StructNewtonExperiments

This project contains the code used to produce the numerical result of the paper Newton Acceleration on Manifolds identified by Proximal-Gradient Methods. It has three main package dependencies:

StructuredProximalOperators.jl implements some nonsmooth functions, their proximity operator, and Riemannian gradients and Hessians;
CompositeProblems.jl implements some additive composite problems, and provide oracles for the smooth and nonsmooth part;
StructuredSover.jl implements solvers designed to tackle composite problems, such as the (Accelerated) Proximal Gradient, or the methods of the above mentioned paper.

Running numerical experiments

The numerical experiments may be run as follows:

first add the three above packages by running

using Pkg
Pkg.add(url="https://github.com/GillesBareilles/StructuredProximalOperators.jl", rev="v0.1")
Pkg.add(url="https://github.com/GillesBareilles/CompositeProblems.jl", rev="v0.1")
Pkg.add(url="https://github.com/GillesBareilles/StructuredSolvers.jl", rev="v0.1")
Pkg.add(url="https://github.com/GillesBareilles/NewtonRiemannAccel-ProxGrad", rev="v0.1")

then by running the following commands:

using StructNewtonExperiments

# running each problem separately
run_expe_maxquad(NUMEXPS_OUTDIR = ".");
run_expe_logistic(NUMEXPS_OUTDIR = ".");
run_expe_tracenorm(NUMEXPS_OUTDIR = ".");

# running all problems
run_expes(NUMEXPS_OUTDIR = ".");

Alternatively, the numerical experiments may be run using docker as follows:

First, build the julia image and install packages: run from a folder containing the Dockerfile

sudo docker build -t julia_structnewton .

Then run the docker image by executing the following command, where /ABSOLUTE/PATH/TO/OUTDIR is the absolute path to the directory where the numerical experiments files will be written:

docker run --mount type=bind,source=/ABSOLUTE/PATH/TO/OUTDIR,target=/root/numexps_output -it julia_structnewton

In the opened julia REPL, run the following

using StructNewtonExperiments
run_expes(NUMEXPS_OUTDIR="/root/numexps_output/")