xgboost-go

XGBoost inference with Golang by means of exporting xgboost model into json format and load model from that json file. This repo only supports DMLC XGBoost model at the moment.

Features

Currently, this repo only supports a few core features such as:

• Read models from json format file (via dump_model API call)
• Support sigmoid and softmax transformation activation.
• Support binary and multiclass predictions.
• Support regressions predictions.
• Support missing values.
• Support libsvm data format.

NOTE: The result from DMLC XGBoost model may slightly differ from this model due to float number precision.

How to use:

To use this repo, first you need to get it:

go get github.com/Elvenson/xgboost-go

Basic example:

package main

import (
	"fmt"

	xgb "github.com/Elvenson/xgboost-go"
	"github.com/Elvenson/xgboost-go/activation"
	"github.com/Elvenson/xgboost-go/mat"
)

func main() {
	ensemble, err := xgb.LoadXGBoostFromJSON("your model path",
		"", 1, 4, &activation.Logistic{})
	if err != nil {
		panic(err)
	}

	input, err := mat.ReadLibsvmFileToSparseMatrix("your libsvm input path")
	if err != nil {
		panic(err)
	}
	predictions, err := ensemble.PredictProba(input)
	if err != nil {
		panic(err)
	}
	fmt.Printf("%+v\n", predictions)
}

Here LoadXGBoostFromJSON requires 5 parameters:

• The json model path.
• DMLC feature map format, if no feature map leave this blank.
• The number of classes (if this is a binary classification, the number of classes should be 1)
• The depth of the tree, if unable to get the tree depth can specify 0 (slightly slower model built time)
• Activation function, for now binary is Logistic multiclass is Softmax and regression is Raw.

For more example, can take a look at xgbensemble_test.go or read this package documentation.

NOTE: This repo only got tested on Python xgboost package version 1.2.0.