jrae源代码解析(二)

本文细述上文引出的RAECost和SoftmaxCost两个类。

SoftmaxCost

我们已经知道。SoftmaxCost类在给定features和label的情况下(超參数给定),衡量给定权重(hidden×catSize)的误差值cost,并指出当前的权重梯度。看代码。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
@Override
    public double valueAt(double[]
x)
    {
        if(
!requiresEvaluation(x) )
            return value;
        int numDataItems
= Features.columns;
         
        int[]
requiredRows = ArraysHelper.makeArray(
0,
CatSize-
2);
        ClassifierTheta
Theta =
new ClassifierTheta(x,FeatureLength,CatSize);
        DoubleMatrix
Prediction = getPredictions (Theta, Features);
         
        double MeanTerm
=
1.0 /
(
double)
numDataItems;
        double Cost
= getLoss (Prediction, Labels).sum() * MeanTerm;
        double RegularisationTerm
=
0.5 *
Lambda * DoubleMatrixFunctions.SquaredNorm(Theta.W);
         
        DoubleMatrix
Diff = Prediction.sub(Labels).muli(MeanTerm);
        DoubleMatrix
Delta = Features.mmul(Diff.transpose());
     
        DoubleMatrix
gradW = Delta.getColumns(requiredRows);
        DoubleMatrix
gradb = ((Diff.rowSums()).getRows(requiredRows));
         
        //Regularizing.
Bias does not have one.
        gradW
= gradW.addi(Theta.W.mul(Lambda));
         
        Gradient
=
new ClassifierTheta(gradW,gradb);
        value
= Cost + RegularisationTerm;
        gradient
= Gradient.Theta;
        return value;
    }<br><br>public DoubleMatrix
getPredictions (ClassifierTheta Theta, DoubleMatrix Features)<br>    {<br>        
int numDataItems
= Features.columns;<br>        DoubleMatrix Input = ((Theta.W.transpose()).mmul(Features)).addColumnVector(Theta.b);<br>        Input = DoubleMatrix.concatVertically(Input, DoubleMatrix.zeros(
1,numDataItems));<br>  
     
return Activation.valueAt(Input);
<br>    }

是个典型的2层神经网络,没有隐层,首先依据features预測labels,预測结果用softmax归一化,然后依据误差反向传播算出权重梯度。

此处添加200字。

这个典型的2层神经网络,label为一列向量,目标label置1,其余为0;转换函数为softmax函数,输出为每一个label的概率。

计算cost的函数为getLoss。如果目标label的预測输出为p∗,则每一个样本的cost也即误差函数为:

cost=E(p∗)=−log(p∗)

依据前述的神经网络后向传播算法,我们得到(j为目标label时,否则为0):

∂E∂wij=∂E∂pj∂hj∂netjxi=−1pjpj(1−pj)xi=−(1−pj)xi=−(labelj−pj)featurei

因此我们便理解了以下代码的含义:

1
DoubleMatrix
Delta = Features.mmul(Diff.transpose());

RAECost

先看实现代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
@Override
    public double valueAt(double[]
x)
    {
        if(!requiresEvaluation(x))
            return value;
         
        Theta
Theta1 =
new Theta(x,hiddenSize,visibleSize,dictionaryLength);
        FineTunableTheta
Theta2 =
new FineTunableTheta(x,hiddenSize,visibleSize,catSize,dictionaryLength);
        Theta2.setWe(
Theta2.We.add(WeOrig) );
         
        final RAEClassificationCost
classificationCost =
new RAEClassificationCost(
                catSize,
AlphaCat, Beta, dictionaryLength, hiddenSize, Lambda, f, Theta2);
        final RAEFeatureCost
featureCost =
new RAEFeatureCost(
                AlphaCat,
Beta, dictionaryLength, hiddenSize, Lambda, f, WeOrig, Theta1);
     
        Parallel.For(DataCell,
            new Parallel.Operation<LabeledDatum<Integer,Integer>>()
{
                public void perform(int index,
LabeledDatum<Integer,Integer> Data)
                {
                    try {
                        LabeledRAETree
Tree = featureCost.Compute(Data);
                        classificationCost.Compute(Data,
Tree);                
                    }
catch (Exception
e) {
                        System.err.println(e.getMessage());
                    }
                }
        });
         
        double costRAE
= featureCost.getCost();
        double[]
gradRAE = featureCost.getGradient().clone();
             
        double costSUP
= classificationCost.getCost();
        gradient
= classificationCost.getGradient();
             
        value
= costRAE + costSUP;
        for(int i=0;
i<gradRAE.length; i++)
            gradient[i]
+= gradRAE[i];
         
        System.gc();   
System.gc();
        System.gc();   
System.gc();
        System.gc();   
System.gc();
        System.gc();   
System.gc();
         
        return value;
    }

cost由两部分组成,featureCost和classificationCost。程序遍历每一个样本,用featureCost.Compute(Data)生成一个递归树,同一时候累加cost和gradient。然后用classificationCost.Compute(Data, Tree)依据生成的树计算并累加cost和gradient。因此关键类为RAEFeatureCost和RAEClassificationCost。

RAEFeatureCost类在Compute函数中调用RAEPropagation的ForwardPropagate函数生成一棵树。然后调用BackPropagate计算梯度并累加。详细的算法过程。下一章分解。

上一篇:J2EE的13个规范之(三) Servlet简单介绍


下一篇:install chrome and chrome driver on ubuntu