More methods for round and rand, fix NNlib

CHANGELOG.md

@@ -1,5 +1,11 @@
 # SparseConnectivityTracer.jl
 
+## Version `v0.6.2`
+
+* ![Feature][badge-feature] More methods for `round` and `rand` ([#162])
+* ![Bugfix][badge-bugfix] Fix Hessian on NNlib activation functions ([#162])
+* ![Bugfix][badge-bugfix] Fix `isless` ([#161])
+
 ## Version `v0.6.1`
 
 * ![Enhancement][badge-enhancement] Improve the performance of Hessian pattern tracing by an order of magnitude:
@@ -33,6 +39,8 @@
 [badge-maintenance]: https://img.shields.io/badge/maintenance-gray.svg
 [badge-docs]: https://img.shields.io/badge/docs-orange.svg
 
+[#162]: https://github.com/adrhill/SparseConnectivityTracer.jl/pull/162
+[#161]: https://github.com/adrhill/SparseConnectivityTracer.jl/pull/161
 [#158]: https://github.com/adrhill/SparseConnectivityTracer.jl/pull/158
 [#155]: https://github.com/adrhill/SparseConnectivityTracer.jl/pull/155
 [#151]: https://github.com/adrhill/SparseConnectivityTracer.jl/pull/151

Project.toml

@@ -1,7 +1,7 @@
 name = "SparseConnectivityTracer"
 uuid = "9f842d2f-2579-4b1d-911e-f412cf18a3f5"
 authors = ["Adrian Hill <[email protected]>"]
-version = "0.6.1"
+version = "0.6.2-DEV"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"

ext/SparseConnectivityTracerNNlibExt.jl

@@ -43,12 +43,12 @@ for op in ops_1_to_1_s
     @eval SCT.is_der2_zero_global(::$T) = false
 end
 
-SCT.is_der2_zero_local(::typeof(celu)) = x > 0
-SCT.is_der2_zero_local(::typeof(elu)) = x > 0
-SCT.is_der2_zero_local(::typeof(selu)) = x > 0
+SCT.is_der2_zero_local(::typeof(celu), x) = x > 0
+SCT.is_der2_zero_local(::typeof(elu), x) = x > 0
+SCT.is_der2_zero_local(::typeof(selu), x) = x > 0
 
-SCT.is_der1_zero_local(::typeof(hardswish)) = x < -3
-SCT.is_der2_zero_local(::typeof(hardswish)) = x < -3 || x > 3
+SCT.is_der1_zero_local(::typeof(hardswish), x) = x < -3
+SCT.is_der2_zero_local(::typeof(hardswish), x) = x < -3 || x > 3
 
 # ops_1_to_1_f:
 # x -> f  != 0

ext/SparseConnectivityTracerSpecialFunctionsExt.jl

@@ -101,8 +101,6 @@ ops_2_to_1_ssc = (
 
 for op in ops_2_to_1_ssc
     T = typeof(op)
-    @eval SCT.is_infl_arg1_zero_global(::$T) = false
-    @eval SCT.is_infl_arg2_zero_global(::$T) = false
     @eval SCT.is_der1_arg1_zero_global(::$T) = false
     @eval SCT.is_der2_arg1_zero_global(::$T) = false
     @eval SCT.is_der1_arg2_zero_global(::$T) = false

src/operators.jl

@@ -1,8 +1,7 @@
 ## Operator definitions
 
 # We use a system of letters to categorize operators:
-#   i: independence - no influence at all
-#   z: influence but first- and second-order derivatives (FOD, SOD) are zero
+#   z: first- and second-order derivatives (FOD, SOD) are zero
 #   f: FOD ∂f/∂x is non-zero, SOD ∂²f/∂x² is zero 
 #   s: FOD ∂f/∂x is non-zero, SOD ∂²f/∂x² is non-zero
 #   c: Cross-derivative ∂²f/∂x∂y is non-zero

src/overloads/gradient_tracer.jl

@@ -211,23 +211,41 @@ for S in (Integer, Rational, Irrational{:ℯ})
     Base.:^(::S, t::T) where {T<:GradientTracer} = t
     function Base.:^(d::D, y::S) where {P,T<:GradientTracer,D<:Dual{P,T}}
         x = primal(d)
-        t = tracer(d)
+        t = gradient_tracer_1_to_1(tracer(d), false)
         return Dual(x^y, t)
     end
     function Base.:^(x::S, d::D) where {P,T<:GradientTracer,D<:Dual{P,T}}
         y = primal(d)
-        t = tracer(d)
+        t = gradient_tracer_1_to_1(tracer(d), false)
         return Dual(x^y, t)
     end
 end
 
