Neuron::Neuron(int inputNum) : m_inputNum(inputNum)
{
//we need an additional weight for the bias hence the +1
for (int i=0; i<inputNum + 1; i++)
{
//set up the weights with an initial random value
m_weights.push_back(RandomWeight());
}
}
TNNet::TNeuron::TNeuron(TInt OutputsN, TInt MyId, TTFunc TransFunc){
TFuncNm = TransFunc;
for(int c = 0; c < OutputsN; ++c){
// define the edges, 0 element is weight, 1 element is weight delta
TFlt RandWeight = RandomWeight();
OutEdgeV.Add(TIntFltFltTr(c, RandWeight, 0.0));
SumDeltaWeight.Add(0.0);
}
Id = MyId;
}
/**
* Construct an ELM
* @param
* elm_type - 0 for regression; 1 for (both binary and multi-classes) classification
*/
void ELMTrain(int elm_type)
{
double starttime,endtime;
double TrainingAccuracy;
int i,j,k = 0;
float **input,**weight,*biase,**tranpI,**tempH,**H;
float **PIMatrix;
float **train_set;
float **T,**Y;
float **out;
train_set = (float **)calloc(DATASET,sizeof(float *));
tranpI = (float **)calloc(INPUT_NEURONS,sizeof(float *));
input = (float **)calloc(DATASET,sizeof(float *)); /*datasize * INPUT_NEURONS*/
weight = (float **)calloc(HIDDEN_NEURONS,sizeof(float *)); /*HIDDEN_NEURONS * INPUT_NEURONS*/
biase = (float *)calloc(HIDDEN_NEURONS,sizeof(float)); /*HIDDEN_NEURONS*/
tempH = (float **)calloc(HIDDEN_NEURONS,sizeof(float *)); /*HIDDEN_NEURONS * datasize*/
PIMatrix = (float **)calloc(HIDDEN_NEURONS,sizeof(float *));
H = (float **)calloc(DATASET,sizeof(float *));
T = (float **)calloc(DATASET,sizeof(float *));
Y = (float **)calloc(DATASET,sizeof(float *));
out = (float **)calloc(HIDDEN_NEURONS,sizeof(float *));
for(i=0;i<DATASET;i++){
train_set[i] = (float *)calloc(NUMROWS,sizeof(float));
input[i] = (float *)calloc(INPUT_NEURONS,sizeof(float));
H[i] = (float *)calloc(HIDDEN_NEURONS,sizeof(float));
}
for(i=0;i<DATASET;i++)
{
T[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
Y[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
}
for(i=0;i<INPUT_NEURONS;i++)
tranpI[i] = (float *)calloc(DATASET,sizeof(float));
for(i=0;i<HIDDEN_NEURONS;i++)
{
weight[i] = (float *)calloc(INPUT_NEURONS,sizeof(float));
tempH[i] = (float *)calloc(DATASET,sizeof(float));
out[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
PIMatrix[i] = (float *)calloc(DATASET,sizeof(float));
}
printf("begin to random weight and biase...\n");
/*得到随机的偏置和权重*/
RandomWeight(weight,HIDDEN_NEURONS,INPUT_NEURONS);
RandomBiase(biase,HIDDEN_NEURONS);
SaveMatrix(weight,"./result/weight",HIDDEN_NEURONS,INPUT_NEURONS);
SaveMatrix_s(biase,"./result/biase",1,HIDDEN_NEURONS);
/*加载数据集到内存*/
printf("begin to load input from the file...\n");
if(LoadMatrix(train_set,"./sample/frieman",DATASET,NUMROWS) == 0){
printf("load input file error!!!\n");
return;
}
/*将数据集划分成输入和输出*/
if(elm_type == 0){ //regression
/*
the first column is the output of the dataset
*/
for(i = 0;i < DATASET ;i++)
{
T[k++][0] = train_set[i][0];
for(j = 1;j <= INPUT_NEURONS;j++)
{
input[i][j-1] = train_set[i][j];
}
}
}else{ //classification
/*
the last column is the lable of class of the dataset
*/
