Distributed training with XGBoost and Dask.distributed

This repository offers a legacy option to perform distributed training with XGBoost on Dask.array and Dask.dataframe collections.

pip install dask-xgboost

Please note that XGBoost now includes a Dask API as part of its official Python package. That API is independent of dask-xgboost and is now the recommended way to use Dask adn XGBoost together. See the xgb.dask documentation here https://xgboost.readthedocs.io/en/latest/tutorials/dask.html for more details on the new API.

from dask.distributed import Client
client = Client('scheduler-address:8786')  # connect to cluster
import dask.dataframe as dd
df = dd.read_csv('...')  # use dask.dataframe to load and
df_train = ...           # preprocess data
labels_train = ...
import dask_xgboost as dxgb
params = {'objective': 'binary:logistic', ...}  # use normal xgboost params
bst = dxgb.train(client, params, df_train, labels_train)
>>> bst  # Get back normal XGBoost result
<xgboost.core.Booster at ... >
predictions = dxgb.predict(client, bst, data_test)

For more information on using Dask.dataframe for preprocessing see the Dask.dataframe documentation.

Once you have created suitable data and labels we are ready for distributed training with XGBoost. Every Dask worker sets up an XGBoost slave and gives them enough information to find each other. Then Dask workers hand their in-memory Pandas dataframes to XGBoost (one Dask dataframe is just many Pandas dataframes spread around the memory of many machines). XGBoost handles distributed training on its own without Dask interference. XGBoost then hands back a single xgboost.Booster result object.

For a more serious example see

• This blogpost
• This notebook
• This screencast

Conversation during development happened at dmlc/xgboost #2032