 ## Rounding
-Base.round(t::T, ::RoundingMode; kwargs...) where {T<:GradientTracer} = myempty(T)
+Base.round(::T, ::RoundingMode; kwargs...) where {T<:GradientTracer} = myempty(T)
 function Base.round(
     d::D, mode::RoundingMode; kwargs...
 ) where {P,T<:GradientTracer,D<:Dual{P,T}}
-    return Dual(round(primal(d), mode; kwargs...), myempty(T))
+    p = round(primal(d), mode; kwargs...)
+    t = myempty(T)
+    return Dual(p, t)
+end
+
+for RR in (Real, Integer, Bool)
+    Base.round(::Type{R}, ::T) where {R<:RR,T<:GradientTracer} = myempty(T)
+    function Base.round(::Type{R}, d::D) where {R<:RR,P,T<:GradientTracer,D<:Dual{P,T}}
+        p = round(R, primal(d))
+        t = myempty(T)
+        return Dual(p, t)
+    end
 end
 
 ## Random numbers 
-Base.rand(::AbstractRNG, ::SamplerType{T}) where {T<:GradientTracer} = myempty(T)  # TODO: was missing Base, add tests
+Base.rand(::AbstractRNG, ::SamplerType{T}) where {T<:GradientTracer} = myempty(T)
+function Base.rand(
+    rng::AbstractRNG, ::SamplerType{D}
+) where {P,T<:GradientTracer,D<:Dual{P,T}}
+    p = rand(rng, P)
+    t = myempty(T)
+    return Dual(p, t)
+end

src/overloads/hessian_tracer.jl

@@ -329,6 +329,29 @@ end
 
 ## Rounding
 Base.round(t::T, ::RoundingMode; kwargs...) where {T<:HessianTracer} = myempty(T)
+function Base.round(
+    d::D, mode::RoundingMode; kwargs...
+) where {P,T<:HessianTracer,D<:Dual{P,T}}
+    p = round(primal(d), mode; kwargs...)
+    t = myempty(T)
+    return Dual(p, t)
+end
+
+for RR in (Real, Integer, Bool)
+    Base.round(::Type{R}, ::T) where {R<:RR,T<:HessianTracer} = myempty(T)
+    function Base.round(::Type{R}, d::D) where {R<:RR,P,T<:HessianTracer,D<:Dual{P,T}}
+        p = round(R, primal(d))
+        t = myempty(T)
+        return Dual(p, t)
+    end
+end
 
 ## Random numbers
-Base.rand(::AbstractRNG, ::SamplerType{T}) where {T<:HessianTracer} = myempty(T)  # TODO: was missing Base, add tests
+Base.rand(::AbstractRNG, ::SamplerType{T}) where {T<:HessianTracer} = myempty(T)
+function Base.rand(
+    rng::AbstractRNG, ::SamplerType{D}
+) where {P,T<:HessianTracer,D<:Dual{P,T}}
+    p = rand(rng, P)
+    t = myempty(T)
+    return Dual(p, t)
+end

src/overloads/ifelse_global.jl

@@ -58,8 +58,6 @@ end
 # Overload only on AbstractTracer, not Dual 
 for op in (isequal, isapprox, isless, ==, <, >, <=, >=)
     T = typeof(op)
-    @eval is_infl_arg1_zero_global(::$T) = false
-    @eval is_infl_arg2_zero_global(::$T) = false
     @eval is_der1_arg1_zero_global(::$T) = true
     @eval is_der2_arg1_zero_global(::$T) = true
     @eval is_der1_arg2_zero_global(::$T) = true

test/test_gradient.jl

@@ -1,9 +1,11 @@
 using SparseConnectivityTracer
 using SparseConnectivityTracer: GradientTracer, Dual, MissingPrimalError, trace_input
+using Test
+
+using Random: rand, GLOBAL_RNG
 using LinearAlgebra: det, dot, logdet
 using SpecialFunctions: erf, beta
 using NNlib: NNlib
-using Test
 
 # Load definitions of GRADIENT_TRACERS, GRADIENT_PATTERNS, HESSIAN_TRACERS and HESSIAN_PATTERNS
 include("tracers_definitions.jl")
@@ -63,7 +65,6 @@ REAL_TYPES = (Float64, Int, Bool, UInt8, Float16, Rational{Int})
         @test J(x -> (2//3)^x, 1) ≈ [1;;]
         @test J(x -> x^ℯ, 1) ≈ [1;;]
         @test J(x -> ℯ^x, 1) ≈ [1;;]
-        @test J(x -> round(x, RoundNearestTiesUp), 1) ≈ [0;;]
         @test J(x -> 0, 1) ≈ [0;;]
 
         # Test special cases on empty tracer
@@ -72,6 +73,18 @@ REAL_TYPES = (Float64, Int, Bool, UInt8, Float16, Rational{Int})
         @test J(x -> zero(x)^ℯ, 1) ≈ [0;;]
         @test J(x -> ℯ^zero(x), 1) ≈ [0;;]
 
