JuliaAI / DecisionTree.jl

Julia implementation of Decision Tree (CART) and Random Forest algorithms
Other
342 stars 101 forks source link

Memory leakage upon repeated training #230

Open CasBex opened 5 months ago

CasBex commented 5 months ago

Hi, I've been creating some random forest regressors lately and I've noticed high memory usage during hyperparameter tuning. It turns out that there is some memory leakage in the package. For some reason Julia does not delete the trees when they become unreachable.

Following is a MWE: after finishing run_forests, some memory should be reclaimed but it doesn't happen and memory usage increases. When running the second loop however, memory usage stays constant.

using DecisionTree
function run_forests(features, labels)
    forest = build_forest(labels, features)
    labels .+= apply_forest(forest, features)
    labels ./= 2
end

function run_something_else(features, labels)
    C = repeat(features, inner=(2,2))
    labels ./= vec(sum(C, dims=2))[1:length(labels)]
end

const features = rand(10_000, 10)
const labels = sum(features, dims=2) |> vec

# notice memory consumption increases every couple of iterations
for i = 1:1_000
    run_forests(features, labels)
    @info "Iteration $i current memory used" Sys.maxrss()
end

# notice memory consumption does not increase every couple of iterations
for i = 1:1_000
    run_something_else(features, labels)
    @info "Iteration $i current memory used" Sys.maxrss()
end

Any idea what might cause this?

rikhuijzer commented 5 months ago

Yes. I can confirm on Julia 1.10 aarch64 Apple Darwin. During execution, htop memory usage will slowly increase each time that I call include("tmp.jl").

Any idea what might cause this?

Could be a (multi-threading) leak somewhere in Julia: https://github.com/search?q=repo%3AJuliaLang%2Fjulia+memory&type=issues&p=2