python训练保存模型后利用jpmml给Java后端调用预测

1.背景

实际工程开发中，算法同学用python对训练模型后需要给Java后端调用，但一般的joblib.dump保存的模型Java后端无法直接使用，因此借助专门的python模型保存库和对应的jar包，来进行模型的保存和读取。

2.具体实现方式

1.Python训练并保存模型

from sklearn2pmml import PMMLPipeline, sklearn2pmml
from sklearn.ensemble import RandomForestClassifier

model = RandomForestClassifier(n_estimators=195, max_depth=14, max_features=11, oob_score=True,
                                 random_state=123, n_jobs=-1, verbose=0)
pipeline_model = PMMLPipeline([("random_forest", model)])
pipeline_model.fit(train_x, train_y)
sklearn2pmml(pipeline_model, "./model/random_forest.pmml", with_repr=True)

2.Java读取模型

需要用到的依赖

    org.jpmml
    pmml-evaluator
    1.6.4

具体实现调用代码

import org.dmg.pmml.Field;
import org.jpmml.model.*;
import org.jpmml.evaluator.*;

import java.io.BufferedReader;
import java.io.File;
import java.io.FileReader;
import java.io.IOException;
import java.util.*;

public class load_model {

    public Evaluator load_pmml (String pmml_model_path) throws Exception{
        Evaluator evaluator = new LoadingModelEvaluatorBuilder().load(new File(pmml_model_path)).build();
        return evaluator;
    }

    public static Float accuracy(int tp, int tn, int fp, int fn){

        float tp1 = (float) tp;
        float tn1 = (float) tn;
        float fp1 = (float) fp;
        float fn1 = (float) fn;

        return (tp1 + tn1) / (tp1 + tn1 + fp1 + fn1);
    }

    public static Float metric_precision(int tp, int tn, int fp, int fn){

        float tp1 = (float) tp;
        float tn1 = (float) tn;
        float fp1 = (float) fp;
        float fn1 = (float) fn;

        return (tp1) / (tp1 + fp1);
    }

    public static Float metric_recall(int tp, int tn, int fp, int fn){

        float tp1 = (float) tp;
        float tn1 = (float) tn;
        float fp1 = (float) fp;
        float fn1 = (float) fn;

        return (tp1) / (tp1 + fn1);
    }

    public boolean check_input(Evaluator evaluator, Map input_sample){
        List inputFields = evaluator.getInputFields();
        List input_name = new ArrayList();
        int index = 0;
        for (InputField inputField : inputFields){
            input_name.add(inputField.getName().toString());
            boolean flag = input_sample.containsKey(inputField.getName().toString());
            if (!flag){
                index ++;
                System.out.println( index + " 输入数据缺少该特征:" + inputField);
            }
        }
        int redundant = 0;
        for (String i : input_sample.keySet()){
            if (!input_name.contains(i)){
                redundant ++;
                System.out.println(redundant + " 输入特征冗余: " + i );
            }
        }
        if (index != 0 || redundant != 0){
            return false;
        }

        return true;
    }

    public List> load_csv_file(String csv_path, int limit){
        BufferedReader reader = null;
        ArrayList> data = new ArrayList();
        try{
            reader = new BufferedReader(new FileReader(csv_path));
            String[] title = reader.readLine().split(",");
            int count = 1;
            String line = null;

            while ((count <= limit) & ((line = reader.readLine()) != null)){
                String[] item = line.split(",");
                Map temp = new HashMap();
                for (int i = 0; i < item.length; i++){
    //                    System.out.println(title[i] + item[i]);
                    if ((!title[i].equals("d")) & (!title[i].equals("ip"))){
                        temp.put(title[i], Float.parseFloat(item[i]));
                    }
                }
                data.add(temp);
                count += 1;
            }
        } catch (IOException e){
            e.printStackTrace();
        }

        return data;
    }

    public void predict(List> data, Evaluator evaluator){
        Long timestamp1 = System.currentTimeMillis();

        List label = new ArrayList();
        List pre = new ArrayList();
        List