+        # Round
+        @test J(round, 1.1) ≈ [0;;]
+        @test J(x -> round(Int, x), 1.1) ≈ [0;;]
+        @test J(x -> round(Bool, x), 1.1) ≈ [0;;]
+        @test J(x -> round(Float16, x), 1.1) ≈ [0;;]
+        @test J(x -> round(x, RoundNearestTiesAway), 1.1) ≈ [0;;]
+        @test J(x -> round(x; digits=3, base=2), 1.1) ≈ [0;;]
+
+        # Random
+        @test J(x -> rand(typeof(x)), 1) ≈ [0;;]
+        @test J(x -> rand(GLOBAL_RNG, typeof(x)), 1) ≈ [0;;]
+
         # Linear Algebra
         @test J(x -> dot(x[1:2], x[4:5]), rand(5)) == [1 1 0 1 1]
 
@@ -202,9 +215,19 @@ end
         @test J(x -> (2//3)^x, 1) ≈ [1;;]
         @test J(x -> x^ℯ, 1) ≈ [1;;]
         @test J(x -> ℯ^x, 1) ≈ [1;;]
-        @test J(x -> round(x, RoundNearestTiesUp), 1) ≈ [0;;]
         @test J(x -> 0, 1) ≈ [0;;]
 
+        # Round
+        @test J(round, 1.1) ≈ [0;;]
+        @test J(x -> round(Int, x), 1.1) ≈ [0;;]
+        @test J(x -> round(Bool, x), 1.1) ≈ [0;;]
+        @test J(x -> round(x, RoundNearestTiesAway), 1.1) ≈ [0;;]
+        @test J(x -> round(x; digits=3, base=2), 1.1) ≈ [0;;]
+
+        # Random
+        @test J(x -> rand(typeof(x)), 1) ≈ [0;;]
+        @test J(x -> rand(GLOBAL_RNG, typeof(x)), 1) ≈ [0;;]
+
         # Linear algebra
         @test J(logdet, [1.0 -1.0; 2.0 2.0]) == [1 1 1 1]  # (#68)
         @test J(x -> log(det(x)), [1.0 -1.0; 2.0 2.0]) == [1 1 1 1]
@@ -213,6 +236,12 @@ end
         # NNlib extension
         @test J(NNlib.relu, -1) ≈ [0;;]
         @test J(NNlib.relu, 1) ≈ [1;;]
+        @test J(NNlib.elu, -1) ≈ [1;;]
+        @test J(NNlib.elu, 1) ≈ [1;;]
+        @test J(NNlib.celu, -1) ≈ [1;;]
+        @test J(NNlib.celu, 1) ≈ [1;;]
+        @test J(NNlib.selu, -1) ≈ [1;;]
+        @test J(NNlib.selu, 1) ≈ [1;;]
 
         @test J(NNlib.relu6, -1) ≈ [0;;]
         @test J(NNlib.relu6, 1) ≈ [1;;]
@@ -221,6 +250,14 @@ end
         @test J(NNlib.trelu, 0.9) ≈ [0;;]
         @test J(NNlib.trelu, 1.1) ≈ [1;;]
 
+        @test J(NNlib.swish, -5) ≈ [1;;]
+        @test J(NNlib.swish, 0) ≈ [1;;]
+        @test J(NNlib.swish, 5) ≈ [1;;]
+
+        @test J(NNlib.hardswish, -5) ≈ [0;;]
+        @test J(NNlib.hardswish, 0) ≈ [1;;]
+        @test J(NNlib.hardswish, 5) ≈ [1;;]
+
         @test J(NNlib.hardσ, -4) ≈ [0;;]
         @test J(NNlib.hardσ, 0) ≈ [1;;]
         @test J(NNlib.hardσ, 4) ≈ [0;;]

test/test_hessian.jl

@@ -1,12 +1,20 @@
 using SparseConnectivityTracer
 using SparseConnectivityTracer: Dual, HessianTracer, MissingPrimalError
 using SparseConnectivityTracer: trace_input, create_tracers, pattern, shared
-using SpecialFunctions: erf, beta
 using Test
 
+using Random: rand, GLOBAL_RNG
+using SpecialFunctions: erf, beta
+using NNlib: NNlib
+
 # Load definitions of GRADIENT_TRACERS, GRADIENT_PATTERNS, HESSIAN_TRACERS and HESSIAN_PATTERNS
 include("tracers_definitions.jl")
 
