Model storage and persistence

[mhangaard]

What would be good ways to store and persist a GP optimized with GP.jl? The purpose would be to make predictions with mean and standard deviation at a later time than training.

[thomaspinder]

Hi @mhangaard, this is something I've thought about before using BSON.jl.

It's certainly something that would be useful in the package, however, I'm not sure I personally have the time currently, so I've left it as an open issue.

[maximerischard]

If you keep hold of the training data and the kernel specification, all you need is the vector of parameters output by get_params. Then you would just re-create the GP object and feed the stored parameters to set_params!. Anything beyond that (such as some kind of serialisation scheme for the kernel and mean specification) would need some serious thought.

[thomaspinder]

One option for this @mhangaard is to use JLD.jl. A minimal working example would look something like

using Random, Distributions, GaussianProcesses, JLD
Random.seed!(13579)               # Set the seed using the 'Random' package
n = 20;                           # number of training points
x = 2π * rand(n);                 # predictors
y = sin.(x) + 0.05*randn(n);      # regressors

# Select mean and covariance function
mZero = MeanZero()                  # Zero mean function
kern = SE(0.0,0.0)                  # Sqaured exponential kernel
logObsNoise = -1.0                  # log standard deviation of observation noise
gp = GP(x,y,mZero,kern,logObsNoise) # Fit the GP
optimize!(gp) #Optimise the parameters

From which the model can be saved by save("gp_model.jld", "gp", gp) and consequently loaded back in by gp = load("gp_model.jld", "gp").