[WIP] Use ReverseDiff/ForwardDiff in KL divergence code

REQUIRE

@@ -2,8 +2,10 @@ julia 0.5-
 DataFrames
 Distributions
 FITSIO 0.8.4
-ForwardDiff 0.2
+DiffBase 0.0.2
+ForwardDiff 0.3.3 0.4
+ReverseDiff 0.0.1 0.1
 JLD
-Optim 0.6
+Optim 0.7
 WCS 0.1.3
 StaticArrays

benchmark/speed/benchmark_elbo.jl

@@ -1,6 +1,7 @@
 #!/usr/bin/env julia
 
 import Celeste.DeterministicVI: elbo, ElboArgs
+using DiffBase
 
 include(string(Pkg.dir("Celeste"), "/test/Synthetic.jl"))
 include(string(Pkg.dir("Celeste"), "/test/SampleData.jl"))
@@ -17,19 +18,21 @@ function main()
                   calculate_gradient=true,
                   calculate_hessian=false)
 
+    param_length = length(Celeste.Model.CanonicalParams)
+    kl_source = Celeste.SensitiveFloats.SensitiveFloat{Float64}(param_length, 1, true, false)
 
     println("Warm-up / compiling.")
     # do a trial run first, so we don't profile/time compling the code
-    elbo(ea)
+    elbo(ea, kl_source)
     Profile.clear_malloc_data()
 
     println("Calculating ELBO and gradient.")
     if isempty(ARGS)
         # let's time it without any overhead from profiling
-        @time elbo(ea)
+        @time elbo(ea, kl_source)
     elseif ARGS[1] == "--profile"
         Profile.init(delay=0.001)
-        @profile elbo(ea)
+        @profile elbo(ea, kl_source)
         Profile.print(format=:flat, sortedby=:count)
     end
 end

src/DeterministicVI.jl

@@ -3,6 +3,7 @@ Calculate value, gradient, and hessian of the variational ELBO.
 """
 module DeterministicVI
 
+using Base.Threads: threadid, nthreads
 using ..Model
 import ..Model: BivariateNormalDerivatives, BvnComponent, GalaxyCacheComponent,
                 GalaxySigmaDerivs, SkyPatch,
@@ -16,6 +17,8 @@ using ..Transform
 import DataFrames
 import Optim
 using StaticArrays
+using ForwardDiff
+using ReverseDiff
 
 export ElboArgs
 

src/DeterministicVIImagePSF.jl

@@ -5,14 +5,14 @@ image rather than a mixture of Gaussians.
 
 module DeterministicVIImagePSF
 
-using StaticArrays
+using StaticArrays, DiffBase
 
 import ..DeterministicVI:
     ElboArgs, ElboIntermediateVariables,
     StarPosParams, GalaxyPosParams, CanonicalParams, VariationalParams,
     SourceBrightness, GalaxyComponent, SkyPatch,
     load_source_brightnesses, add_elbo_log_term!,
-    accumulate_source_pixel_brightness!, subtract_kl!
+    accumulate_source_pixel_brightness!, subtract_kl_all_sources!, KL_HELPER_POOL
 
 import ..Model:
     populate_gal_fsm!, getids, ParamSet, linear_world_to_pix, lidx,
@@ -23,7 +23,7 @@ import ..Model:
 import ..SensitiveFloats:
     SensitiveFloat, zero_sensitive_float_array,
     multiply_sfs!, add_scaled_sfs!, clear!
-    
+
 import ..Infer: load_active_pixels!
 
 import ..PSF: get_psf_at_point

src/SensitiveFloats.jl

@@ -208,11 +208,12 @@ function add_sources_sf!{NumType <: Number}(
     @assert size(sf_all.d, 1) == size(sf_s.d, 1)
 
     P = sf_all.local_P
+    P_shifted = P * (s - 1)
 
     if sf_all.has_gradient
         @assert size(sf_s.d) == (P, 1)
-        @inbounds for s_ind1 in 1:sf_all.local_P
-            s_all_ind1 = P * (s - 1) + s_ind1
+        @inbounds for s_ind1 in 1:P
+            s_all_ind1 = P_shifted + s_ind1
             sf_all.d[s_all_ind1] = sf_all.d[s_all_ind1] + sf_s.d[s_ind1]
         end
     end
@@ -224,12 +225,11 @@ function add_sources_sf!{NumType <: Number}(
         @assert Ph >= P * s
 
         @inbounds for s_ind1 in 1:P
-            s_all_ind1 = P * (s - 1) + s_ind1
+            s_all_ind1 = P_shifted + s_ind1
 
             @inbounds for s_ind2 in 1:s_ind1
-                s_all_ind2 = P * (s - 1) + s_ind2
-                sf_all.h[s_all_ind2, s_all_ind1] =
-                    sf_all.h[s_all_ind2, s_all_ind1] + sf_s.h[s_ind2, s_ind1]
+                s_all_ind2 = P_shifted + s_ind2
+                sf_all.h[s_all_ind2, s_all_ind1] += sf_s.h[s_ind2, s_ind1]
                 # TODO: move outside the loop?
                 sf_all.h[s_all_ind1, s_all_ind2] = sf_all.h[s_all_ind2, s_all_ind1]
             end