add indirection for update step, add projection for LocationScale

ext/AdvancedVIBijectorsExt.jl

@@ -4,13 +4,37 @@ module AdvancedVIBijectorsExt
 if isdefined(Base, :get_extension)
     using AdvancedVI
     using Bijectors
+    using LinearAlgebra
+    using Optimisers
     using Random
 else
     using ..AdvancedVI
     using ..Bijectors
+    using ..LinearAlgebra
+    using ..Optimisers
     using ..Random
 end
 
+function AdvancedVI.update_variational_params!(
+    ::Type{<:Bijectors.TransformedDistribution{<:AdvancedVI.MvLocationScale}},
+    opt_st,
+    params,
+    restructure,
+    grad
+)
+    opt_st, params = Optimisers.update!(opt_st, params, grad)
+    q = restructure(params)
+    ϵ = q.dist.scale_eps
+
+    # Project the scale matrix to the set of positive definite triangular matrices
+    diag_idx = diagind(q.dist.scale)
+    @. q.dist.scale[diag_idx] = max(q.dist.scale[diag_idx], ϵ)
+
+    params, _ = Optimisers.destructure(q)
+
+    opt_st, params
+end
+
 function AdvancedVI.reparam_with_entropy(
     rng      ::Random.AbstractRNG,
     q        ::Bijectors.TransformedDistribution,

src/AdvancedVI.jl

@@ -37,6 +37,33 @@ Evaluate the value and gradient of a function `f` at `θ` using the automatic di
 """
 function value_and_gradient! end
 
+# Update for gradient descent step
+"""
+    update_variational_params!(family_type, opt_st, params, restructure, grad)
+
+Update variational distribution according to the update rule in the optimizer state `opt_st` and the variational family `family_type`.
+
+This is a wrapper around `Optimisers.update!` to provide some indirection.
+For example, depending on the optimizer and the variational family, this may do additional things such as applying projection or proximal mappings.
+Same as the default behavior of `Optimisers.update!`, `params` and `opt_st` may be updated by the routine and are no longer valid after calling this functino.
+Instead, the return values should be used.
+
+# Arguments
+- `family_type::Type`: Type of the variational family `typeof(restructure(params))`.
+- `opt_st`: Optimizer state returned by `Optimisers.setup`.
+- `params`: Current set of parameters to be updated.
+- `restructure`: Callable for restructuring the varitional distribution from `params`.
+- `grad`: Gradient to be used by the update rule of `opt_st`.
+
+# Returns
+- `opt_st`: Updated optimizer state.
+- `params`: Updated parameters.
+"""
+function update_variational_params! end
+
+update_variational_params!(::Type, opt_st, params, restructure, grad) =
+    Optimisers.update!(opt_st, params, grad)
+
 # estimators
 """
     AbstractVariationalObjective

