Ready for the first release!!!

Author: YingboMa

Project.toml

@@ -1,6 +1,6 @@
 name = "ForwardDiff2"
 uuid = "994df76e-a4c1-5e1f-bd5c-23b9b5303d4f"
-authors = ["Yingbo Ma <mayingbo5@gmail.com>"]
+authors = ["Yingbo Ma <mayingbo5@gmail.com>", "Shashi Gowda <gowda@mit.edu>"]
 version = "0.1.0"
 
 [deps]
@@ -14,9 +14,15 @@ StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 
 [compat]
 julia = "1"
+Calculus = "0.5.1"
 Cassette = "0.3.0"
 ChainRules = "0.3.1"
 ChainRulesCore = "0.5.3"
+DiffRules = "1.0.0"
+MacroTools = "0.5.3"
+NaNMath = "0.3.3"
+SafeTestsets = "0.0.1"
+SpecialFunctions = "0.9"
 StaticArrays = "0.11, 0.12"
 
 [extras]

README.md

@@ -3,16 +3,60 @@
 [![Build Status](https://travis-ci.org/YingboMa/ForwardDiff2.jl.svg?branch=master)](https://travis-ci.org/YingboMa/ForwardDiff2.jl)
 [![codecov](https://codecov.io/gh/YingboMa/ForwardDiff2.jl/branch/master/graph/badge.svg)](https://codecov.io/gh/YingboMa/ForwardDiff2.jl)
 
-`ForwardDiff2` = `ForwardDiff.jl` + `ChainRules.jl` + Struct of arrays + `DualCache`
+`ForwardDiff2` = `ForwardDiff.jl` + `ChainRules.jl` + Struct of arrays
+
+### Warning!!!: This package is still work-in-progress
+
+User API:
+```julia
+julia> using ForwardDiff2: D
+
+julia> v = rand(2)
+2-element Array{Float64,1}:
+ 0.22260830987887537
+ 0.6397089507287486
+
+julia> D(prod)(v) # gradient
+1×2 LinearAlgebra.Adjoint{Float64,Array{Float64,1}}:
+ 0.639709  0.222608
+
+julia> D(cumsum)(v) # Jacobian
+2×2 Array{Float64,2}:
+ 1.0  0.0
+ 1.0  1.0
+
+julia> D(D(prod))(v) # Hessian
+2×2 LinearAlgebra.Adjoint{Float64,Array{Float64,2}}:
+ 0.0  1.0
+ 1.0  0.0
+```
+
+Note that `ForwardDiff2.jl` also works with `ModelingToolkit.jl`:
+```julia
+julia> using ModelingToolkit
+
+julia> @variables v[1:2]
+(Operation[v₁, v₂],)
+
+julia> D(prod)(v) # gradient
+1×2 LinearAlgebra.Adjoint{Operation,Array{Operation,1}}:
+ conj(1v₂ + v₁ * identity(0))  conj(identity(0) * v₂ + v₁ * 1)
+
+julia> D(cumsum)(v) # Jacobian
+2×2 Array{Expression,2}:
+     Constant(1)      identity(0)
+ identity(0) + 1  1 + identity(0)
+
+julia> D(D(prod))(v) # Hessian
+2×2 LinearAlgebra.Adjoint{Operation,Array{Operation,2}}:
+ conj((1 * identity(0) + v₁ * 0) + (1 * identity(0) + v₂ * 0))  conj((identity(0) * identity(0) + v₁ * 0) + (1 * 1 + v₂ * 0))
+ conj((1 * 1 + v₁ * 0) + (identity(0) * identity(0) + v₂ * 0))  conj((identity(0) * 1 + v₁ * 0) + (identity(0) * 1 + v₂ * 0))
+```
 
 Planned features:
 
 - works both on GPU and CPU
-- scalar forward mode AD
-- vectorized forward mode AD
 - [Dual cache](http://docs.juliadiffeq.org/latest/basics/faq.html#I-get-Dual-number-errors-when-I-solve-my-ODE-with-Rosenbrock-or-SDIRK-methods...?-1)
-- nested differentiation
-- hyper duals (?)
 - user-extensible scalar and tensor derivative definitions
 - in-place function
 - sparsity exploitation (color vector support)

test/api.jl

@@ -16,6 +16,7 @@ using StaticArrays
     @test D(D(x->x[1]^x[2] + x[3]^3 + x[3]*x[2]*x[1]))(@SVector[1,2,3]) === @SMatrix [2 4 2; 4 0 1; 2 1 18.]
     @test D(D(x->x[1]^x[2] + x[3]^3 + x[3]*x[2]*x[1]))([1,2,3]) == [2 4 2; 4 0 1; 2 1 18.]
     # inference
+    @inferred D(x->exp(x) + x^x + cos(x) + tan(x) + 2^x)(1)
     # broken due to `Core._apply`
     @test_broken @inferred D(x->exp(x) + x^x + cos(x) + tan(x) + 2^x + log(cos(x)) + sec(pi*x) - angle(x) + one(x) / log1p(sin(x)))(1)
 end

test/dualtest.jl

@@ -415,8 +415,8 @@ for N in (3), M in (4), V in (Int, Float32)
                     x, y = rand(1:10), rand()
                     dx = dualrun(()->$M.$f(Dual(x, one(x)), y))
                     dy = dualrun(()->$M.$f(x, Dual(y, one(y))))
-                    actualdx = $(derivs[1])
-                    actualdy = $(derivs[2])
+                    actualdx ≈ $(derivs[1])
+                    actualdy ≈ $(derivs[2])
                     @dtest value(dx) == $M.$f(x, y)
                     @dtest value(dy) == value(dx)
                     if isnan(actualdx)