Map Training Routines

CreateGaussianKLObjective(train)
CreateGaussianKLObjective(train, outputDim)
CreateGaussianKLObjective(train, test)
CreateGaussianKLObjective(train, test, outputDim)

Create an objective for the optimization problem of training a transport map.

Currently only supports variational simulation-based inference, i.e. creating a map that uses samples train from target distribution and a Gaussian reference distribution of dimension outputDim. If outputDim is zero, this is equivalent to when outputDim == size(train,1).

Arguments

• train::Matrix{Float64}: mandatory training dataset
• test::Matrix{Float64}: optional test dataset
• outputDim::Int = 0: Dimensions of output of map

Examples

julia> using Random; rng = Random.Xoshiro(1);

julia> # Replace RNG type with MersenneTwister for <1.7

julia> N, inDim, outDim = 1000, 4, 2;

julia> samples = 2*randn(rng, inDim, N) .+ 5;

julia> obj1 = CreateGaussianKLObjective(samples);

julia> obj2 = CreateGaussianKLObjective(samples, outDim);

julia> train, test = samples[:,1:500], samples[:,501:end];

julia> obj3 = CreateGaussianKLObjective(train, test);

julia> obj4 = CreateGaussianKLObjective(train, test, outDim);

See also TrainMap, TrainOptions

`TrainOptions(;kwargs...)`

Creates options for using TrainMap to train a transport map.

See example for possible arguments.

Examples

julia> TrainOptions(opt_alg="LD_SLSQP", opt_maxeval = 1_000_000)
opt_alg = LD_SLSQP
opt_stopval = -inf
opt_ftol_rel = 0.001
opt_ftol_abs = 0.001
opt_xtol_rel = 0.0001
opt_xtol_abs = 0.0001
opt_maxeval = 1000000
opt_maxtime = inf
verbose = 0

See also TrainMap, CreateGaussianKLObjective

TrainMap(map, obj::MapObjective, opts::TrainOptions)

Trains map according to the objective obj with training options opts.

TestError(obj::MapObjective, map)

Uses the test dataset in obj to evaluate the error of map.