add Tapir support

.github/workflows/CI.yml

@@ -19,8 +19,8 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1'
-          - '1.6'
+          - '1.7'
+          - '1.10'
         os:
           - ubuntu-latest
           - macOS-latest

Project.toml

@@ -25,13 +25,15 @@ Bijectors = "76274a88-744f-5084-9051-94815aaf08c4"
 Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
+Tapir = "07d77754-e150-4737-8c94-cd238a1fb45b"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [extensions]
 AdvancedVIBijectorsExt = "Bijectors"
 AdvancedVIEnzymeExt = "Enzyme"
 AdvancedVIForwardDiffExt = "ForwardDiff"
 AdvancedVIReverseDiffExt = "ReverseDiff"
+AdvancedVITapirExt = "Tapir"
 AdvancedVIZygoteExt = "Zygote"
 
 [compat]
@@ -55,15 +57,17 @@ Requires = "1.0"
 ReverseDiff = "1.15.1"
 SimpleUnPack = "1.1.0"
 StatsBase = "0.32, 0.33, 0.34"
+Tapir = "0.2.23"
 Zygote = "0.6.63"
-julia = "1.6"
+julia = "1.7"
 
 [extras]
 Bijectors = "76274a88-744f-5084-9051-94815aaf08c4"
 Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
+Tapir = "07d77754-e150-4737-8c94-cd238a1fb45b"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 

ext/AdvancedVIForwardDiffExt.jl

@@ -14,10 +14,11 @@ end
 getchunksize(::ADTypes.AutoForwardDiff{chunksize}) where {chunksize} = chunksize
 
 function AdvancedVI.value_and_gradient!(
-    ad ::ADTypes.AutoForwardDiff,
+    ad   ::ADTypes.AutoForwardDiff,
+         ::Any,
     f,
-    x  ::AbstractVector{<:Real},
-    out::DiffResults.MutableDiffResult
+    x    ::AbstractVector,
+    out  ::DiffResults.MutableDiffResult
 )
     chunk_size = getchunksize(ad)
     config = if isnothing(chunk_size)
@@ -30,13 +31,14 @@ function AdvancedVI.value_and_gradient!(
 end
 
 function AdvancedVI.value_and_gradient!(
-    ad ::ADTypes.AutoForwardDiff,
+    ad    ::ADTypes.AutoForwardDiff,
+    st_ad,
     f,
-    x  ::AbstractVector,
+    x     ::AbstractVector,
     aux, 
-    out::DiffResults.MutableDiffResult
+    out   ::DiffResults.MutableDiffResult
 )
-    AdvancedVI.value_and_gradient!(ad, x′ -> f(x′, aux), x, out)
+    AdvancedVI.value_and_gradient!(ad, st_ad, x′ -> f(x′, aux), x, out)
 end
 
 end

ext/AdvancedVIReverseDiffExt.jl

@@ -13,9 +13,10 @@ end
 
 # ReverseDiff without compiled tape
 function AdvancedVI.value_and_gradient!(
-    ad::ADTypes.AutoReverseDiff,
+       ::ADTypes.AutoReverseDiff,
+       ::Any,
     f,
-    x::AbstractVector{<:Real},
+    x  ::AbstractVector{<:Real},
     out::DiffResults.MutableDiffResult
 )
     tp = ReverseDiff.GradientTape(f, x)
@@ -24,13 +25,14 @@ function AdvancedVI.value_and_gradient!(
 end
 
 function AdvancedVI.value_and_gradient!(
-    ad::ADTypes.AutoReverseDiff,
+    ad    ::ADTypes.AutoReverseDiff,
+    st_ad,
     f,
-    x::AbstractVector{<:Real},
+    x     ::AbstractVector{<:Real},
     aux,
-    out::DiffResults.MutableDiffResult
+    out   ::DiffResults.MutableDiffResult
 )
-    AdvancedVI.value_and_gradient!(ad, x′ -> f(x′, aux), x, out)
+    AdvancedVI.value_and_gradient!(ad, st_ad, x′ -> f(x′, aux), x, out)
 end
 
