1 /*!
2 * Copyright 2015-2019 by Contributors
3 * \file custom_metric.cc
4 * \brief This is an example to define plugin of xgboost.
5 * This plugin defines the additional metric function.
6 */
7 #include <xgboost/base.h>
8 #include <xgboost/parameter.h>
9 #include <xgboost/objective.h>
10 #include <xgboost/json.h>
11
12 namespace xgboost {
13 namespace obj {
14
15 // This is a helpful data structure to define parameters
16 // You do not have to use it.
17 // see http://dmlc-core.readthedocs.org/en/latest/parameter.html
18 // for introduction of this module.
19 struct MyLogisticParam : public XGBoostParameter<MyLogisticParam> {
20 float scale_neg_weight;
21 // declare parameters
DMLC_DECLARE_PARAMETERxgboost::obj::MyLogisticParam22 DMLC_DECLARE_PARAMETER(MyLogisticParam) {
23 DMLC_DECLARE_FIELD(scale_neg_weight).set_default(1.0f).set_lower_bound(0.0f)
24 .describe("Scale the weight of negative examples by this factor");
25 }
26 };
27
28 DMLC_REGISTER_PARAMETER(MyLogisticParam);
29
30 // Define a customized logistic regression objective in C++.
31 // Implement the interface.
32 class MyLogistic : public ObjFunction {
33 public:
Configure(const std::vector<std::pair<std::string,std::string>> & args)34 void Configure(const std::vector<std::pair<std::string, std::string> >& args) override {
35 param_.UpdateAllowUnknown(args);
36 }
GetGradient(const HostDeviceVector<bst_float> & preds,const MetaInfo & info,int iter,HostDeviceVector<GradientPair> * out_gpair)37 void GetGradient(const HostDeviceVector<bst_float> &preds,
38 const MetaInfo &info,
39 int iter,
40 HostDeviceVector<GradientPair> *out_gpair) override {
41 out_gpair->Resize(preds.Size());
42 const std::vector<bst_float>& preds_h = preds.HostVector();
43 std::vector<GradientPair>& out_gpair_h = out_gpair->HostVector();
44 const std::vector<bst_float>& labels_h = info.labels_.HostVector();
45 for (size_t i = 0; i < preds_h.size(); ++i) {
46 bst_float w = info.GetWeight(i);
47 // scale the negative examples!
48 if (labels_h[i] == 0.0f) w *= param_.scale_neg_weight;
49 // logistic transformation
50 bst_float p = 1.0f / (1.0f + std::exp(-preds_h[i]));
51 // this is the gradient
52 bst_float grad = (p - labels_h[i]) * w;
53 // this is the second order gradient
54 bst_float hess = p * (1.0f - p) * w;
55 out_gpair_h.at(i) = GradientPair(grad, hess);
56 }
57 }
DefaultEvalMetric() const58 const char* DefaultEvalMetric() const override {
59 return "logloss";
60 }
PredTransform(HostDeviceVector<bst_float> * io_preds) const61 void PredTransform(HostDeviceVector<bst_float> *io_preds) const override {
62 // transform margin value to probability.
63 std::vector<bst_float> &preds = io_preds->HostVector();
64 for (auto& pred : preds) {
65 pred = 1.0f / (1.0f + std::exp(-pred));
66 }
67 }
ProbToMargin(bst_float base_score) const68 bst_float ProbToMargin(bst_float base_score) const override {
69 // transform probability to margin value
70 return -std::log(1.0f / base_score - 1.0f);
71 }
72
SaveConfig(Json * p_out) const73 void SaveConfig(Json* p_out) const override {
74 auto& out = *p_out;
75 out["name"] = String("my_logistic");
76 out["my_logistic_param"] = ToJson(param_);
77 }
78
LoadConfig(Json const & in)79 void LoadConfig(Json const& in) override {
80 FromJson(in["my_logistic_param"], ¶m_);
81 }
82
83 private:
84 MyLogisticParam param_;
85 };
86
87 // Finally register the objective function.
88 // After it succeeds you can try use xgboost with objective=mylogistic
89 XGBOOST_REGISTER_OBJECTIVE(MyLogistic, "mylogistic")
90 .describe("User defined logistic regression plugin")
__anon01b959c70102() 91 .set_body([]() { return new MyLogistic(); });
92
93 } // namespace obj
94 } // namespace xgboost
95