You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
The key has expired.
This is a minor release, mainly motivated by issues concerning pip install, e.g. #2426, #3189, #3118, and #3194. With this release, users of Linux and MacOS will be able to run pip install for the most part.
Refactored linear booster class (gblinear), so as to support multiple coordinate descent updaters (#3103, #3134). See BREAKING CHANGES below.
Fix slow training for multiclass classification with high number of classes (#3109)
Fix a corner case in approximate quantile sketch (#3167). Applicable for 'hist' and 'gpu_hist' algorithms
Monotonic constraints for 'hist' algorithm (#3085)
GPU support
Create an abtract 1D vector class that moves data seamlessly between the main and GPU memory (#2935, #3116, #3068). This eliminates unnecessary PCIe data transfer during training time.
Compatibility fixes for latest Spark versions (#3062, #3093)
BREAKING CHANGES: Updated linear modelling algorithms. In particular L1/L2 regularisation penalties are now normalised to number of training examples. This makes the implementation consistent with sklearn/glmnet. L2 regularisation has also been removed from the intercept. To produce linear models with the old regularisation behaviour, the alpha/lambda regularisation parameters can be manually scaled by dividing them by the number of training examples.