Fix constructors to allow non unicode

Author: gdalle

src/frank_wolfe/differentiable_frank_wolfe.jl

@@ -8,7 +8,7 @@ Parameterized version of the Frank-Wolfe algorithm `θ -> argmin_{x ∈ C} f(x,
 # Fields
 - `f::F`: function `f(x, θ)` to minimize wrt `x`
 - `f_grad1::G`: gradient `∇ₓf(x, θ)` of `f` wrt `x`
-- `lmo::M`: linear minimization oracle `θ -> argmin_{x ∈ C} θᵀx` which implicitly defines the polytope `C`
+- `lmo::M`: linear minimization oracle `θ -> argmin_{x ∈ C} θᵀx`, implicitly defines the polytope `C`
 - `linear_solver::S`: solver for linear systems of equations, used during implicit differentiation
 
 # Applicable methods
@@ -24,7 +24,7 @@ struct DifferentiableFrankWolfe{F,G,M<:LinearMinimizationOracle,S}
     linear_solver::S
 end
 
-function DifferentiableFrankWolfe(f, f_grad1, lmo; linear_solver=gmres)
+function DifferentiableFrankWolfe(f, f_grad1, lmo, linear_solver=gmres)
     return DifferentiableFrankWolfe(f, f_grad1, lmo, linear_solver)
 end
 

src/perturbed/abstract_perturbed.jl

@@ -17,9 +17,9 @@ These subtypes share the following fields:
 
 - `maximizer`: black box optimizer
 - `ε`: magnitude of the perturbation
+- `nb_samples::Int`: number of random samples for Monte-Carlo computations
 - `rng::AbstractRNG`: random number generator
 - `seed::Union{Nothing,Int}`: random seed
-- `nb_samples::Int`: number of random samples for Monte-Carlo computations
 """
 abstract type AbstractPerturbed end
 

src/perturbed/additive.jl

@@ -1,7 +1,7 @@
 """
     PerturbedAdditive{F}
 
-Differentiable normal perturbation of a black-box optimizer: the input undergoes `θ -> θ + εZ` where `Z ∼ N(0, I)`.
+Differentiable normal perturbation of a black-box optimizer of type `F`: the input undergoes `θ -> θ + εZ` where `Z ∼ N(0, I)`.
 
 See also: [`AbstractPerturbed`](@ref).
 
@@ -10,25 +10,30 @@ Reference: <https://arxiv.org/abs/2002.08676>
 struct PerturbedAdditive{F,R<:AbstractRNG,S<:Union{Nothing,Int}} <: AbstractPerturbed
     maximizer::F
     ε::Float64
+    nb_samples::Int
     rng::R
     seed::S
-    nb_samples::Int
 end
 
 function Base.show(io::IO, perturbed::PerturbedAdditive)
     (; maximizer, ε, rng, seed, nb_samples) = perturbed
     return print(
-        io, "PerturbedAdditive($maximizer, $ε, $(typeof(rng)), $seed, $nb_samples)"
+        io, "PerturbedAdditive($maximizer, $ε, $nb_samples, $(typeof(rng)), $seed)"
     )
 end
 
+"""
+    PerturbedAdditive(maximizer[; ε=1.0, nb_samples=1])
+
+Shorter constructor with defaults.
+"""
 function PerturbedAdditive(
-    maximizer; ε=1.0, epsilon=nothing, rng=MersenneTwister(0), seed=nothing, nb_samples=2
+    maximizer; ε=1.0, epsilon=nothing, nb_samples=1, rng=MersenneTwister(0), seed=nothing
 )
     if isnothing(epsilon)
-        return PerturbedAdditive(maximizer, float(ε), rng, seed, nb_samples)
+        return PerturbedAdditive(maximizer, float(ε), nb_samples, rng, seed)
     else
-        return PerturbedAdditive(maximizer, float(epsilon), rng, seed, nb_samples)
+        return PerturbedAdditive(maximizer, float(epsilon), nb_samples, rng, seed)
     end
 end
 

src/perturbed/multiplicative.jl

@@ -1,7 +1,7 @@
 """
     PerturbedMultiplicative{F}
 
-Differentiable log-normal perturbation of a black-box optimizer: the input undergoes `θ -> θ ⊙ exp[εZ - ε²/2]` where `Z ∼ N(0, I)`.
+Differentiable log-normal perturbation of a black-box optimizer of type `F`: the input undergoes `θ -> θ ⊙ exp[εZ - ε²/2]` where `Z ∼ N(0, I)`.
 
 See also: [`AbstractPerturbed`](@ref).
 
