void NeuralNetwork::backprop(const arma::mat& input, const arma::mat& output)
{
std::vector<arma::mat> gradients;
gradients.push_back(m_activationOnLayer[m_numLayers - 1] - output);
unsigned int prevErrorIndex = 0;
for (int layer = m_numLayers - 2; layer > 0; --layer)
{
arma::mat error;
error = gradients[prevErrorIndex++] * m_theta[layer].cols(1, m_theta[layer].n_cols - 1);
error = error % sigmoidGradient(m_partialOnLayer[layer - 1]);
gradients.push_back(error);
}
int errorIndex = 0;
for (int layer = m_numLayers - 2; layer >= 0; --layer)
{
gradients[errorIndex] = (1.0 / input.n_rows) * (gradients[errorIndex].t() * m_activationOnLayer[layer]);
++errorIndex;
}
std::reverse(gradients.begin(), gradients.end());
for (unsigned int layer = 0; layer < m_numLayers - 1; ++layer)
{
int lastCol = gradients[layer].n_cols - 1;
gradients[layer].cols(1, lastCol) += (m_regFactor / input.n_rows) * m_theta[layer].cols(1, lastCol);
}
//checkGradients(input, output, gradients);
gradientDescent(gradients);
}
int main()
{
populateTrainingSet();
printf("\nEnter the learning rate ");
scanf("%f",&learningRate);
gradientDescent(0,0);
}
int main(int argc, char **argv){
char* filename = argv[1];
/*This is the number of features provided on the command line*/
/*The csv should contain values for each x value and the result, per example*/
int features = atoi(argv[2]);
int examples = atoi(argv[3]); /*Number of training examples provided*/
double cost = DBL_MAX; /*Cost for the current hypothesis, set arbitratily high*/
/*Values for the coefficients in the hypothesis function*/
double *theta = malloc(features * sizeof(double));
double **X = malloc(examples * sizeof(double*));
for(int i = 0; i < examples; i++){
X[i] = malloc(features * sizeof(double));
}
double *Y = malloc(examples * sizeof(double));
parse(features, examples, X, Y, filename);
for(int i = 0; i < features; i++){
theta[i] = 0;
}
theta[0] = 1;
double **meanAndRange = malloc((features - 1) * sizeof(double*));
for(int i = 0; i < features - 1; i++){
meanAndRange[i] = malloc(2 * sizeof(double));
}
/*meanNormalization(X, Y, meanAndRange, features, examples);*/
clock_t begin, end;
begin = clock();
double timeElapsed;
gradientDescent(X, Y, theta, meanAndRange, features, examples);
int *values = malloc((features - 1) * sizeof(int));
values[0] = 1;
char val[5];
/*Print the learned formula*/
printf("Learned function: %f", theta[0]);
for(int i = 1; i < features; i++){
printf(" + %f(x%d)",theta[i], i);
}
printf("\n");
end = clock();
timeElapsed = (double)(end - begin) / CLOCKS_PER_SEC;
printf("Elapsed Time: %f\n", timeElapsed);
/*Obtain experimental values*/
for(int i = 1; i < features ; i++){
printf("Value for x%d:", i);
scanf("%s", val);
values[i] = atoi(val);
}
/*Print out the estimate for given values*/
float output = 0;
for(int i = 0; i < features; i++){
output += values[i] * theta[i];
}
printf("\nOutput: %f\n", output);
}
void gradientDescent(float x0,float x1)
{
float a=x0,b=x1,temp0,temp1;
temp0=x0-(learningRate*der_x0(x0,x1));
temp1=x1-(learningRate*der_x1(x0,x1));
x0=temp0;
x1=temp1;
printf("x0=%f\tx1=%f\n",x0,x1);
getchar();
temp0=a-x0;
temp1=b-x1;
temp0=temp0<0?temp0*-1:temp0;
temp1=temp1<0?temp1*-1:temp1;
if(temp1<.001&&temp0<.001)
{
printf(" %f %f\n ",x0,x1);
exit(0);
}
gradientDescent(x0,x1);
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray * prhs[]) {
/* compute the infimum in CSFM, provided that the W
* (which is 2 x nPoint x nFrame) is centered
*/
double *wOri, *q, *res;
int i, ii, j, k, l, m, n, p, iFrame, jFrame, nFrame, nPoint;
mxArray *pIsDone;
const mwSize *dim_array;
double *errArray, *quaternionArray ;
char *isDone;
mxArray *pWPairwiseProd, *pTmp2By2, *pRTransposeR, *pDRTransposeR,
*pQuaternionTmp, *pGradient, *pWPrimeWPrimeTranspose;
double *wPairwiseProd, *tmp2By2, *RTransposeR, *dRTransposeR,
*quaternionTmp, *gradient, *wPrimeWPrimeTranspose;
/* Retrieve the useful data */
wOri= mxGetPr(prhs[0]);
dim_array = mxGetDimensions(prhs[0]);
nPoint = dim_array[1];
nFrame = dim_array[2];
/* Create the output data */
plhs[0] = mxCreateDoubleMatrix(nFrame, nFrame, mxREAL);
errArray = mxGetPr(plhs[0]);
plhs[1] = mxCreateDoubleMatrix(4, nFrame * nFrame, mxREAL);
quaternionArray = mxGetPr(plhs[1]);
pIsDone = mxCreateLogicalMatrix(nFrame, nFrame);
isDone = (char*) mxGetPr(pIsDone);
for (iFrame = 0, k = 0; iFrame < nFrame; ++iFrame)
