Remove dependency DistributionsAD

Project.toml

@@ -6,14 +6,14 @@ version = "0.2.3"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
-DistributionsAD = "ced4e74d-a319-5a8a-b0ac-84af2272839c"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 IterativeSolvers = "42fd0dbc-a981-5370-80f2-aaf504508153"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 LinearMaps = "7a12625a-238d-50fd-b39a-03d52299707e"
 Optim = "429524aa-4258-5aef-a3af-852621145aeb"
 PositiveFactorizations = "85a6dd25-e78a-55b7-8502-1745935b8125"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 ValueShapes = "136a8f8c-c49b-4edb-8b98-f3d64d48be8f"
@@ -22,7 +22,6 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 [compat]
 ChainRulesCore = "0.9.44"
 Distributions = "0.23, 0.24, 0.25"
-DistributionsAD = "0.6"
 ForwardDiff = "0.10"
 IterativeSolvers = "0.8, 0.9"
 LinearMaps = "3"

README.md

@@ -8,9 +8,11 @@
 
 *This is an implementation of the [Metric Gaussian Variational Inference](https://arxiv.org/abs/1901.11033) (MGVI) algorithm in julia*
 
-
 MGVI is an iterative method that performs a series of Gaussian approximations to the posterior. It alternates between approximating the covariance with the inverse Fisher information metric evaluated at an intermediate mean estimate and optimizing the KL-divergence for the given covariance with respect to the mean. This procedure is iterated until the uncertainty estimate is self-consistent with the mean parameter. We achieve linear scaling by avoiding to store the covariance explicitly at any time. Instead we draw samples from the approximating distribution relying on an implicit representation and numerical schemes to approximately solve linear equations. Those samples are used to approximate the KL-divergence and its gradient. The usage of natural gradient descent allows for rapid convergence. Formulating the Bayesian model in standardized coordinates makes MGVI applicable to any inference problem with continuous parameters.
 
+Depending on the distributions used in your application, you may need to use the package [DistributionsAD](https://github.com/TuringLang/DistributionsAD.jl).
+
+
 ## Documentation
 * [Documentation for stable version](https://bat.github.io/MGVI.jl/stable)
 * [Documentation for development version](https://bat.github.io/MGVI.jl/dev)

docs/src/index.md

@@ -2,6 +2,8 @@
 
 MGVI is an iterative method that performs a series of Gaussian approximations to the posterior. We alternate between approximating the covariance with the inverse Fisher information metric evaluated at an intermediate mean estimate and optimizing the KL-divergence for the given covariance with respect to the mean. This procedure is iterated until the uncertainty estimate is self-consistent with the mean parameter. We achieve linear scaling by avoiding to store the covariance explicitly at any time. Instead we draw samples from the approximating distribution relying on an implicit representation and numerical schemes to approximately solve linear equations. Those samples are used to approximate the KL-divergence and its gradient. The usage of natural gradient descent allows for rapid convergence. Formulating the Bayesian model in standardized coordinates makes MGVI applicable to any inference problem with continuous parameters. 
 
+Depending on the distributions used in your application, you may need to use the package [DistributionsAD](https://github.com/TuringLang/DistributionsAD.jl).
+
 
 ## Citing MGVI.jl
 

src/MGVI.jl

@@ -16,23 +16,29 @@ using Random
 using SparseArrays
 using Base.Iterators
 using Distributions
-using DistributionsAD
 import ForwardDiff
 using LinearMaps
 using IterativeSolvers
 using Optim
 using PositiveFactorizations
+using Requires
 import SparseArrays: blockdiag
 using SparseArrays
 using StaticArrays
 using ValueShapes
 import Zygote
 
+import Requires
+
 include("custom_linear_maps.jl")
 include("shapes.jl")
 include("jacobian_maps.jl")
 include("information.jl")
 include("residual_samplers.jl")
 include("mgvi_impl.jl")
 
+function __init__()
+    @require DistributionsAD = "ced4e74d-a319-5a8a-b0ac-84af2272839c" include("distributionsad_support.jl")
+end
+
 end # module

src/distributionsad_support.jl

@@ -0,0 +1,8 @@
+# This file is a part of MGVI.jl, licensed under the MIT License (MIT).
+
+
+function unshaped_params(d::DistributionsAD.TuringDenseMvNormal)
+    μ = d.m
+    σ = convert(AbstractMatrix, d.C)
+    vcat(μ, _uppertriang_to_vec(σ))
+end

src/shapes.jl

@@ -52,9 +52,3 @@ function unshaped_params(d::MvNormal)
     μ, σ = params(d)
     vcat(μ, _uppertriang_to_vec(σ))
 end
-
-function unshaped_params(d::TuringDenseMvNormal)
-    μ = d.m
-    σ = d.C.L*d.C.U
-    vcat(μ, _uppertriang_to_vec(σ))
-end

test/Project.toml

@@ -1,7 +1,6 @@
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
-DistributionsAD = "ced4e74d-a319-5a8a-b0ac-84af2272839c"
 FFTW = "7a1cc6ca-52ef-59f5-83cd-3a7055c09341"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 HypothesisTests = "09f84164-cd44-5f33-b23f-e6b0d136a0d5"

test/information/information_utils.jl

@@ -1,7 +1,6 @@
 # This file is a part of MGVI.jl, licensed under the MIT License (MIT).
 
 using Distributions
-using DistributionsAD
 using LinearAlgebra
 import Zygote