Created
March 5, 2021 00:39
-
-
Save jiweiqi/ecb8fd57728f333dea4cf56a2a28d877 to your computer and use it in GitHub Desktop.
demo of lm optimization
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
using LsqFit | |
using ForwardDiff | |
using Plots | |
using Flux | |
using MINPACK | |
xdata = 0.0:0.1:10.0 | |
xdata = xdata' | |
ydata = Float64.(vec(sin.(xdata))) | |
nn = Chain(Dense(1, 5, tanh), Dense(5, 1)) | |
p, re = Flux.destructure(nn) | |
function f!(fvec, p) | |
ypred = vec(re(p)(xdata)) | |
@. fvec = abs(ypred - ydata) | |
return fvec | |
end | |
fvec = similar(ydata) | |
f!(fvec, p) | |
function g!(fjac, p) | |
ForwardDiff.jacobian!(fjac, (fvec, p) -> f!(fvec, p), fvec, p) | |
return fjac | |
end | |
fjac = zeros(length(ydata), length(p)) | |
g!(fjac, p) | |
p0 = Float64.(p); | |
m = length(ydata) | |
res = fsolve( | |
f!, | |
g!, | |
p0, | |
m, | |
iterations = 300, | |
tol = 1e-8, | |
show_trace = true, | |
tracing = true; | |
method = :lm, | |
) | |
l_loss = ones(res.trace.f_calls, 4) | |
for i = 1:res.trace.f_calls | |
l_loss[i, 1] = res.trace.trace[i].iteration | |
l_loss[i, 2] = res.trace.trace[i].fnorm | |
l_loss[i, 3] = res.trace.trace[i].xnorm | |
l_loss[i, 4] = res.trace.trace[i].step_time | |
end | |
l_plt = [] | |
plt = scatter(xdata[:], ydata[:], label = "data"); | |
plot!(plt, xdata[:], vec(re(res.x)(xdata)), label = "pred"); | |
push!(l_plt, plt) | |
plt = plot( | |
l_loss[:, 1], | |
l_loss[:, 2], | |
xscale = :identity, | |
yscale = :log10, | |
label = "fnorm", | |
); | |
push!(l_plt, plt) | |
plt = plot( | |
l_loss[:, 1], | |
l_loss[:, 3] .+ 1.e-6, | |
xscale = :identity, | |
yscale = :log10, | |
label = "xnorm", | |
); | |
push!(l_plt, plt) | |
plt = plot(l_plt...) | |
png(plt, "lmnn") |
Author
jiweiqi
commented
Mar 5, 2021
Using LsqFit.jl
using LsqFit
using ForwardDiff
using Plots
using Flux
xdata = 0.0:0.1:10.0
xdata = xdata'
ydata = Float64.(vec(sin.(xdata)))
nn = Chain(
Dense(1, 5, tanh),
Dense(5, 1))
x, re = Flux.destructure(nn)
function f(p)
ypred = vec(re(p)(xdata))
return @. abs(ypred - ydata)
end
f(x)
function g(p)
return ForwardDiff.jacobian(x -> f(x), p)
end
g(x)
res = LsqFit.lmfit(f, g, Float64.(x), Float64[];
show_trace=true, maxIter=1000, x_tol=1e-8)
display(res.param)
plt = scatter(xdata[:], ydata[:], label="data");
plot!(plt, xdata[:], vec(re(res.param)(xdata)), label="pred");
png(plt, "lmnn")
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment