package edu.hitsz.c102c.cnn;
import java.io.Serializable;
import edu.hitsz.c102c.util.Log;
import edu.hitsz.c102c.util.Util;
/**
* CNN network layer.
*
* @author jiqunpeng
*
*/
public class Layer implements Serializable {
/**
*
*/
private static final long serialVersionUID = 1L;
private LayerType type;
private int outMapNum; ;// output map number
private Size mapSize;// map size
private Size kernelSize; // convolution kernel size, only the convolution layer
private Size scaleSize;// sample size, only the sampling layer
private double[][][][] kernel; // Convolution kernel, only convolutional layer and output layer
private double[] bias; // Each map corresponds to a bias, only the convolutional and output layers have
// Save the output map of each batch, outmaps[0][0] represents the 0th output map of the first record training
private double[][][][] outmaps;
// Residual, corresponding to d in matlab toolbox
private double[][][][] errors;
private static int recordInBatch = 0;
private int classNum = -1; // Category number
private Layer() {
}
/**
* Prepare for the next batch of training·
*/
public static void prepareForNewBatch() {
recordInBatch = 0;
}
/**
* Prepare the next recorded workout.
*/
public static void prepareForNewRecord() {
recordInBatch++;
}
/**
* Build impot layer.
*
* @param mapSize
* @return
*/
public static Layer buildInputLayer(Size mapSize) {
Layer layer = new Layer();
layer.type = LayerType.input;
layer.outMapNum = 1;// ÊäÈë²ãµÄmap¸öÊýΪ1£¬¼´Ò»ÕÅͼ
layer.setMapSize(mapSize);//
return layer;
}
/**
* Build convolution layer.
*
* @return
*/
public static Layer buildConvLayer(int outMapNum, Size kernelSize) {
Layer layer = new Layer();
layer.type = LayerType.conv;
layer.outMapNum = outMapNum;
layer.kernelSize = kernelSize;
return layer;
}
/**
* Build sampling layer.
*
* @param scaleSize
* @return
*/
public static Layer buildSampLayer(Size scaleSize) {
Layer layer = new Layer();
layer.type = LayerType.samp;
layer.scaleSize = scaleSize;
return layer;
}
/**
* Construct the output layer, the number of categories, determine the number of output units according to the number of categories.
* @return
*/
public static Layer buildOutputLayer(int classNum) {
Layer layer = new Layer();
layer.classNum = classNum;
layer.type = LayerType.output;
layer.mapSize = new Size(1, 1);
layer.outMapNum = classNum;
// int outMapNum = 1;
// while ((1 << outMapNum) < classNum)
// outMapNum += 1;
// layer.outMapNum = outMapNum;
Log.i("outMapNum:" + layer.outMapNum);
return layer;
}
/**
* Get map size.
*
* @return
*/
public Size getMapSize() {
return mapSize;
}
/**
* Set map size.
*
* @param mapSize
*/
public void setMapSize(Size mapSize) {
this.mapSize = mapSize;
}
/**
* Get type.
*
* @return
*/
public LayerType getType() {
return type;
}
/**
* Get number of maps.
*
* @return
*/
public int getOutMapNum() {
return outMapNum;
}
/**
* ÉèÖÃÊä³ömapµÄ¸öÊý
*
* @param outMapNum
*/
public void setOutMapNum(int outMapNum) {
this.outMapNum = outMapNum;
}
/**
* Get the size of the convolution kernel, only the kernel layer has kernelSize, and all other layers are null.
*
* @return
*/
public Size getKernelSize() {
return kernelSize;
}
/**
* Get sample size, only sampling layer has scaleSize, other layers are not null.
*
* @return
*/
public Size getScaleSize() {
return scaleSize;
}
enum LayerType {
// The type of network layer: input layer, output layer, convolution layer, sampling layer
input, output, conv, samp
}
/**
* Get sample size, only sampling layer has scaleSize, other layers are not null.
*
* @author jiqunpeng
*
*/
public static class Size implements Serializable {
private static final long serialVersionUID = 1L;
public final int x;
public final int y;
public Size(int x, int y) {
this.x = x;
this.y = y;
}
public String toString() {
StringBuilder s = new StringBuilder("Size(").append(" x = ")
.append(x).append(" y= ").append(y).append(")");
return s.toString();
}
/**
* Dividing scaleSize to get a new Size requires this.x, this.
* y can be divided by scaleSize.x, scaleSize.y
* @param scaleSize
* @return
*/
public Size divide(Size scaleSize) {
int x = this.x / scaleSize.x;
int y = this.y / scaleSize.y;
if (x * scaleSize.x != this.x || y * scaleSize.y != this.y)
throw new RuntimeException(this + "²»ÄÜÕû³ý" + scaleSize);
return new Size(x, y);
}
/**
* ¼õÈ¥size´óС£¬²¢xºÍy·Ö±ð¸½¼ÓÒ»¸öÖµappend
*
* @param size
* @param append
* @return
*/
public Size subtract(Size size, int append) {
int x = this.x - size.x + append;
int y = this.y - size.y + append;
return new Size(x, y);
}
}
/**
* Subtract the size size and append a value append to x and y respectively.