InitMatrix(T,DATASET,OUTPUT_NEURONS,-1);
for(i = 0;i < DATASET ;i++)
{
//get the last column
T[k++][0] = train_set[i][0] - 1;//class label starts from 0,so minus one
for(j = 1;j <= INPUT_NEURONS;j++)
{
input[i][j-1] = train_set[i][j];
}
}
for(i = 0;i < DATASET ;i++)
{
for(j = 0;j < OUTPUT_NEURONS;j++)
{
k = T[i][0];
if(k < OUTPUT_NEURONS && k >= 0){
T[i][k] = 1;
}
if(k != 0){
T[i][0] = -1;
}
}
}
}
/*ELM*/
printf("begin to compute...\n");
starttime = omp_get_wtime();
TranspositionMatrix(input,tranpI,DATASET,INPUT_NEURONS);
printf("begin to compute step 1...\n");
MultiplyMatrix(weight,HIDDEN_NEURONS,INPUT_NEURONS, tranpI,INPUT_NEURONS,DATASET,tempH);
printf("begin to compute setp 2...\n");
AddMatrix_bais(tempH,biase,HIDDEN_NEURONS,DATASET);
printf("begin to compute step 3...\n");
SigmoidHandle(tempH,HIDDEN_NEURONS,DATASET);
printf("begin to compute step 4...\n");
TranspositionMatrix(tempH,H,HIDDEN_NEURONS,DATASET);
PseudoInverseMatrix(H,DATASET,HIDDEN_NEURONS,PIMatrix);
MultiplyMatrix(PIMatrix,HIDDEN_NEURONS,DATASET,T,DATASET,OUTPUT_NEURONS,out);
//SaveMatrix(H,"./result/H",DATASET,HIDDEN_NEURONS);
//SaveMatrix(PIMatrix,"./result/PIMatrix",HIDDEN_NEURONS,DATASET);
printf("begin to compute step 5...\n");
endtime = omp_get_wtime();
//保存输出权值
SaveMatrix(out,"./result/result",HIDDEN_NEURONS,OUTPUT_NEURONS);
MultiplyMatrix(H,DATASET,HIDDEN_NEURONS,out,HIDDEN_NEURONS,OUTPUT_NEURONS,Y);
printf("use time :%f\n",endtime - starttime);
printf("train complete...\n");
if(elm_type == 0){
//检测准确率
double MSE = 0;
for(i = 0;i< DATASET;i++)
{
MSE += (Y[i][0] - T[i][0])*(Y[i][0] - T[i][0]);
}
TrainingAccuracy = sqrt(MSE/DATASET);
printf("Regression/trainning accuracy :%f\n",TrainingAccuracy);
}else{
float MissClassificationRate_Training=0;
double maxtag1,maxtag2;
int tag1 = 0,tag2 = 0;
for (i = 0; i < DATASET; i++) {
maxtag1 = Y[i][0];
tag1 = 0;
maxtag2 = T[i][0];
tag2 = 0;
for (j = 1; j < OUTPUT_NEURONS; j++) {
if(Y[i][j] > maxtag1){
maxtag1 = Y[i][j];
tag1 = j;
}
if(T[i][j] > maxtag2){
maxtag2 = T[i][j];
tag2 = j;
}
}
if(tag1 != tag2)
MissClassificationRate_Training ++;
}
TrainingAccuracy = 1 - MissClassificationRate_Training*1.0f/DATASET;
printf("Classification/training accuracy :%f\n",TrainingAccuracy);
}
FreeMatrix(train_set,DATASET);
FreeMatrix(tranpI,INPUT_NEURONS);
FreeMatrix(input,DATASET);
FreeMatrix(weight,HIDDEN_NEURONS);
free(biase);
FreeMatrix(tempH,HIDDEN_NEURONS);
FreeMatrix(PIMatrix,HIDDEN_NEURONS);
FreeMatrix(H,DATASET);
FreeMatrix(T,DATASET);
FreeMatrix(Y,DATASET);
FreeMatrix(out,HIDDEN_NEURONS);
}
//回归
void ELMTrain()
{
double starttime,endtime;
int i,j,k = 0;
double MSE = 0,TrainingAccuracy;
float **input,**weight,*biase,**tranpI,**tempH,**H;
float **PIMatrix;
float **train_set;
float **T,**Y;
float **out;
train_set = (float **)calloc(DATASET,sizeof(float *));
tranpI = (float **)calloc(INPUT_NEURONS,sizeof(float *));
input = (float **)calloc(DATASET,sizeof(float *)); /*datasize * INPUT_NEURONS*/
weight = (float **)calloc(HIDDEN_NEURONS,sizeof(float *)); /*HIDDEN_NEURONS * INPUT_NEURONS*/
