/**
* This file is part of FNLP (formerly FudanNLP).
*
* FNLP is free software: you can redistribute it and/or modify
* it under the terms of the GNU Lesser General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* FNLP is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU Lesser General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with FudanNLP. If not, see <http://www.gnu.org/licenses/>.
*
* Copyright 2009-2014 www.fnlp.org. All rights reserved.
*/
package org.fnlp.ml.classifier.linear;
import java.io.IOException;
import java.util.Arrays;
import java.util.Random;
import org.fnlp.ml.classifier.Predict;
import org.fnlp.ml.classifier.linear.inf.Inferencer;
import org.fnlp.ml.classifier.linear.inf.LinearMax;
import org.fnlp.ml.classifier.linear.update.LinearMaxPAUpdate;
import org.fnlp.ml.classifier.linear.update.Update;
import org.fnlp.ml.feature.Generator;
import org.fnlp.ml.feature.SFGenerator;
import org.fnlp.ml.loss.Loss;
import org.fnlp.ml.loss.ZeroOneLoss;
import org.fnlp.ml.types.Instance;
import org.fnlp.ml.types.InstanceSet;
import org.fnlp.ml.types.alphabet.AlphabetFactory;
import org.fnlp.util.MyArrays;
/**
* 在线参数训练类，
* 可能问题：收敛控制，参数c设置过小，可能会导致“假收敛”的情况 2012.8.6
*
*/
public class OnlineTrainer extends AbstractTrainer {
/**
* 收敛控制，保留最近的错误率个数
*/
private static final int historyNum = 5;
/**
* 收敛控制，最小误差
*/
public static float eps = 1e-10f;
public boolean DEBUG = false;
public boolean shuffle = true;
public boolean finalOptimized = false;
public boolean innerOptimized = false;
public boolean simpleOutput = false;
public boolean interim = false;
public float c=0.1f;
public float threshold = 0.99f;
protected Linear classifier;
protected Inferencer inferencer;
protected Loss loss;
protected Update update;
protected Random random;
public int iternum;
protected float[] weights;
public OnlineTrainer(AlphabetFactory af, int iternum) {
//默认特征生成器
Generator gen = new SFGenerator();
//默认推理器
this.inferencer = new LinearMax(gen, af.getLabelSize());
//默认损失函数
this.loss = new ZeroOneLoss();
//默认参数更新策略
this.update = new LinearMaxPAUpdate(loss);
this.iternum = iternum;
this.c = 0.1f;
weights = (float[]) inferencer.getWeights();
if (weights == null) {
weights = new float[af.getFeatureSize()];
inferencer.setWeights(weights);
}
random = new Random(1l);
}
/**
* 构造函数
* @param af 字典
*/
public OnlineTrainer(AlphabetFactory af) {
this(af,50);
}
/**
* 构造函数
* @param inferencer 推理算法
* @param update 参数更新方法
* @param loss 损失计算方法
* @param fsize 特征数量
* @param iternum 最大迭代次数
* @param c 步长阈值
*/
public OnlineTrainer(Inferencer inferencer, Update update,
Loss loss, int fsize, int iternum, float c) {
this.inferencer = inferencer;
this.update = update;
this.loss = loss;
this.iternum = iternum;
this.c = c;
weights = (float[]) inferencer.getWeights();
if (weights == null) {
weights = new float[fsize];
inferencer.setWeights(weights);
}else if(weights.length<fsize){
weights = Arrays.copyOf(weights, fsize);
inferencer.setWeights(weights);
}
random = new Random(1l);
}
/**
* 构造函数，可根据已训练得到的模型重新开始训练
* @param classifier 分类器
* @param update 参数更新方法
* @param loss 损失计算方法
* @param fsize 特征数量
* @param iternum 最大迭代次数
* @param c 步长阈值
*/
public OnlineTrainer(Linear classifier, Update update, Loss loss, int fsize, int iternum, float c) {
this(classifier.getInferencer(), update, loss, fsize, iternum, c);
}
/**
* 参数训练方法
* @return 线性分类器
*/
@Override
public Linear train(InstanceSet trainset) {
return train(trainset,null);
}
/**
* 参数训练方法
* @return 线性分类器
*/
@Override
public Linear train(InstanceSet trainset, InstanceSet devset) {