src/families/location_scale.jl

@@ -14,11 +14,21 @@ represented as follows:
 ```
 """
 struct MvLocationScale{
-    S, D <: ContinuousDistribution, L
+    S, D <: ContinuousDistribution, L, E <: Real
 } <: ContinuousMultivariateDistribution
-    location::L
-    scale   ::S
-    dist    ::D
+    location ::L
+    scale    ::S
+    dist     ::D
+    scale_eps::E
+end
+
+function MvLocationScale(
+    location ::AbstractVector{T},
+    scale    ::AbstractMatrix{T},
+    dist     ::ContinuousDistribution;
+    scale_eps::T = sqrt(eps(T))
+) where {T <: Real}
+    MvLocationScale(location, scale, dist, scale_eps)
 end
 
 Functors.@functor MvLocationScale (location, scale)
@@ -36,14 +46,14 @@ function (re::RestructureMeanField)(flat::AbstractVector)
     n_dims   = div(length(flat), 2)
     location = first(flat, n_dims)
     scale    = Diagonal(last(flat, n_dims))
-    MvLocationScale(location, scale, re.q.dist)
+    MvLocationScale(location, scale, re.q.dist, re.q.scale_eps)
 end
 
 function Optimisers.destructure(
     q::MvLocationScale{<:Diagonal, D, L}
 ) where {D, L}
     @unpack location, scale, dist = q
-    flat   = vcat(location, diag(scale))
+    flat = vcat(location, diag(scale))
     flat, RestructureMeanField(q)
 end
 # end
@@ -57,17 +67,17 @@ Base.eltype(::Type{<:MvLocationScale{S, D, L}}) where {S, D, L} = eltype(D)
 function StatsBase.entropy(q::MvLocationScale)
     @unpack  location, scale, dist = q
     n_dims = length(location)
-    n_dims*convert(eltype(location), entropy(dist)) + first(logdet(scale))
+    n_dims*convert(eltype(location), entropy(dist)) + logdet(scale)
 end
 
 function Distributions.logpdf(q::MvLocationScale, z::AbstractVector{<:Real})
     @unpack location, scale, dist = q
-    sum(Base.Fix1(logpdf, dist), scale \ (z - location)) - first(logdet(scale))
+    sum(Base.Fix1(logpdf, dist), scale \ (z - location)) - logdet(scale)
 end
 
 function Distributions._logpdf(q::MvLocationScale, z::AbstractVector{<:Real})
     @unpack location, scale, dist = q
-    sum(Base.Fix1(logpdf, dist), scale \ (z - location)) - first(logdet(scale))
+    sum(Base.Fix1(logpdf, dist), scale \ (z - location)) - logdet(scale)
 end
 
 function Distributions.rand(q::MvLocationScale)
@@ -128,14 +138,11 @@ Construct a Gaussian variational approximation with a dense covariance matrix.
 function FullRankGaussian(
     μ::AbstractVector{T},
     L::LinearAlgebra.AbstractTriangular{T};
-    check_args::Bool = true
+    scale_eps::T = sqrt(eps(T))
 ) where {T <: Real}
-    @assert minimum(diag(L)) > eps(eltype(L)) "Scale must be positive definite"
-    if check_args && (minimum(diag(L)) < sqrt(eps(eltype(L))))
-        @warn "Initial scale is too small (minimum eigenvalue is $(minimum(diag(L)))). This might result in unstable optimization behavior."
-    end
+    @assert minimum(diag(L)) ≥ sqrt(scale_eps) "Initial scale is too small (smallest diagonal value is $(minimum(diag(L)))). This might result in unstable optimization behavior."
     q_base = Normal{T}(zero(T), one(T))
-    MvLocationScale(μ, L, q_base)
+    MvLocationScale(μ, L, q_base, scale_eps)
 end
 
 """
@@ -153,12 +160,25 @@ Construct a Gaussian variational approximation with a diagonal covariance matrix
 function MeanFieldGaussian(
     μ::AbstractVector{T},
     L::Diagonal{T};
-    check_args::Bool = true
+    scale_eps::T = sqrt(eps(T)),
 ) where {T <: Real}
-    @assert minimum(diag(L)) > eps(eltype(L)) "Scale must be a Cholesky factor"
-    if check_args && (minimum(diag(L)) < sqrt(eps(eltype(L))))
-        @warn "Initial scale is too small (minimum eigenvalue is $(minimum(diag(L)))). This might result in unstable optimization behavior."
-    end
+    @assert minimum(diag(L)) ≥ sqrt(eps(eltype(L))) "Initial scale is too small (smallest diagonal value is $(minimum(diag(L)))). This might result in unstable optimization behavior."
     q_base = Normal{T}(zero(T), one(T))
-    MvLocationScale(μ, L, q_base)
+    MvLocationScale(μ, L, q_base, scale_eps)
+end
+
+function update_variational_params!(
+    ::Type{<:MvLocationScale}, opt_st, params, restructure, grad
+)
+    opt_st, params = Optimisers.update!(opt_st, params, grad)
+    q = restructure(params)
+    ϵ = q.scale_eps
+
+    # Project the scale matrix to the set of positive definite triangular matrices
+    diag_idx = diagind(q.scale)
+    @. q.scale[diag_idx] = max(q.scale[diag_idx], ϵ)
+
+    params, _ = Optimisers.destructure(q)
+
+    opt_st, params
 end

src/optimize.jl

@@ -80,7 +80,9 @@ function optimize(
         stat = merge(stat, stat′)
 
         grad = DiffResults.gradient(grad_buf)
-        opt_st, params = Optimisers.update!(opt_st, params, grad)
+        opt_st, params = update_variational_params!(
+            typeof(q_init), opt_st, params, restructure, grad
+        )
 
         if !isnothing(callback)
             stat′ = callback(

test/interface/location_scale.jl

@@ -138,3 +138,35 @@
         @test q         == re(λ)
     end
 end
+
+@testset "scale positive definite projection" begin
+    @testset "$(string(covtype)) $(realtype) $(bijector)" for
+        covtype     = [:meanfield, :fullrank],
+        realtype    = [Float32,     Float64],
+        bijector    = [nothing,    :identity]
+
+        d = 5
+        μ = zeros(realtype, d)
+        ϵ = sqrt(realtype(0.5))
+        q = if covtype == :fullrank
+            L = LowerTriangular(Matrix{realtype}(I,d,d))
+            FullRankGaussian(μ, L; scale_eps=ϵ)
+        elseif covtype == :meanfield
+            L = Diagonal(ones(realtype, d))
+            MeanFieldGaussian(μ, L; scale_eps=ϵ)
+        end
+        q_trans = if isnothing(bijector) 
+            q
+        else
+            Bijectors.TransformedDistribution(q, identity)
+        end
+        g = deepcopy(q)
+
+        λ, re   = Optimisers.destructure(q)
+        grad, _ = Optimisers.destructure(g)
+        opt_st  = Optimisers.setup(Descent(one(realtype)), λ)
+        _, λ′    = AdvancedVI.update_variational_params!(typeof(q), opt_st, λ, re, grad)
+        q′       = re(λ′)
+        @test all(diag(var(q′)) .≥ ϵ^2)
+    end
+end