biase = (float *)calloc(HIDDEN_NEURONS,sizeof(float)); /*HIDDEN_NEURONS*/
tempH = (float **)calloc(HIDDEN_NEURONS,sizeof(float *)); /*HIDDEN_NEURONS * datasize*/
PIMatrix = (float **)calloc(HIDDEN_NEURONS,sizeof(float *));
H = (float **)calloc(DATASET,sizeof(float *));
T = (float **)calloc(DATASET,sizeof(float *));
Y = (float **)calloc(DATASET,sizeof(float *));
out = (float **)calloc(HIDDEN_NEURONS,sizeof(float *));
for(i=0;i<DATASET;i++){
train_set[i] = (float *)calloc(NUMROWS,sizeof(float));
input[i] = (float *)calloc(INPUT_NEURONS,sizeof(float));
H[i] = (float *)calloc(HIDDEN_NEURONS,sizeof(float));
}
for(i=0;i<DATASET;i++)
{
T[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
Y[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
}
for(i=0;i<INPUT_NEURONS;i++)
tranpI[i] = (float *)calloc(DATASET,sizeof(float));
for(i=0;i<HIDDEN_NEURONS;i++)
{
weight[i] = (float *)calloc(INPUT_NEURONS,sizeof(float));
tempH[i] = (float *)calloc(DATASET,sizeof(float));
out[i] = (float *)calloc(OUTPUT_NEURONS,sizeof(float));
PIMatrix[i] = (float *)calloc(DATASET,sizeof(float));
}
printf("begin to random weight and biase...\n");
/*得到随机的偏置和权重*/
RandomWeight(weight,HIDDEN_NEURONS,INPUT_NEURONS);
RandomBiase(biase,HIDDEN_NEURONS);
/*加载数据集到内存*/
printf("begin to load input from the file...\n");
if(LoadMatrix(train_set,"../TrainingDataSet/covtype",DATASET,NUMROWS,1) == 0){
printf("load input file error!!!\n");
return;
}
/*将数据集划分成输入和输出*/
for(i = 0;i < DATASET ;i++)
{
T[k++][0] = train_set[i][0];
for(j = 1;j <= INPUT_NEURONS;j++)
{
input[i][j-1] = train_set[i][j];
}
}
SaveMatrix(input,"./result/input",DATASET,INPUT_NEURONS);
/*ELM*/
printf("begin to compute...\n");
starttime = omp_get_wtime();
TranspositionMatrix(input,tranpI,DATASET,INPUT_NEURONS);
printf("begin to compute step 1...\n");
MultiplyMatrix(weight,HIDDEN_NEURONS,INPUT_NEURONS, tranpI,INPUT_NEURONS,DATASET,tempH);
printf("begin to compute setp 2...\n");
AddMatrix_bais(tempH,biase,HIDDEN_NEURONS,DATASET);
printf("begin to compute step 3...\n");
SigmoidHandle(tempH,HIDDEN_NEURONS,DATASET);
printf("begin to compute step 4...\n");
TranspositionMatrix(tempH,H,HIDDEN_NEURONS,DATASET);
PseudoInverseMatrix(H,DATASET,HIDDEN_NEURONS,PIMatrix);
MultiplyMatrix(PIMatrix,HIDDEN_NEURONS,DATASET,T,DATASET,OUTPUT_NEURONS,out);
//SaveMatrix(H,"./result/H",DATASET,HIDDEN_NEURONS);
//SaveMatrix(PIMatrix,"./result/PIMatrix",HIDDEN_NEURONS,DATASET);
printf("begin to compute step 5...\n");
endtime = omp_get_wtime();
//保存输出权值
SaveMatrix(out,"./result/result",HIDDEN_NEURONS,OUTPUT_NEURONS);
//检测准确率
MultiplyMatrix(H,DATASET,HIDDEN_NEURONS,out,HIDDEN_NEURONS,OUTPUT_NEURONS,Y);
for(i = 0;i< DATASET;i++)
{
MSE += (Y[i][0] - T[i][0])*(Y[i][0] - T[i][0]);
}
SaveMatrix(T,"./result/t",DATASET,OUTPUT_NEURONS);
SaveMatrix(Y,"./result/y",DATASET,OUTPUT_NEURONS);
TrainingAccuracy = sqrt(MSE/DATASET);
printf("use time :%f\n",endtime - starttime);
printf("trainning accuracy :%f\n",TrainingAccuracy);
printf("train complete...\n");
//print(PIMatrix,DATASET,HIDDEN_NEURONS);
}