 end

ext/AdvancedVITapirExt.jl

@@ -0,0 +1,47 @@
+
+module AdvancedVITapirExt
+
+if isdefined(Base, :get_extension)
+    using AdvancedVI
+    using AdvancedVI: ADTypes, DiffResults
+    using Tapir
+else
+    using ..AdvancedVI
+    using ..AdvancedVI: ADTypes, DiffResults
+    using ..Tapir
+end
+
+AdvancedVI.init_adbackend(::ADTypes.AutoTapir, f, x) = Tapir.build_rrule(f, x)
+
+function AdvancedVI.value_and_gradient!(
+          ::ADTypes.AutoTapir,
+    st_ad,
+    f,
+    x     ::AbstractVector{<:Real},
+    out   ::DiffResults.MutableDiffResult
+)
+    rule = st_ad
+    y, g = Tapir.value_and_gradient!!(rule, f, x)
+    DiffResults.value!(out, y)
+    DiffResults.gradient!(out, last(g))
+    return out
+end
+
+AdvancedVI.init_adbackend(::ADTypes.AutoTapir, f, x, aux) = Tapir.build_rrule(f, x, aux)
+
+function AdvancedVI.value_and_gradient!(
+          ::ADTypes.AutoTapir,
+    st_ad,
+    f,
+    x     ::AbstractVector{<:Real},
+    aux,
+    out   ::DiffResults.MutableDiffResult
+)
+    rule = st_ad
+    y, g = Tapir.value_and_gradient!!(rule, f, x, aux)
+    DiffResults.value!(out, y)
+    DiffResults.gradient!(out, last(g))
+    return out
+end
+
+end

ext/AdvancedVIZygoteExt.jl

@@ -14,9 +14,10 @@ else
 end
 
 function AdvancedVI.value_and_gradient!(
-    ::ADTypes.AutoZygote,
+       ::ADTypes.AutoZygote,
+       ::Any,
     f,
-    x::AbstractVector{<:Real},
+    x  ::AbstractVector{<:Real},
     out::DiffResults.MutableDiffResult
 )
     y, back = Zygote.pullback(f, x)
@@ -27,13 +28,14 @@ function AdvancedVI.value_and_gradient!(
 end
 
 function AdvancedVI.value_and_gradient!(
-    ad::ADTypes.AutoZygote,
+    ad    ::ADTypes.AutoZygote,
+    st_ad,
     f,
-    x::AbstractVector{<:Real},
+    x     ::AbstractVector{<:Real},
     aux,
-    out::DiffResults.MutableDiffResult
+    out   ::DiffResults.MutableDiffResult
 )
-    AdvancedVI.value_and_gradient!(ad, x′ -> f(x′, aux), x, out)
+    AdvancedVI.value_and_gradient!(ad, st_ad, x′ -> f(x′, aux), x, out)
 end
 
 end

src/AdvancedVI.jl

@@ -25,14 +25,15 @@ using StatsBase
 
 # derivatives
 """
-    value_and_gradient!(ad, f, x, out)
-    value_and_gradient!(ad, f, x, aux, out)
+    value_and_gradient!(adtype, ad_st, f, x, out)
+    value_and_gradient!(adtype, ad_st, f, x, aux, out)
 
-Evaluate the value and gradient of a function `f` at `x` using the automatic differentiation backend `ad` and store the result in `out`.
+Evaluate the value and gradient of a function `f` at `x` using the automatic differentiation (AD) backend `ad` and store the result in `out`.
 `f` may receive auxiliary input as `f(x,aux)`.
 
 # Arguments
-- `ad::ADTypes.AbstractADType`: Automatic differentiation backend. 
+- `adtype::ADTypes.AbstractADType`: AD backend. 
+- `ad_st`: State used by the AD backend. (This will often be pre-compiled tapes/caches.)
 - `f`: Function subject to differentiation.
 - `x`: The point to evaluate the gradient.
 - `aux`: Auxiliary input passed to `f`.
@@ -41,18 +42,22 @@ Evaluate the value and gradient of a function `f` at `x` using the automatic dif
 function value_and_gradient! end
 
 """
-    stop_gradient(x)
+    init_adbackend(adtype, f, x)
+    init_adbackend(adtype, f, x, aux)
 
-Stop the gradient from propagating to `x` if the selected ad backend supports it.
-Otherwise, it is equivalent to `identity`.
+Initialize the AD backend and setup states necessary.
 
 # Arguments
-- `x`: Input
+- `ad::ADTypes.AbstractADType`: Automatic differentiation backend. 
+- `f`: Function subject to differentiation.
+- `x`: The point to evaluate the gradient.
+- `aux`: Auxiliary input passed to `f`.
 
 # Returns
-- `x`: Same value as the input.
+- `ad_st`: State of the AD backend. (This will often be pre-compiled tapes/caches.)
 """
-function stop_gradient end
+init_adbackend(::ADTypes.AbstractADType, ::Any, ::Any)        = nothing
+init_adbackend(::ADTypes.AbstractADType, ::Any, ::Any, ::Any) = nothing
 
 # Update for gradient descent step
 """
@@ -95,14 +100,17 @@ If the estimator is stateful, it can implement `init` to initialize the state.
 abstract type AbstractVariationalObjective end
 
 """
-    init(rng, obj, prob, params, restructure)
+    init(rng, obj, adtype, prob, params, restructure)
 
-Initialize a state of the variational objective `obj` given the initial variational parameters `λ`.
+Initialize a state of the variational objective `obj`.
 This function needs to be implemented only if `obj` is stateful.
+The state of the AD backend `adtype` shall also be initialized here.
 
