static void softmaxGradientForWholeBatch(
DataSet training_set,
FeatureType* gradient) {
memset(gradient, 0, training_set.num_features * sizeof(FeatureType) * LABEL_CLASS);
// computes softmax function for each data point in the training set
std::vector<float> probabilities_of_each;
size_t idx = 0;
for (size_t i = 0; i < training_set.num_data_points; i++) {
idx = i * training_set.num_features;
//http://stackoverflow.com/questions/3177241/what-is-the-best-way-to-concatenate-two-vectors
// concatenate-two-vectors to make probability vector
probabilities_of_each+=softmaxFunctionFloat(
training_set.parameter_vector,
&training_set.data_points[idx],
training_set.num_features);
}
//change a vector to an array
//http://stackoverflow.com/questions/2923272/how-to-convert-vector-to-array-c
float* probabilities_array = &probabilities_of_each[0];
//initialize to 0 by c++
std::vector<float> groundTruth(LABEL_CLASS * training_set.num_data_points);
//establish groundtruth function to see softmaxExercise at UFLDL Tutorial
for (size_t i = 0; i < training_set.num_data_points; i++) {
int idx1 = training_set.labels[i];
groundTruth[idx1 + i * LABEL_CLASS]=1.0f;
}
//change a vector to an array
float* groundTruth_array = &groundTruth[0];
//computing 1{P(y=k|x,theta)-1{y=k}}
addVectors(probabilities_array,
groundTruth_array,
training_set.num_data_points * LABEL_CLASS,
-1);
float factor = 1.0f / training_set.num_data_points;
//finish computing x*sigma (i=1..k){P(y=k|x,theta)-1{y=k}}
matrixMatrixMultiply(probabilities_array,
training_set.data_points,
factor,
training_set.num_data_points,
training_set.num_features,
gradient);
}
char checkLU(){
/*
Tests LU factorization.
Parameters:
int n Matrix dimension.
double tolMax Tolerance for max difference.
Output:
char pass (returned) Whether all norms are under supplied tolerances.
*/
int j;
int n = 100 ;
double tolMax = 1e-10 ;
char pass = 1;
double cond[3] = {10.0, 1.0 / sqrt(1e-14), 1.0 / 1e-14} ;
int numConds = 3;
int type = 4;
double startTime, endTime;
matrix A = allocAndInitRandomizedMatrix(n,n) ;
matrix L = allocAndInitZeroMatrix(n, n) ;
matrix U = allocAndInitZeroMatrix(n, n) ;
matrix product = allocAndInitZeroMatrix(n,n) ;
for(j=0; j<numConds; j++){
printf("Tests for cond = %e :\n\n", cond[j]) ;
for(type = 0; type < 5; type++){
getCondNumberMatrix(&A, cond[j], type) ;
// perform the matrix factorization
startTime = read_timer();
LUFactorize(A, &L, &U) ;
endTime = read_timer();
printf("Elapsed time = %f seconds.\n", endTime - startTime);
matrixMatrixMultiply(L, U, &product) ;
char output = 0;
if(output){
printf("Input matrix:\n");
printMatrix(A);
printf("U = \n") ;
printMatrix(U);
printf("L = \n") ;
printMatrix(L);
printMatrix(product) ;
}
double maxDiff = maxDiffMatrix(product, A) ;
printf("maxDiff = %e\n", maxDiff) ;
if(maxDiff < tolMax){
printf("LU factorization test passed on matrix of type %d.\n\n", type) ;
}
else{
printf("LU factorization test failed on matrix of type %d.\n\n", type) ;
pass = 0;
}
}
}
if(pass)
printf("LU factorization test passed.\n\n") ;
else
printf("LU factorization test failed.\n\n") ;
// free resources
freeMatrix(A) ;
freeMatrix(L) ;
freeMatrix(U) ;
freeMatrix(product) ;
return pass;
}
