search

dmlc / XGBoost.jl

XGBoost Julia Package
Other
289 stars 110 forks source link

Is support for GPU available using parameters? #70

Closed caxelrud closed 4 years ago

caxelrud commented 5 years ago

Is it possible to use parameters as "tree_method= gpu_hist"? Are examples available?

dpk1729 commented 5 years ago

In order to be able to train on gpu, you should first compile xgboost with cuda flag and then replace the .so file in your julia packages folder with this new file. Here's how you can do that

Clone xgboost repo, and pull the submodules

git clone https://github.com/dmlc/xgboost.git
cd xgboost/
git submodule update --init --recursive

Build xgboost

mkdir build
cd build
cmake .. -DUSE_CUDA=ON
make -j4

This creates a libxgboost.so file in the lib folder. Copy this file to the XGBoost package directory to the location where julia installed xgboost cp xgboost/lib/libxgboost.so .julia/packages/XGBoost/...../deps/usr/lib

After this you should be able to train on gpu with bst = xgboost(data, ....,tree_method="gpu_hist")

caxelrud commented 5 years ago

I am using Windows 10. I got the dll from this site: https://picnet.com.au/blog/xgboost-windows-x64-binaries-for-download/ And replaced at the Julia package location. It is not complaining when using method="gpu_hist" so it looks like is working.

aviks commented 4 years ago

closing this issue, this is the way to do this for now.