 # Arguments
 - `rng::Random.AbstractRNG`: Random number generator.
 - `obj::AbstractVariationalObjective`: Variational objective.
+- `adtype::ADTypes.ADType`:Automatic differentiation backend.
+- `prob`: The target log-joint likelihood implementing the `LogDensityProblem` interface.
 - `params`: Initial variational parameters.
 - `restructure`: Function that reconstructs the variational approximation from `λ`.
 """
@@ -112,6 +120,7 @@ init(
     ::Any,
     ::Any,
     ::Any,
+    ::Any,
 ) = nothing
 
 """

src/objectives/elbo/repgradelbo.jl

@@ -106,6 +106,20 @@ function estimate_repgradelbo_ad_forward(params′, aux)
     -elbo
 end
 
+function init(
+    rng         ::Random.AbstractRNG,
+    obj         ::RepGradELBO,
+    adtype      ::ADTypes.AbstractADType,
+    prob,
+    params,
+    restructure,
+)
+    q_stop = restructure(params)
+    aux = (rng=rng, obj=obj, problem=prob, restructure=restructure, q_stop=q_stop)
+    ad_st = init_adbackend(adtype, estimate_repgradelbo_ad_forward, params, aux)
+    (ad_st=ad_st,)
+end
+
 function estimate_gradient!(
     rng   ::Random.AbstractRNG,
     obj   ::RepGradELBO,
@@ -117,11 +131,12 @@ function estimate_gradient!(
     state,
 )
     q_stop = restructure(params)
+    ad_st  = state.ad_st
     aux = (rng=rng, obj=obj, problem=prob, restructure=restructure, q_stop=q_stop)
     value_and_gradient!(
-        adtype, estimate_repgradelbo_ad_forward, params, aux, out
+        adtype, ad_st, estimate_repgradelbo_ad_forward, params, aux, out
     )
     nelbo = DiffResults.value(out)
     stat  = (elbo=-nelbo,)
-    out, nothing, stat
+    out, state, stat
 end

src/optimize.jl

@@ -66,7 +66,9 @@ function optimize(
 )
     params, restructure = Optimisers.destructure(deepcopy(q_init))
     opt_st   = maybe_init_optimizer(state_init, optimizer, params)
-    obj_st   = maybe_init_objective(state_init, rng, objective, problem, params, restructure)
+    obj_st   = maybe_init_objective(
+        state_init, rng, adtype, objective, problem, params, restructure
+    )
     grad_buf = DiffResults.DiffResult(zero(eltype(params)), similar(params))
     stats    = NamedTuple[]
 

src/utils.jl

@@ -18,6 +18,7 @@ end
 function maybe_init_objective(
     state_init::NamedTuple,
     rng       ::Random.AbstractRNG,
+    adtype    ::ADTypes.AbstractADType,
     objective ::AbstractVariationalObjective,
     problem,
     params,
@@ -26,7 +27,7 @@ function maybe_init_objective(
     if haskey(state_init, :objective)
         state_init.objective
     else
-        init(rng, objective, problem, params, restructure)
+        init(rng, objective, adtype, problem, params, restructure)
     end
 end
 

test/Project.toml

@@ -18,6 +18,7 @@ ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
 SimpleUnPack = "ce78b400-467f-4804-87d8-8f486da07d0a"
 StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
+Tapir = "07d77754-e150-4737-8c94-cd238a1fb45b"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
@@ -40,7 +41,8 @@ ReverseDiff = "1.15.1"
 SimpleUnPack = "1.1.0"
 StableRNGs = "1.0.0"
 Statistics = "1"
+Tapir = "0.2.23"
 Test = "1"
 Tracker = "0.2.20"
 Zygote = "0.6.63"
-julia = "1.6"
+julia = "1.7"

test/inference/repgradelbo_distributionsad.jl

@@ -14,6 +14,7 @@
             :ForwarDiff  => AutoForwardDiff(),
             #:ReverseDiff => AutoReverseDiff(),
             :Zygote      => AutoZygote(), 
+            :Tapir       => AutoTapir(), 
             #:Enzyme      => AutoEnzyme(),
         )
 

test/inference/repgradelbo_locationscale.jl

@@ -14,6 +14,7 @@
         (adbackname, adtype) in Dict(
             :ForwarDiff  => AutoForwardDiff(),
             :ReverseDiff => AutoReverseDiff(),
+            :Tapir       => AutoTapir(safe_mode=false), 
             :Zygote      => AutoZygote(), 
             #:Enzyme      => AutoEnzyme(),
         )

test/inference/repgradelbo_locationscale_bijectors.jl

@@ -13,7 +13,8 @@
         (adbackname, adtype) in Dict(
             :ForwarDiff  => AutoForwardDiff(),
             :ReverseDiff => AutoReverseDiff(),
-            #:Zygote      => AutoZygote(), 
+            :Zygote      => AutoZygote(), 
+            :Tapir       => AutoTapir(safe_mode=false), 
             #:Enzyme      => AutoEnzyme(),
         )