loss2 = mean((test_cut.OLSPrediction - test_cut.MPG) .^ 2)

println("Training set:
    VI loss: $vi_loss1
    Bayes loss: $bayes_loss1
    OLS loss: $ols_loss1
Test set: 
    VI loss: $vi_loss2
    Bayes loss: $bayes_loss2
    OLS loss: $ols_loss2")
Training set:
    VI loss: 3.1038739071675976
    Bayes loss: 3.073768383569947
    OLS loss: 3.0709261248930093
Test set: 
    VI loss: 26.403208513058
    Bayes loss: 25.88734633454868
    OLS loss: 27.09481307076057

Interestingly the squared difference between true- and mean-prediction on the test-set is actually better for the mean-field variational posterior than for the “true” posterior obtained by MCMC sampling using NUTS. But, as Bayesians, we know that the mean doesn’t tell the entire story. One quick check is to look at the mean predictions ± standard deviation of the two different approaches:

z = rand(q, 1000);
preds = mapreduce(hcat, eachcol(z)) do zi
    return unstandardize(prediction(zi, sym2range, test), train_unstandardized.MPG)
end

scatter(
    1:size(test, 1),
    mean(preds; dims=2);
    yerr=std(preds; dims=2),
    label="prediction (mean ± std)",
    size=(900, 500),
    markersize=8,
)
scatter!(1:size(test, 1), unstandardize(test_label, train_unstandardized.MPG); label="true")
xaxis!(1:size(test, 1))
ylims!(10, 40)
title!("Mean-field ADVI (Normal)")
preds = mapreduce(hcat, 1:5:size(chain, 1)) do i
    return unstandardize(prediction_chain(chain[i], test), train_unstandardized.MPG)
end

scatter(
    1:size(test, 1),
    mean(preds; dims=2);
    yerr=std(preds; dims=2),
    label="prediction (mean ± std)",
    size=(900, 500),
    markersize=8,
)
scatter!(1:size(test, 1), unstandardize(test_label, train_unstandardized.MPG); label="true")
xaxis!(1:size(test, 1))
ylims!(10, 40)
title!("MCMC (NUTS)")

Indeed we see that the MCMC approach generally provides better uncertainty estimates than the mean-field ADVI approach! Good. So all the work we’ve done to make MCMC fast isn’t for nothing.

Alternative: provide parameter-to-distribution instead of \(q\) with update implemented

As mentioned earlier, it’s also possible to just provide the mapping \(\theta \mapsto q_{\theta}\) rather than the variational family / initial variational posterior q, i.e. use the interface vi(m, advi, getq, θ_init) where getq is the mapping \(\theta \mapsto q_{\theta}\)

In this section we’re going to construct a mean-field approximation to the model by hand using a composition ofShift and Scale from Bijectors.jl togheter with a standard multivariate Gaussian as the base distribution.

using Bijectors
using Bijectors: Scale, Shift
d = length(q)
base_dist = Turing.DistributionsAD.TuringDiagMvNormal(zeros(d), ones(d))
DistributionsAD.TuringDiagMvNormal{Vector{Float64}, Vector{Float64}}(
m: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
σ: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
)

bijector(model::Turing.Model) is defined by Turing, and will return a bijector which takes you from the space of the latent variables to the real space. In this particular case, this is a mapping ((0, ∞) × ℝ × ℝ¹⁰) → ℝ¹². We’re interested in using a normal distribution as a base-distribution and transform samples to the latent space, thus we need the inverse mapping from the reals to the latent space:

to_constrained = inverse(bijector(m));
function getq(θ)
    d = length(θ) ÷ 2
    A = @inbounds θ[1:d]
    b = @inbounds θ[(d + 1):(2 * d)]

    b = to_constrained ∘ Shift(b) ∘ Scale(exp.(A))

    return transformed(base_dist, b)
end
getq (generic function with 1 method)
q_mf_normal = vi(m, advi, getq, randn(2 * d));
┌ Info: [ADVI] Should only be seen once: optimizer created for θ
└   objectid(θ) = 0x426e7d9036da32ba
p1 = plot_variational_marginals(rand(q_mf_normal, 10_000), sym2range) # MvDiagNormal + Affine transformation + to_constrained
p2 = plot_variational_marginals(rand(q, 10_000), sym2range)  # Turing.meanfield(m)

plot(p1, p2; layout=(1, 2), size=(800, 2000))

As expected, the fits look pretty much identical.