* @param frontMapNum
*/
public void initKernel(int frontMapNum) {
// int fan_out = getOutMapNum() * kernelSize.x * kernelSize.y;
// int fan_in = frontMapNum * kernelSize.x * kernelSize.y;
// double factor = 2 * Math.sqrt(6 / (fan_in + fan_out));
this.kernel = new double[frontMapNum][outMapNum][kernelSize.x][kernelSize.y];
for (int i = 0; i < frontMapNum; i++)
for (int j = 0; j < outMapNum; j++)
kernel[i][j] = Util.randomMatrix(kernelSize.x, kernelSize.y,true);
}
/**
* Randomly initialize the convolution kernel.
*
* @param frontMapNum
* @param size
*/
public void initOutputKerkel(int frontMapNum, Size size) {
kernelSize = size;
// int fan_out = getOutMapNum() * kernelSize.x * kernelSize.y;
// int fan_in = frontMapNum * kernelSize.x * kernelSize.y;
// double factor = 2 * Math.sqrt(6 / (fan_in + fan_out));
this.kernel = new double[frontMapNum][outMapNum][kernelSize.x][kernelSize.y];
for (int i = 0; i < frontMapNum; i++)
for (int j = 0; j < outMapNum; j++)
kernel[i][j] = Util.randomMatrix(kernelSize.x, kernelSize.y,false);
}
/**
* ³õʼ»¯Æ«ÖÃ
*
* @param frontMapNum
*/
public void initBias(int frontMapNum) {
this.bias = Util.randomArray(outMapNum);
}
/**
* ³õʼ»¯Êä³ömap
*
* @param batchSize
*/
public void initOutmaps(int batchSize) {
outmaps = new double[batchSize][outMapNum][mapSize.x][mapSize.y];
}
/**
* ÉèÖÃmapÖµ
*
* @param mapNo
* µÚ¼¸¸ömap
* @param mapX
* mapµÄ¸ß
* @param mapY
* mapµÄ¿í
* @param value
*/
public void setMapValue(int mapNo, int mapX, int mapY, double value) {
outmaps[recordInBatch][mapNo][mapX][mapY] = value;
}
static int count = 0;
/**
* ÒÔ¾ØÕóÐÎʽÉèÖõÚmapNo¸ömapµÄÖµ
*
* @param mapNo
* @param outMatrix
*/
public void setMapValue(int mapNo, double[][] outMatrix) {
// Log.i(type.toString());
// Util.printMatrix(outMatrix);
outmaps[recordInBatch][mapNo] = outMatrix;
}
/**
* Get the index map matrix. In performance considerations, no duplicate objects are returned, but the reference is returned directly, and the caller must be careful.
* Avoid modifying outmaps. If you need to modify, please call setMapValue(...)
* @param index
* @return
*/
public double[][] getMap(int index) {
return outmaps[recordInBatch][index];
}
/**
* Get the convolution kernel of the previous layer i map to the current layer j map.
*
* @param i
* ÉÏÒ»²ãµÄmapϱê
* @param j
* µ±Ç°²ãµÄmapϱê
* @return
*/
public double[][] getKernel(int i, int j) {
return kernel[i][j];
}
/**
*
* @param mapNo
* @param mapX
* @param mapY
* @param value
*/
public void setError(int mapNo, int mapX, int mapY, double value) {
errors[recordInBatch][mapNo][mapX][mapY] = value;
}
/**
*
* @param mapNo
* @param matrix
*/
public void setError(int mapNo, double[][] matrix) {
// Log.i(type.toString());
// Util.printMatrix(matrix);
errors[recordInBatch][mapNo] = matrix;
}
/**
* Get the mapNo map of the residual. Does not return a copy of the object, but directly returns a reference, call the cautious,
* Avoid modifying errors. If you need to modify, please call setError(...)
* @param mapNo
* @return
*/
public double[][] getError(int mapNo) {
return errors[recordInBatch][mapNo];
}
/**
* Get all the residuals (each record and each map).
*
* @return
*/
public double[][][][] getErrors() {
return errors;
}
/**
* Initiazlie errors.
*
* @param batchSize
*/
public void initErros(int batchSize) {
errors = new double[batchSize][outMapNum][mapSize.x][mapSize.y];
}
/**
*
* @param lastMapNo
* @param mapNo
* @param kernel
*/
public void setKernel(int lastMapNo, int mapNo, double[][] kernel) {
this.kernel[lastMapNo][mapNo] = kernel;
}
/**
* Get Bias.
*
* @param mapNo
* @return
*/
public double getBias(int mapNo) {
return bias[mapNo];
}
/**
* Set bias.
*
* @param mapNo
* @param value
*/
public void setBias(int mapNo, double value) {
bias[mapNo] = value;
}
/**
* »ñÈ¡batch¸÷¸ömap¾ØÕó
*
* @return
*/
public double[][][][] getMaps() {
return outmaps;
}
/**
* Get the residual of mapNo in the recordId record.
*
* @param recordId
* @param mapNo
* @return
*/
public double[][] getError(int recordId, int mapNo) {
return errors[recordId][mapNo];
}
/**
* Set map.
*
* @param recordId
* @param mapNo
* @return
*/
public double[][] getMap(int recordId, int mapNo) {
return outmaps[recordId][mapNo];
}
/**
* Get type.
*
* @return
*/
public int getClassNum() {
return classNum;
}
/**
* Get kernel.
*
* @return
*/
public double[][][][] getKernel() {
return kernel;
}
}
Ads help maintain this website.