update to ForwardDiff 0.2.x

jeff-regier · Jul 26, 2016 · 1f03f15 · 1f03f15
1 parent 1c7033f
commit 1f03f15
Show file tree

Hide file tree

Showing 4 changed files with 39 additions and 50 deletions.
diff --git a/REQUIRE b/REQUIRE
@@ -3,7 +3,7 @@ Compat
 DataFrames
 Distributions
 DocOpt
-ForwardDiff 0.1.8 0.2
+ForwardDiff 0.2
 JLD
 Lumberjack
 Optim 0.5 0.5.1

diff --git a/src/elbo_kl.jl b/src/elbo_kl.jl
@@ -173,23 +173,16 @@ function subtract_kl!{NumType <: Number}(
 
     if calculate_derivs
         if calculate_hessian
-            hess, all_results =
-                ForwardDiff.hessian(subtract_kl_value_wrapper,
-                                    vp_vec,
-                                    ForwardDiff.AllResults)
-            accum.h += hess
-            accum.d += reshape(ForwardDiff.gradient(all_results), P, Sa)
-            accum.v[1] += ForwardDiff.value(all_results)
+            result = ForwardDiff.HessianResult(vp_vec)
+            ForwardDiff.hessian!(result, subtract_kl_value_wrapper, vp_vec)
+            accum.h += ForwardDiff.hessian(result)
         else
-            grad, all_results =
-                ForwardDiff.gradient(subtract_kl_value_wrapper,
-                                     vp_vec,
-                                     ForwardDiff.AllResults)
-            accum.d += reshape(grad, P, Sa)
-            accum.v[1] += ForwardDiff.value(all_results)
+            result = ForwardDiff.GradientResult(vp_vec)
+            ForwardDiff.gradient!(result, subtract_kl_value_wrapper, vp_vec)
         end
+        accum.d += reshape(ForwardDiff.gradient(result), P, Sa)
+        accum.v[1] += ForwardDiff.value(result)
     else
         accum.v[1] += subtract_kl_value_wrapper(vp_vec)
     end
 end
-
diff --git a/test/derivative_utils.jl b/test/derivative_utils.jl
@@ -1,39 +1,36 @@
 import Base.convert
 
-
-function convert(FDType::Type{ForwardDiff.GradientNumber},
-                 ea::ElboArgs{Float64})
-    x = ea.vp[1]
-    P = length(x)
-    FDType = ForwardDiff.GradientNumber{length(ea.vp[1]), Float64}
-
-    fd_x = [ ForwardDiff.GradientNumber(x[i], zeros(Float64, P)...) for i=1:P ]
-    convert(FDType, x[1])
-
-    vp_fd = convert(Array{Array{FDType, 1}, 1}, ea.vp[1])
-    ea_fd = ElboArgs(vp_fd)
-end
-
-function convert(FDType::Type{ForwardDiff.HessianNumber},
-                 ea::ElboArgs{Float64})
-    x = ea.vp[1]
-    P = length(x)
-    FDType = ForwardDiff.HessianNumber{length(ea.vp[1]), Float64}
-
-    fd_x = [ ForwardDiff.HessianNumber(x[i], zeros(Float64, P)...) for i=1:P ]
-    convert(FDType, x[1])
-
-    vp_fd = convert(Array{Array{FDType, 1}, 1}, ea.vp[1])
-    ea_fd = ElboArgs(vp_fd)
-end
+# function convert(FDType::Type{ForwardDiff.GradientNumber},
+#                  ea::ElboArgs{Float64})
+#     x = ea.vp[1]
+#     P = length(x)
+#     FDType = ForwardDiff.GradientNumber{length(ea.vp[1]), Float64}
+#
+#     fd_x = [ ForwardDiff.GradientNumber(x[i], zeros(Float64, P)...) for i=1:P ]
+#     convert(FDType, x[1])
+#
+#     vp_fd = convert(Array{Array{FDType, 1}, 1}, ea.vp[1])
+#     ea_fd = ElboArgs(vp_fd)
+# end
+#
+# function convert(FDType::Type{ForwardDiff.HessianNumber},
+#                  ea::ElboArgs{Float64})
+#     x = ea.vp[1]
+#     P = length(x)
+#     FDType = ForwardDiff.HessianNumber{length(ea.vp[1]), Float64}
+#
+#     fd_x = [ ForwardDiff.HessianNumber(x[i], zeros(Float64, P)...) for i=1:P ]
+#     convert(FDType, x[1])
+#
+#     vp_fd = convert(Array{Array{FDType, 1}, 1}, ea.vp[1])
+#     ea_fd = ElboArgs(vp_fd)
+# end
 
 
 # Maybe write it as a convert()?
-function forward_diff_model_params{T <: Number}(
-            FDType::Type{T},
-            ea0::ElboArgs{Float64})
+function forward_diff_model_params{T<:Number}(::Type{T}, ea0::ElboArgs{Float64})
     P = length(ea0.vp[1])
-    vp = Vector{FDType}[zeros(FDType, P) for s=1:ea0.S]
+    vp = Vector{T}[zeros(T, P) for s=1:ea0.S]
     # Set the values (but not gradient numbers) for parameters other
     # than the galaxy parameters.
     for s=1:ea0.S, i=1:length(ids)
@@ -46,4 +43,3 @@ function forward_diff_model_params{T <: Number}(
              ea0.patches,
              ea0.active_sources)
 end
-
diff --git a/test/test_derivatives.jl b/test/test_derivatives.jl
@@ -823,12 +823,12 @@ function test_galaxy_cache_component()
         f_wrap(par))
 
     # Check the gradient.
-    ad_grad_fun = ForwardDiff.gradient(f_wrap)
+    ad_grad_fun = x -> ForwardDiff.gradient(f_wrap, x)
     ad_grad = ad_grad_fun(par)
     bvn_derivs = elbo_vars.bvn_derivs
     @test_approx_eq ad_grad [bvn_derivs.bvn_u_d; bvn_derivs.bvn_s_d]
 
-    ad_hess_fun = ForwardDiff.hessian(f_wrap)
+    ad_hess_fun = x -> ForwardDiff.hessian(f_wrap, x)
     ad_hess = ad_hess_fun(par)
 
     @test_approx_eq ad_hess[1:2, 1:2] bvn_derivs.bvn_uu_h
@@ -871,11 +871,11 @@ function test_galaxy_sigma_derivs()
 
         gal_derivs = ElboDeriv.GalaxySigmaDerivs(e_angle, e_axis, e_scale, XiXi)
 
-        ad_grad_fun = ForwardDiff.gradient(f_wrap)
+        ad_grad_fun = x -> ForwardDiff.gradient(f_wrap, x)
         ad_grad = ad_grad_fun(par)
         @test_approx_eq gal_derivs.j[si, :][:] ad_grad
 
-        ad_hess_fun = ForwardDiff.hessian(f_wrap)
+        ad_hess_fun = x -> ForwardDiff.hessian(f_wrap, x)
         ad_hess = ad_hess_fun(par)
         @test_approx_eq(
             ad_hess,