public ImageRecord[] update(ImageRecord[] deltas, Integer batchSize, Double learningRate)
throws ScaleNotMatchException {
int imageSize = deltas.length;
ImageRecord[] deltaResult = new ImageRecord[this.inMapSize];
for (int i = 0; i < imageSize; i++)
deltas[i] =
new ImageRecord(
MathUtils.MatrixElementMultiply(
deltas[i].getImage(),
MathUtils.MatrixElementMultiply(
outputs[i].getImage(),
MathUtils.ContantMinusArray(outputs[i].getImage(), 1.0))),
-999);
this.d = deltas; // 保存卷积层的灵敏度
for (int i = 0; i < inMapSize; i++) {
Double[][] image = new Double[inputSize.GetLeftElement()][inputSize.GetRightElement()];
for (int j = 0; j < inputSize.GetLeftElement(); j++)
for (int k = 0; k < inputSize.GetRightElement(); k++) image[j][k] = 0.0;
for (int j = 0; j < imageSize; j++) {
image =
MathUtils.addArrays(
image,
MathUtils.conv(deltas[j].getImage(), kernels[i][j].getRotatedKernel(), "full"));
}
deltaResult[i] = new ImageRecord(image, -999);
}
calculateGradient();
this.batchCount += 1;
// 按batchsize更新权重
if (this.batchCount == batchSize - 1) {
updateWeight(learningRate);
this.batchCount = 0;
}
return deltaResult;
}
public ImageRecord[] calculateOutputs(ImageRecord[] inputs) throws ScaleNotMatchException {
this.inputs = inputs;
for (int i = 0; i < outMapSize; i++) {
Double[][] z =
new Double[outputSize.GetLeftElement()][outputSize.GetRightElement()]; // 初始化输出图像
for (int j = 0; j < inputs.length; j++) {
ImageRecord input = inputs[j];
z =
MathUtils.addArrays(
z, MathUtils.conv(input.getImage(), kernels[j][i].getKernel(), "valid")); // 图像卷积
}
z = MathUtils.AddConstantToArray(z, b[i]);
z = MathUtils.SigmoidForImage(z);
outputs[i] = new ImageRecord(z, -999);
}
return outputs;
}
public ConvLayer(
Pair<Integer, Integer> inputSize,
Integer inMapSize,
Integer outMapSize,
Integer kernelSize,
String activationFunction) {
super();
type = "Convolutional Layer";
this.activationFunction = activationFunction;
this.inputSize = new Pair<Integer, Integer>(inputSize);
this.outputSize =
new Pair<Integer, Integer>(
inputSize.GetLeftElement() - kernelSize + 1,
inputSize.GetRightElement() - kernelSize + 1);
this.inMapSize = inMapSize;
this.kernelSize = kernelSize;
this.outMapSize = outMapSize;
this.fan_in = inMapSize * (kernelSize * kernelSize); // 该卷积层的一个map对应的卷积核权重数
this.fan_out = outMapSize * (kernelSize * kernelSize); // 该卷积层的卷积核权重数
this.outputs = new ImageRecord[outMapSize];
init_Kernels();
init_b();
}