@@ -10,25 +10,30 @@ Reference: preprint coming soon.
 struct PerturbedMultiplicative{F,R<:AbstractRNG,S<:Union{Nothing,Int}} <: AbstractPerturbed
     maximizer::F
     ε::Float64
+    nb_samples::Int
     rng::R
     seed::S
-    nb_samples::Int
 end
 
 function Base.show(io::IO, perturbed::PerturbedMultiplicative)
     (; maximizer, ε, rng, seed, nb_samples) = perturbed
     return print(
-        io, "PerturbedMultiplicative($maximizer, $ε, $(typeof(rng)), $seed, $nb_samples)"
+        io, "PerturbedMultiplicative($maximizer, $ε, $nb_samples, $(typeof(rng)), $seed)"
     )
 end
 
+"""
+    PerturbedMultiplicative(maximizer[; ε=1.0, nb_samples=1])
+
+Shorter constructor with defaults.
+"""
 function PerturbedMultiplicative(
-    maximizer; ε=1.0, epsilon=nothing, rng=MersenneTwister(0), seed=nothing, nb_samples=2
+    maximizer; ε=1.0, epsilon=nothing, nb_samples=1, rng=MersenneTwister(0), seed=nothing
 )
     if isnothing(epsilon)
-        return PerturbedMultiplicative(maximizer, float(ε), rng, seed, nb_samples)
+        return PerturbedMultiplicative(maximizer, float(ε), nb_samples, rng, seed)
     else
-        return PerturbedMultiplicative(maximizer, float(epsilon), rng, seed, nb_samples)
+        return PerturbedMultiplicative(maximizer, float(epsilon), nb_samples, rng, seed)
     end
 end
 

src/regularized/regularized_generic.jl

@@ -6,7 +6,7 @@ Differentiable regularized prediction function `ŷ(θ) = argmax_{y ∈ C} {θ
 Relies on the Frank-Wolfe algorithm to minimize a concave objective on a polytope.
 
 # Fields
-- `maximizer::M`: linear maximization oracle `θ -> argmax_{x ∈ C} θᵀx` which implicitly defines the polytope `C`
+- `maximizer::M`: linear maximization oracle `θ -> argmax_{x ∈ C} θᵀx`, implicitly defines the polytope `C`
 - `Ω::RF`: regularization function `Ω(y)`
 - `Ω_grad::RG`: gradient of the regularization function `∇Ω(y)`
 - `f::F`: objective function `f(x, θ) = Ω(y) - θᵀy` minimized by Frank-Wolfe (computed automatically)
@@ -34,17 +34,32 @@ function Base.show(io::IO, regularized::RegularizedGeneric)
     return print(io, "RegularizedGeneric($maximizer, $Ω, $Ω_grad, $linear_solver)")
 end
 
-"""
-    RegularizedGeneric(maximizer, Ω, Ω_grad[; linear_solver=gmres])
-
-Short form constructor with a default linear solver.
-"""
-function RegularizedGeneric(maximizer, Ω, Ω_grad; linear_solver=gmres)
+function RegularizedGeneric(maximizer, Ω, Ω_grad, linear_solver=gmres)
     f(y, θ) = Ω(y) - dot(θ, y)
     f_grad1(y, θ) = Ω_grad(y) - θ
     return RegularizedGeneric(maximizer, Ω, Ω_grad, f, f_grad1, linear_solver)
 end
 
+"""
+    RegularizedGeneric(maximizer[; Ω, Ω_grad, linear_solver=gmres])
+
+Shorter constructor with defaults.
+"""
+function RegularizedGeneric(
+    maximizer;
+    Ω=zero_regularization,
+    Ω_grad=zero_gradient,
+    omega=nothing,
+    omega_grad=nothing,
+    linear_solver=gmres,
+)
+    if isnothing(omega) || isnothing(omega_grad)
+        return RegularizedGeneric(maximizer, Ω, Ω_grad, linear_solver)
+    else
+        return RegularizedGeneric(maximizer, omega, omega_grad, linear_solver)
+    end
+end
+
 @traitimpl IsRegularized{RegularizedGeneric}
 
 function compute_regularization(regularized::RegularizedGeneric, y::AbstractArray{<:Real})

src/regularized/regularized_utils.jl

@@ -64,3 +64,6 @@ Compute the vector `r` such that `rᵢ` is the rank of `θᵢ` in `θ`.
 function ranking(θ::AbstractVector{<:Real}; rev::Bool=false, kwargs...)
     return invperm(sortperm(θ; rev=rev))
 end
+
+zero_regularization(y) = zero(eltype(y))
+zero_gradient(y) = zero(y)