But using this interface it becomes trivial to go beyond the mean-field assumption we made for the variational posterior, as we’ll see in the next section.

Relaxing the mean-field assumption

Here we’ll instead consider the variational family to be a full non-diagonal multivariate Gaussian. As in the previous section we’ll implement this by transforming a standard multivariate Gaussian using Scale and Shift, but now Scale will instead be using a lower-triangular matrix (representing the Cholesky of the covariance matrix of a multivariate normal) in contrast to the diagonal matrix we used in for the mean-field approximate posterior.

# Using `ComponentArrays.jl` together with `UnPack.jl` makes our lives much easier.
using ComponentArrays, UnPack
proto_arr = ComponentArray(; L=zeros(d, d), b=zeros(d))
proto_axes = getaxes(proto_arr)
num_params = length(proto_arr)

function getq(θ)
    L, b = begin
        @unpack L, b = ComponentArray(θ, proto_axes)
        LowerTriangular(L), b
    end
    # For this to represent a covariance matrix we need to ensure that the diagonal is positive.
    # We can enforce this by zeroing out the diagonal and then adding back the diagonal exponentiated.
    D = Diagonal(diag(L))
    A = L - D + exp(D) # exp for Diagonal is the same as exponentiating only the diagonal entries

    b = to_constrained ∘ Shift(b) ∘ Scale(A)

    return transformed(base_dist, b)
end
getq (generic function with 1 method)
advi = ADVI(10, 20_000)
ADVI{AutoForwardDiff{nothing, Nothing}}(10, 20000, AutoForwardDiff{nothing, Nothing}(nothing))
q_full_normal = vi(
    m, advi, getq, randn(num_params); optimizer=Variational.DecayedADAGrad(1e-2)
);

Let’s have a look at the learned covariance matrix:

A = q_full_normal.transform.inner.a
12×12 LowerTriangular{Float64, Matrix{Float64}}:
  0.319082      ⋅             ⋅         …    ⋅         ⋅           ⋅ 
 -0.00655617   0.0956971      ⋅              ⋅         ⋅           ⋅ 
 -0.00616398  -0.00981992    0.419181        ⋅         ⋅           ⋅ 
 -0.00535257   0.00448312   -0.0357844       ⋅         ⋅           ⋅ 
 -0.00891624  -0.000871926  -0.0418763       ⋅         ⋅           ⋅ 
 -0.00522732  -0.0065781     0.118129   …    ⋅         ⋅           ⋅ 
 -0.00318409  -0.00320813    0.0221968       ⋅         ⋅           ⋅ 
  0.00503986   0.00491964    0.097568        ⋅         ⋅           ⋅ 
  0.00235194  -0.00128335    0.104888        ⋅         ⋅           ⋅ 
 -0.00998891  -0.00261624    0.0591931      0.141154   ⋅           ⋅ 
 -0.0103496    0.00444161    0.0187501  …  -0.107031  0.101531     ⋅ 
  0.00974616  -0.00205971   -0.04809        0.024989  0.00160166  0.100039
heatmap(cov(A * A'))
zs = rand(q_full_normal, 10_000);
p1 = plot_variational_marginals(rand(q_mf_normal, 10_000), sym2range)
p2 = plot_variational_marginals(rand(q_full_normal, 10_000), sym2range)

plot(p1, p2; layout=(1, 2), size=(800, 2000))

So it seems like the “full” ADVI approach, i.e. no mean-field assumption, obtain the same modes as the mean-field approach but with greater uncertainty for some of the coefficients. This

