Skip to content

Use safe_sqrt for norms #19

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Use safe_sqrt for norms #19

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ecbd858
Use safe_sqrt for norms
lkdvos Jun 12, 2024
f3c06a8
Compare to absolute eps instead of relative eps
lkdvos Jun 14, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion src/OptimKit.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,9 +12,18 @@ _add!(vdst, vsrc, α) = LinearAlgebra.axpy!(α, vsrc, vdst)
_precondition(x, g) = g
_finalize!(x, f, g, numiter) = x, f, g

abstract type OptimizationAlgorithm
# print error message and return eps(x) if x is negative
function safe_sqrt(x::Real)
return if x >= 0
sqrt(x)
else
-x < eps(x)^(3 / 4) || @error "sqrt of negative number: $x"
eps(x)
end
end

abstract type OptimizationAlgorithm
end
const _xlast = Ref{Any}()
const _glast = Ref{Any}()
const _dlast = Ref{Any}()
Expand Down
6 changes: 3 additions & 3 deletions src/cg.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,13 +24,13 @@ function optimize(fg, x, alg::ConjugateGradient;
f, g = fg(x)
numfg = 1
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
fhistory = [f]
normgradhistory = [normgrad]

# compute here once to define initial value of α in scale-invariant way
Pg = precondition(x, g)
normPg = sqrt(inner(x, Pg, Pg))
normPg = safe_sqrt(inner(x, Pg, Pg))
α = 1/(normPg) # initial guess: scale invariant
# α = one(normgrad)

Expand Down Expand Up @@ -71,7 +71,7 @@ function optimize(fg, x, alg::ConjugateGradient;
numiter += 1
x, f, g = finalize!(x, f, g, numiter)
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
push!(fhistory, f)
push!(normgradhistory, normgrad)

Expand Down
6 changes: 3 additions & 3 deletions src/gd.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,13 +19,13 @@ function optimize(fg, x, alg::GradientDescent;
f, g = fg(x)
numfg = 1
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
fhistory = [f]
normgradhistory = [normgrad]

# compute here once to define initial value of α in scale-invariant way
Pg = precondition(x, g)
normPg = sqrt(inner(x, Pg, Pg))
normPg = safe_sqrt(inner(x, Pg, Pg))
α = 1/(normPg) # initial guess: scale invariant

numiter = 0
Expand All @@ -46,7 +46,7 @@ function optimize(fg, x, alg::GradientDescent;
numiter += 1
x, f, g = finalize!(x, f, g, numiter)
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
push!(fhistory, f)
push!(normgradhistory, normgrad)

Expand Down
12 changes: 6 additions & 6 deletions src/lbfgs.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@ function optimize(fg, x, alg::LBFGS;
f, g = fg(x)
numfg = 1
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
fhistory = [f]
normgradhistory = [normgrad]

Expand All @@ -43,7 +43,7 @@ function optimize(fg, x, alg::LBFGS;
η = scale!(Hg, -1)
else
Pg = precondition(x, deepcopy(g))
normPg = sqrt(inner(x, Pg, Pg))
normPg = safe_sqrt(inner(x, Pg, Pg))
η = scale!(Pg, -1/normPg) # initial guess: scale invariant
end

Expand All @@ -64,7 +64,7 @@ function optimize(fg, x, alg::LBFGS;
numiter += 1
x, f, g = finalize!(x, f, g, numiter)
innergg = inner(x, g, g)
normgrad = sqrt(innergg)
normgrad = safe_sqrt(innergg)
push!(fhistory, f)
push!(normgradhistory, normgrad)

Expand Down Expand Up @@ -95,8 +95,8 @@ function optimize(fg, x, alg::LBFGS;
# define new isometric transport such that, applying it to transported ηprev,
# it returns a vector proportional to ξ but with the norm of ηprev
# still has norm normη because transport is isometric
normη = sqrt(inner(x, ηprev, ηprev))
normξ = sqrt(inner(x, ξ, ξ))
normη = safe_sqrt(inner(x, ηprev, ηprev))
normξ = safe_sqrt(inner(x, ξ, ξ))
β = normη/normξ
if !(inner(x, ξ, ηprev) ≈ normξ * normη) # ξ and η are not parallel
ξ₁ = ηprev
Expand Down Expand Up @@ -131,7 +131,7 @@ function optimize(fg, x, alg::LBFGS;
innerss = inner(x, s, s)

if innersy/innerss > normgrad/10000
norms = sqrt(innerss)
norms = safe_sqrt(innerss)
ρ = innerss/innersy
push!(H, (scale!(s, 1/norms), scale!(y, 1/norms), ρ))
end
Expand Down