Xgboost: Predicting in submilliseconds

Created on 26 Jul 2017  路  1Comment  路  Source: dmlc/xgboost

Most helpful comment

Lately there's been greater interest in faster prediction. In the near future, the dmlc group will host a new project (provisionally named) tree-lite that is dedicated to improving prediction speed.

Some facts about tree-lite:

  • It will read your (already trained) tree ensemble model and produce an optimized prediction library.
  • We aim to provide a well assorted toolbox of methods that are known to improve prediction speed. People will be able to experiment with it on their own. Good engineering will be essential in this effort.
  • It will support multiple tree formats, so that you can use other packages (e.g. LightGBM) to train your model.
  • The end goal is to maximize prediction throughput (# rows / second) given the latency constraint.

The project is currently private, but we are working hard to make it public in the near future. Stay tuned!

>All comments

Lately there's been greater interest in faster prediction. In the near future, the dmlc group will host a new project (provisionally named) tree-lite that is dedicated to improving prediction speed.

Some facts about tree-lite:

  • It will read your (already trained) tree ensemble model and produce an optimized prediction library.
  • We aim to provide a well assorted toolbox of methods that are known to improve prediction speed. People will be able to experiment with it on their own. Good engineering will be essential in this effort.
  • It will support multiple tree formats, so that you can use other packages (e.g. LightGBM) to train your model.
  • The end goal is to maximize prediction throughput (# rows / second) given the latency constraint.

The project is currently private, but we are working hard to make it public in the near future. Stay tuned!

Was this page helpful?
0 / 5 - 0 ratings