Unify interface for linesearches (#121)

* Standardise line search arguments

* Update example with new interface.

Author: anriseth

docs/src/examples/customoptimizer.jl

@@ -44,8 +44,8 @@ function gdoptimize(f, g!, fg!, x0::AbstractArray{T}, linesearch,
         s .= -gvec
 
         dϕ_0 = dot(s, gvec)
-        α, fx = perform_linesearch(ϕ, dϕ, ϕdϕ, 1.0,
-                                   fx, dϕ_0, linesearch)
+        α, fx = linesearch(ϕ, dϕ, ϕdϕ, 1.0, fx, dϕ_0)
+
         @. x = x + α*s
         g!(gvec, x)
         gnorm = norm(gvec)
@@ -62,29 +62,12 @@ end
 # LineSearches algorithms do.
 
 # The Gradient Descent `gdoptimize` method selects a descent direction and calls
-# a method `perform_linesearch` that returns the step length `α` and the
+# the line search algorithm  `linesearch` which returns the step length `α` and the
 # objective value `fx = f(x + α*s)`.
 #
-# We use multiple dispatch on `linesearch` to call the different line search procedures:
-
-using LineSearches
-perform_linesearch(ϕ, dϕ, ϕdϕ, α0, ϕ_0, dϕ_0,
-                   linesearch::BackTracking) =
-                       linesearch(ϕ, α0, ϕ_0, dϕ_0)
-perform_linesearch(ϕ, dϕ, ϕdϕ, α0, ϕ_0, dϕ_0,
-                   linesearch::HagerZhang) =
-                       linesearch(ϕ, ϕdϕ, α0, ϕ_0, dϕ_0)
-perform_linesearch(ϕ, dϕ, ϕdϕ, α0, ϕ_0, dϕ_0,
-                   linesearch::MoreThuente) =
-                       linesearch(ϕdϕ, α0, ϕ_0, dϕ_0)
-perform_linesearch(ϕ, dϕ, ϕdϕ, α0, ϕ_0, dϕ_0,
-                   linesearch::StrongWolfe) =
-                       linesearch(ϕ, dϕ, ϕdϕ, α0, ϕ_0, dϕ_0)
-
 # The functions ϕ and dϕ represent a univariate objective
 # and its derivative, which is used by the line search algorithms.
-# To utilize the `fg!` function call in the optimizer, `HagerZhang`,
-# `MoreThuente`, and `StrongWolfe` also
+# To utilize the `fg!` function call in the optimizer, some of the line searches
 # require a function ϕdϕ which returns the univariate objective and the
 # derivative at the same time.
 
@@ -111,6 +94,8 @@ end
 # from a given initial condition `x0`.
 
 x0 = [-1., 1.0]
+
+using LineSearches
 ls = BackTracking(order=3)
 fx_bt3, x_bt3, iter_bt3 = gdoptimize(f, g!, fg!, x0, ls)
 

src/backtracking.jl

@@ -32,8 +32,11 @@ function (ls::BackTracking)(df::AbstractObjective, x::AbstractArray{T}, s::Abstr
     α_0 = min(α_0, min(alphamax, ls.maxstep / vecnorm(s, Inf)))
     ls(ϕ, α_0, ϕ_0, dϕ_0)
 end
-function (ls::BackTracking)(ϕ, α_0::Tα, ϕ_0, dϕ_0) where Tα
 
+(ls::BackTracking)(ϕ, dϕ, ϕdϕ, αinitial, ϕ_0, dϕ_0) = ls(ϕ, αinitial, ϕ_0, dϕ_0)
+
+# TODO: Should we deprecate the interface that only uses the ϕ argument?
+function (ls::BackTracking)(ϕ, αinitial::Tα, ϕ_0, dϕ_0) where Tα
     @unpack c_1, ρ_hi, ρ_lo, iterations, order = ls
 
     iterfinitemax = -log2(eps(real(Tα)))
@@ -52,7 +55,7 @@ function (ls::BackTracking)(ϕ, α_0::Tα, ϕ_0, dϕ_0) where Tα
 
     ϕx_0, ϕx_1 = ϕ_0, ϕ_0
 
-    α_1, α_2 = α_0, α_0
+    α_1, α_2 = αinitial, αinitial
 
     ϕx_1 = ϕ(α_1)
 

src/hagerzhang.jl

@@ -101,11 +101,13 @@ function (ls::HagerZhang)(df::AbstractObjective, x::AbstractArray{T},
     ls(ϕ, ϕdϕ, α::Real, phi_0, dphi_0)
 end
 
-function (ls::HagerZhang)(ϕ, ϕdϕ,
-                     c::T,
-                     phi_0::Real,
-                     dphi_0::Real) where T
+(ls::HagerZhang)(ϕ, dϕ, ϕdϕ, c, phi_0, dphi_0) = ls(ϕ, ϕdϕ, c, phi_0, dphi_0)
 
+# TODO: Should we deprecate the interface that only uses the ϕ and ϕd\phi arguments?
+function (ls::HagerZhang)(ϕ, ϕdϕ,
+                          c::T,
+                          phi_0::Real,
+                          dphi_0::Real) where T # Should c and phi_0 be same type?
     @unpack delta, sigma, alphamax, rho, epsilon, gamma,
             linesearchmax, psi3, display, mayterminate = ls
 

src/morethuente.jl

@@ -153,10 +153,14 @@ function (ls::MoreThuente)(df::AbstractObjective, x::AbstractArray{T},
     ϕdϕ = make_ϕdϕ(df, x_new, x, s)
     ls(ϕdϕ, alpha, ϕ_0, dϕ_0)
 end
+
+(ls::MoreThuente)(ϕ, dϕ, ϕdϕ, alpha, ϕ_0, dϕ_0) = ls(ϕdϕ, alpha, ϕ_0, dϕ_0)
+
+# TODO: Should we deprecate the interface that only uses the ϕdϕ argument?
 function (ls::MoreThuente)(ϕdϕ,
-                  alpha::T,
-                  ϕ_0,
-                  dϕ_0) where T
+                           alpha::T,
+                           ϕ_0,
+                           dϕ_0) where T
     @unpack f_tol, gtol, x_tol, alphamin, alphamax, maxfev = ls
 
     iterfinitemax = -log2(eps(T))

src/static.jl

@@ -37,6 +37,8 @@ function (ls::DeprecatedStatic)(df::AbstractObjective, x, s, α, x_new = similar
     ls(ϕ, α)
 end
 
+(ls::DeprecatedStatic)(ϕ, dϕ, ϕdϕ, α, ϕ_0, dϕ_0) = ls(ϕ, α)
+
 function (ls::DeprecatedStatic)(ϕ, α::Tα) where Tα
     @unpack α, scaled = ls
     @assert α > Tα(0) # This should really be done at the constructor level
@@ -73,10 +75,13 @@ immutable NewStatic end
 
 function (ls::NewStatic)(df::AbstractObjective, x, s, α, x_new = similar(x), ϕ_0 = nothing, dϕ_0 = nothing)
     ϕ = make_ϕ(df, x_new, x, s)
-    ls(ϕ, x, s, α)
+    ls(ϕ, α)
 end
 
-function (ls::NewStatic)(ϕ, x, s, α::Tα) where Tα
+(ls::NewStatic)(ϕ, dϕ, ϕdϕ, α, ϕ_0, dϕ_0) = ls(ϕ, α)
+
+# TODO: Should we deprecate the interface that only uses the ϕ argument?
+function (ls::NewStatic)(ϕ, α::Tα) where Tα
     @assert α > real(Tα(0))
     ϕα = ϕ(α)