进一步梳理理解了一下正向和反向传播。Forward 是利用当前网络对一条数据进行预测的过程,BackPropagation 是根据误差进行网络权重调节的过程。
完整的代码在72天,这里只粘贴Forward和BackPropagation两个方法。
/**
* ***********************************************************
* Forward prediction.
*
* @param paraInput The input data of one instance.
* @return The data at the output end.
* ***********************************************************
*/
public double[] forward(double[] paraInput) {
// Initialize the input layer.
for (int i = 0; i < layerNodeValues[0].length; i++) {
layerNodeValues[0][i] = paraInput[i];
}//of for i
// Calculate the node values of each layer.
double z;
for (int l = 1; l < numLayers; l++) {
for (int j = 0; j < layerNodeValues[l].length; j++) {
// Initialize according to the offset, which is always +1
//(l-1)层的第(该层节点个数)个节点(偏置)指向下一层[j]节点的值;z等于该边的权值。
z = edgeWeights[l - 1][layerNodeValues[l - 1].length][j];//数组edgeWeights的第三维表示下一层节点的索引
// Weighted sum on all edges for this node.
for (int i = 0; i < layerNodeValues[l - 1].length; i++) {
z += edgeWeights[l - 1][i][j] * layerNodeValues[l - 1][i];
}//of for i (循环上一层的节点)
// Sigmoid activation.
// This line should be changed for other activation functions.
layerNodeValues[l][j] = 1 / (1 + Math.exp(-z));
}//of for j(循环要计算的节点的当前层节点)
}//of for l(循环神经网络的层)
return layerNodeValues[numLayers - 1];
}//of forward
/**
* ******************************************************
* Back propagation and change the edge weights.
*
* @param paraTarget For 3-class data, it is [0, 0, 1], [0, 1, 0] or [1, 0, 0].
* ******************************************************
*/
public void backPropagation(double[] paraTarget) {
// Step 1. Initialize the output layer error.
int l = numLayers - 1; //第l层,即输出层
for (int j = 0; j < layerNodeErrors[l].length; j++) {
layerNodeErrors[l][j] = layerNodeValues[l][j] * (1 - layerNodeValues[l][j])
* (paraTarget[j] - layerNodeValues[l][j]);
//套用输出层误差公式。上游传过来的值(即误差)乘以激活函数的倒数,这里sigmod的倒数为y(1-y)。)
}//of for j
// Step 2. Back-propagation even for l == 0
while (l > 0) {
l--;
// Layer l, for each node.
for (int j = 0; j < layerNumNodes[l]; j++) {
double z = 0.0;
// For each node of the next layer.
for (int i = 0; i < layerNumNodes[l + 1]; i++) {
if (l > 0) {
z += layerNodeErrors[l + 1][i] * edgeWeights[l][j][i];
//(l+1)层的第i个节点,乘以l层第j个节点指向下一层第i个节点的边的权重。
}//of if
// Weight adjusting.
edgeWeightsDelta[l][j][i] = mobp * edgeWeightsDelta[l][j][i] +
learningRate * layerNodeErrors[l + 1][i] * layerNodeValues[l][j];
edgeWeights[l][j][i] += edgeWeightsDelta[l][j][i];
if (j == layerNumNodes[l] - 1) {
// Weight adjusting for the offset part.
//偏置节点没包含在每层节点个数里,所以要加1.
edgeWeightsDelta[l][j + 1][i] = mobp * edgeWeightsDelta[l][j + 1][i]
+ learningRate * layerNodeErrors[l + 1][i];
edgeWeights[l][j + 1][i] += edgeWeightsDelta[l][j + 1][i];
}//of if
}//of for i
// Record the error according to the differential of Sigmoid.
// This line should be changed for other activation functions.
layerNodeErrors[l][j] = layerNodeValues[l][j] * (1 - layerNodeValues[l][j]) * z;
}//of for j
}//of while
}//of backPropagation
edgeWeights与edgeWeightsDelta两个三维数组,再声明的时候第二维大小就是“layerNumNodes[l] + 1”。所以,偏置节点没包含在layerNumNodes[l] 里。因此在偏置调整时第二维的下标是j+1。