# Unfortunately, it seems like this has quite a high variance which is likely to be due to numerical instability, 
# so we consider a larger number of samples. If we get a couple of outliers due to numerical issues, 
# these kind affect the mean prediction greatly.
z = rand(q_full_normal, 10_000);
train_cut.VIFullPredictions = unstandardize(
    prediction(z, sym2range, train), train_unstandardized.MPG
)
test_cut.VIFullPredictions = unstandardize(
    prediction(z, sym2range, test), train_unstandardized.MPG
);
vi_loss1 = mean((train_cut.VIPredictions - train_cut.MPG) .^ 2)
vifull_loss1 = mean((train_cut.VIFullPredictions - train_cut.MPG) .^ 2)
bayes_loss1 = mean((train_cut.BayesPredictions - train_cut.MPG) .^ 2)
ols_loss1 = mean((train_cut.OLSPrediction - train_cut.MPG) .^ 2)

vi_loss2 = mean((test_cut.VIPredictions - test_cut.MPG) .^ 2)
vifull_loss2 = mean((test_cut.VIFullPredictions - test_cut.MPG) .^ 2)
bayes_loss2 = mean((test_cut.BayesPredictions - test_cut.MPG) .^ 2)
ols_loss2 = mean((test_cut.OLSPrediction - test_cut.MPG) .^ 2)

println("Training set:
    VI loss: $vi_loss1
    Bayes loss: $bayes_loss1
    OLS loss: $ols_loss1
Test set: 
    VI loss: $vi_loss2
    Bayes loss: $bayes_loss2
    OLS loss: $ols_loss2")
Training set:
    VI loss: 3.1038739071675976
    Bayes loss: 3.073768383569947
    OLS loss: 3.0709261248930093
Test set: 
    VI loss: 26.403208513058
    Bayes loss: 25.88734633454868
    OLS loss: 27.09481307076057
z = rand(q_mf_normal, 1000);
preds = mapreduce(hcat, eachcol(z)) do zi
    return unstandardize(prediction(zi, sym2range, test), train_unstandardized.MPG)
end

p1 = scatter(
    1:size(test, 1),
    mean(preds; dims=2);
    yerr=std(preds; dims=2),
    label="prediction (mean ± std)",
    size=(900, 500),
    markersize=8,
)
scatter!(1:size(test, 1), unstandardize(test_label, train_unstandardized.MPG); label="true")
xaxis!(1:size(test, 1))
ylims!(10, 40)
title!("Mean-field ADVI (Normal)")
z = rand(q_full_normal, 1000);
preds = mapreduce(hcat, eachcol(z)) do zi
    return unstandardize(prediction(zi, sym2range, test), train_unstandardized.MPG)
end

p2 = scatter(
    1:size(test, 1),
    mean(preds; dims=2);
    yerr=std(preds; dims=2),
    label="prediction (mean ± std)",
    size=(900, 500),
    markersize=8,
)
scatter!(1:size(test, 1), unstandardize(test_label, train_unstandardized.MPG); label="true")
xaxis!(1:size(test, 1))
ylims!(10, 40)
title!("Full ADVI (Normal)")
preds = mapreduce(hcat, 1:5:size(chain, 1)) do i
    return unstandardize(prediction_chain(chain[i], test), train_unstandardized.MPG)
end

p3 = scatter(
    1:size(test, 1),
    mean(preds; dims=2);
    yerr=std(preds; dims=2),
    label="prediction (mean ± std)",
    size=(900, 500),
    markersize=8,
)
scatter!(1:size(test, 1), unstandardize(test_label, train_unstandardized.MPG); label="true")
xaxis!(1:size(test, 1))
ylims!(10, 40)
title!("MCMC (NUTS)")
plot(p1, p2, p3; layout=(1, 3), size=(900, 250), label="")

Here we actually see that indeed both the full ADVI and the MCMC approaches does a much better job of quantifying the uncertainty of predictions for never-before-seen samples, with full ADVI seemingly underestimating the variance slightly compared to MCMC.

So now you know how to do perform VI on your Turing.jl model! Great isn’t it?

Back to top