int numSamples = trainset.size();
System.out.println("Instance Number: "+numSamples);
float[] hisErrRate = new float[historyNum];
long beginTime, endTime;
long beginTimeIter, endTimeIter;
int iter = 0;
int frac = numSamples / 10;
//平均化感知器需要减去的权重
float[] extraweight = null;
extraweight = new float[weights.length];
beginTime = System.currentTimeMillis();
//遍历的总样本数
int k=0;
while (iter++ < iternum) {
if (!simpleOutput) {
System.out.print("iter "+iter+": ");
}
float err = 0;
float errtot = 0;
int cnt = 0;
int cnttot = 0;
int progress = frac;
if (shuffle)
trainset.shuffle(random);
beginTimeIter = System.currentTimeMillis();
for (int ii = 0; ii < numSamples; ii++) {
k++;
Instance inst = trainset.getInstance(ii);
Predict pred = (Predict) inferencer.getBest(inst,2);
float l = loss.calc(pred.getLabel(0), inst.getTarget());
if (l > 0) {
err += l;
errtot++;
update.update(inst, weights, k, extraweight, pred.getLabel(0), c);
}else{
if (pred.size() > 1)
update.update(inst, weights, k, extraweight, pred.getLabel(1), c);
}
cnt += inst.length();
cnttot++;
if (!simpleOutput && progress != 0 && ii % progress == 0) {
System.out.print('.');
progress += frac;
}
}//end for
float curErrRate = err / cnt;
endTimeIter = System.currentTimeMillis();
if (!simpleOutput) {
System.out.println(" time: " + (endTimeIter - beginTimeIter)
/ 1000.0 + "s");
System.out.print("Train:");
System.out.print(" Tag acc: ");
}
System.out.print(1 - curErrRate);
if (!simpleOutput) {
System.out.print(" Sentence acc: ");
System.out.print(1 - errtot / cnttot);
System.out.println();
}
System.out.print("Weight Numbers: "
+ MyArrays.countNoneZero(weights));
if (innerOptimized) {
int[] idx = MyArrays.getTop(weights.clone(), threshold, false);
MyArrays.set(weights, idx, 0.0f);
System.out.print(" After Optimized: "
+ MyArrays.countNoneZero(weights));
}
System.out.println();
if (devset != null) {
evaluate(devset);
}
System.out.println();
if (interim) {
Linear p = new Linear(inferencer, trainset.getAlphabetFactory());
try {
p.saveTo("tmp.model");
} catch (IOException e) {
System.err.println("write model error!");
}
}
hisErrRate[iter%historyNum] = curErrRate;
if(MyArrays.viarance(hisErrRate) < eps){
System.out.println("convergence!");
break;
}
}// end while 外循环
//平均化参数
for (int i = 0; i < weights.length; i++) {
weights[i] -= extraweight[i]/k;
}
System.out.print("Non-Zero Weight Numbers: " + MyArrays.countNoneZero(weights));
if (finalOptimized) {
int[] idx = MyArrays.getTop(weights.clone(), threshold, false);
MyArrays.set(weights, idx, 0.0f);
System.out.print(" After Optimized: "
+ MyArrays.countNoneZero(weights));
}
System.out.println();
endTime = System.currentTimeMillis();
System.out.println("time escape:" + (endTime - beginTime) / 1000.0
+ "s");
System.out.println();
Linear p = new Linear(inferencer, trainset.getAlphabetFactory());
return p;
}
@Override
public void evaluate(InstanceSet devset) {
float err = 0;
float errtot = 0;
int total = 0;
for (int i = 0; i < devset.size(); i++) {
Instance inst = devset.getInstance(i);
total += inst.length();
Predict pred = (Predict) inferencer.getBest(inst);
float l = loss.calc(pred.getLabel(0), inst.getTarget());
if (l > 0) {
errtot += 1.0;
err += l;
}
}
if (!simpleOutput) {
System.out.print("Test:");
System.out.print(total - err);
System.out.print('/');
System.out.print(total);
System.out.print(" Tag acc:");
} else {
System.out.print(" ");
}
System.out.print(1 - err / total);
if (!simpleOutput) {
System.out.print(" Sentence acc:");
System.out.println(1 - errtot / devset.size());
}
}
}