afps(f, x; iters::Int = 5000, vel::Float64 = 0.9, ep::Float64 = 0.01, tol::Float64 = 1e-12, grad_norm=x->maximum(abs,x))
solve equation f(x) = x according to:
`f` : function to find fixed point for
`x` : initial condition, ideally it should be close to the final solution
`vel` : amount of Nesterov acceleration in [0,1]
`ep` : learning rate, typically in ]0,1[
`tol` : absolute tolerance on |f(x)-x|
`grad_norm` : function to evaluate the norm |f(x)-x|
returns a named tuple (x, error, iters) where:
`x` : is the solution found for f(x)=x
`error` : is the norm of f(x)-x at the solution point
`iters` : total number of iterations performed
using FixedPoint
s = afps(x -> 2 - x^2 + x, 1.3)
@show s.x, √2using FixedPoint, LinearAlgebra
Ts = LinRange(0.01, 2.0, 500)
βs = 1 ./ Ts
f(x) = tanh.(βs .* x)
s = afps(
f,
zero(βs) .+ 1,
iters = 5000,
)
@show norm(f(s.x).-s.x)for the inplace method use afps! as
Ts = LinRange(0.01, 2.0, 500)
βs = 1 ./ Ts
function f!(out,x)
@. out = tanh(βs * x)
end
x = zero(βs) .+ 1
afps!(
f!,
x,
grad_norm = x -> maximum(abs, x),
iters = 5000,
)
@show maximum(abs, x .- tanh.(βs .* x))