+# This exists to be able to quickly run tests in the REPL.
+# NOTE: H gets overwritten inside the testsets.
+method = TracerSparsityDetector()
+H(f, x) = hessian_sparsity(f, x, method)
+
 @testset "Global Hessian" begin
     @testset "$P" for P in HESSIAN_PATTERNS
         T = HessianTracer{P}
@@ -26,9 +34,20 @@ include("tracers_definitions.jl")
         @test H(x -> (2//3)^x, 1) ≈ [1;;]
         @test H(x -> x^ℯ, 1) ≈ [1;;]
         @test H(x -> ℯ^x, 1) ≈ [1;;]
-        @test H(x -> round(x, RoundNearestTiesUp), 1) ≈ [0;;]
         @test H(x -> 0, 1) ≈ [0;;]
 
+        # Round
+        @test H(round, 1.1) ≈ [0;;]
+        @test H(x -> round(Int, x), 1.1) ≈ [0;;]
+        @test H(x -> round(Bool, x), 1.1) ≈ [0;;]
+        @test H(x -> round(Float16, x), 1.1) ≈ [0;;]
+        @test H(x -> round(x, RoundNearestTiesAway), 1.1) ≈ [0;;]
+        @test H(x -> round(x; digits=3, base=2), 1.1) ≈ [0;;]
+
+        # Random
+        @test H(x -> rand(typeof(x)), 1) ≈ [0;;]
+        @test H(x -> rand(GLOBAL_RNG, typeof(x)), 1) ≈ [0;;]
+
         @test H(x -> x[1] / x[2] + x[3] / 1 + 1 / x[4], rand(4)) == [
             0 1 0 0
             1 1 0 0
@@ -318,11 +337,59 @@ end
         @test H(x -> ℯ^x, 1) ≈ [1;;]
         @test H(x -> 0, 1) ≈ [0;;]
 
+        # Round
+        @test H(round, 1.1) ≈ [0;;]
+        @test H(x -> round(Int, x), 1.1) ≈ [0;;]
+        @test H(x -> round(Bool, x), 1.1) ≈ [0;;]
+        @test H(x -> round(x, RoundNearestTiesAway), 1.1) ≈ [0;;]
+        @test H(x -> round(x; digits=3, base=2), 1.1) ≈ [0;;]
+
+        # Random
+        @test H(x -> rand(typeof(x)), 1) ≈ [0;;]
+        @test H(x -> rand(GLOBAL_RNG, typeof(x)), 1) ≈ [0;;]
+
         # Test special cases on empty tracer
         @test H(x -> zero(x)^(2//3), 1) ≈ [0;;]
         @test H(x -> (2//3)^zero(x), 1) ≈ [0;;]
         @test H(x -> zero(x)^ℯ, 1) ≈ [0;;]
         @test H(x -> ℯ^zero(x), 1) ≈ [0;;]
+
+        # NNlib extension
+        @test H(NNlib.relu, -1) ≈ [0;;]
+        @test H(NNlib.relu, 1) ≈ [0;;]
+        @test H(NNlib.elu, -1) ≈ [1;;]
+        @test H(NNlib.elu, 1) ≈ [0;;]
+        @test H(NNlib.celu, -1) ≈ [1;;]
+        @test H(NNlib.celu, 1) ≈ [0;;]
+        @test H(NNlib.selu, -1) ≈ [1;;]
+        @test H(NNlib.selu, 1) ≈ [0;;]
+
+        @test H(NNlib.relu6, -1) ≈ [0;;]
+        @test H(NNlib.relu6, 1) ≈ [0;;]
+        @test H(NNlib.relu6, 7) ≈ [0;;]
+
+        @test H(NNlib.trelu, 0.9) ≈ [0;;]
+        @test H(NNlib.trelu, 1.1) ≈ [0;;]
+
+        @test H(NNlib.swish, -5) ≈ [1;;]
+        @test H(NNlib.swish, 0) ≈ [1;;]
+        @test H(NNlib.swish, 5) ≈ [1;;]
+
+        @test H(NNlib.hardswish, -5) ≈ [0;;]
+        @test H(NNlib.hardswish, 0) ≈ [1;;]
+        @test H(NNlib.hardswish, 5) ≈ [0;;]
+
+        @test H(NNlib.hardσ, -4) ≈ [0;;]
+        @test H(NNlib.hardσ, 0) ≈ [0;;]
+        @test H(NNlib.hardσ, 4) ≈ [0;;]
+
+        @test H(NNlib.hardtanh, -2) ≈ [0;;]
+        @test H(NNlib.hardtanh, 0) ≈ [0;;]
+        @test H(NNlib.hardtanh, 2) ≈ [0;;]
+
+        @test H(NNlib.softshrink, -1) ≈ [0;;]
+        @test H(NNlib.softshrink, 0) ≈ [0;;]
+        @test H(NNlib.softshrink, 1) ≈ [0;;]
         yield()
     end
 end