Hi , i'm using java XGBoost in ADAMS-WEKA !

[Daniel Debuisson]

setNumThreads(16) does'nt seems to work !  

only one thread is running and of course it's take a very long time

looking for doc but nothing explained about that ???
need help thank's !

[Peter Reutemann]

I just ran XGBoost on my Linux machine and I can see a difference in CPU usage and speed between using 1 thread and more than 1. However, on the dataset that I've tested it with, there is no linear relationship between number of threads and the speed up in building/evaluating the models.

The following numbers are taken from a single run of 10-fold cross-validation in the Weka Investigator:
- 1 thread: 6.7s
- 2 threads: 4.9s
- 8 threads: 4.2s

The dataset that I used had only numeric attributes:
# Attributes: 1193
# Instances: 10000

The XGBoost classifier in ADAMS just wraps around the following library:
  https://github.com/dmlc/xgboost
Specifically version 2.1.0 of xgboost4j_2.12.

Documentation for the parameters is available here:
https://xgboost.readthedocs.io/en/release_2.1.0/parameter.html

The "numThreads" option in the ADAMS wrapper gets translated to the "nthread" parameter for the underlying xgboost algorithm.

Cheers, Peter

--
Peter Reutemann
Dept. of Computer Science
University of Waikato, Hamilton, NZ
Mobile +64 22 190 2375
https://profiles.waikato.ac.nz/peter.reutemann
http://www.data-mining.co.nz/