for (jFrame = 0; jFrame < nFrame; ++jFrame, ++k) {
errArray[k] = 0;
if (iFrame == jFrame) {
/* set the rotation to identity */
isDone[k] = 1;
quaternionArray[4 * k + 0] = 1;
for (i = 1; i < 4; ++i)
quaternionArray[4 * k + i] = 0;
} else
isDone[k] = 0;
}
/* create the pairwise W multiplications between frames */
pWPairwiseProd = mxCreateDoubleMatrix(2*nFrame, 2*nFrame, mxREAL);
wPairwiseProd = mxGetPr(pWPairwiseProd);
pTmp2By2 = mxCreateDoubleMatrix(2, 2, mxREAL);
tmp2By2 = mxGetPr(pTmp2By2);
for (i = 0; i < nFrame; ++i )
for (j = 0; j <= i; ++j ) {
matrixMultiply(wOri + i * 2 * nPoint, wOri + j * 2 * nPoint, tmp2By2, 2, nPoint, 2);
for( k = 0; k < 2; ++k )
for( l = 0; l < 2; ++l )
wPairwiseProd[(2*i+k)*2*nFrame + 2*j + l] = tmp2By2[l+2*k];
}
/* symmetrize the matrix */
for (i = 0; i < 2*nFrame; ++i )
for (j = i+1; j < 2*nFrame; ++j )
wPairwiseProd[i*2*nFrame+j] = wPairwiseProd[j*2*nFrame+i];
/* create some cache some temporary matrices */
pRTransposeR = mxCreateDoubleMatrix(4, 4, mxREAL);
RTransposeR = mxGetPr(pRTransposeR);
pDRTransposeR = mxCreateDoubleMatrix(4, 4*4, mxREAL);
dRTransposeR = mxGetPr(pDRTransposeR);
pQuaternionTmp = mxCreateDoubleMatrix(4, 1, mxREAL);
quaternionTmp = mxGetPr(pQuaternionTmp);
pGradient = mxCreateDoubleMatrix(4, 1, mxREAL);
gradient = mxGetPr(pGradient);
pWPrimeWPrimeTranspose = mxCreateDoubleMatrix(4, 4, mxREAL);
wPrimeWPrimeTranspose = mxGetPr(pWPrimeWPrimeTranspose);
/* do everything else */
for (iFrame = 0; iFrame < nFrame; ++iFrame) {
for (jFrame = iFrame+1; jFrame < nFrame; ++jFrame)
gradientDescent(wPrimeWPrimeTranspose, iFrame, jFrame, wPairwiseProd, errArray, quaternionArray, isDone, nFrame, nPoint, RTransposeR, dRTransposeR, quaternionTmp, gradient);
++iFrame;
for (jFrame = nFrame - 1; jFrame > iFrame; --jFrame)
gradientDescent(wPrimeWPrimeTranspose, iFrame, jFrame, wPairwiseProd, errArray, quaternionArray, isDone, nFrame, nPoint, RTransposeR, dRTransposeR, quaternionTmp, gradient);
}
/* do everything else in a different order */
for (jFrame = nFrame-1; jFrame > 0; --jFrame) {
for (iFrame = jFrame+1; iFrame < nFrame; ++iFrame)
gradientDescent(wPrimeWPrimeTranspose, iFrame, jFrame, wPairwiseProd, errArray, quaternionArray, isDone, nFrame, nPoint, RTransposeR, dRTransposeR, quaternionTmp, gradient);
--jFrame;
for (iFrame = nFrame - 1; iFrame > jFrame; --iFrame)
gradientDescent(wPrimeWPrimeTranspose, iFrame, jFrame, wPairwiseProd, errArray, quaternionArray, isDone, nFrame, nPoint, RTransposeR, dRTransposeR, quaternionTmp, gradient);
}
/* Free memory */
mxDestroyArray(pIsDone);
mxDestroyArray(pWPrimeWPrimeTranspose);
mxDestroyArray(pTmp2By2);
mxDestroyArray(pWPairwiseProd);
mxDestroyArray(pRTransposeR);
mxDestroyArray(pDRTransposeR);
mxDestroyArray(pQuaternionTmp);
mxDestroyArray(pGradient);
}
void GMExperiment6_1::displayBezierSpline()
{
//Trying a new way of computing inflection points
//First look for long straight bits on the curve and add
//one of the end points to the list of knots
float inflection_thresh = 1.5f;
int32 inflection_filter = 4;
Vector2i inflection_range;
inflection_range[0] = -1; inflection_range[1]= -1;
vector<Vector2i> searchRanges;
vector<int> inflection_idx;
for(int i = 0; i < curvature.size(); ++i)
{
if(abs(curvature[i]) < inflection_thresh)
{ if(inflection_range[0]<0)
{ inflection_range[0]=i;
}
}
else
{
if((inflection_range[0]>=0)) //Infection range was being observed
{
inflection_idx.push_back(inflection_range[0]);
inflection_range[1]=i;
//inflection_idx.push_back(inflection_range[1]);
searchRanges.push_back(inflection_range);
inflection_range[0]=-1;
inflection_range[1]=-1;
}
}
}
//Now add endpoints to the vector and filter out any indices extremely close to each other
//See if the endpoints already exist. If not, then push them onto the data structure
if(inflection_idx.empty())
{
inflection_idx.push_back(0);
inflection_idx.push_back(curvature.size()-1);
}
else
{
if(inflection_idx[0]!=0)
{
inflection_idx.insert( inflection_idx.begin(), 0);
}
if(inflection_idx[inflection_idx.size()-1] != curvature.size()-1)
{
inflection_idx.push_back(curvature.size()-1);
}
}
qDebug()<<"Number of inflection range points "<<inflection_idx.size()<<"\n";
//Then look for sign changes in the segments not spanned by the long straight curves
//and add those and sort the vector
if(!searchRanges.empty())
{
if(searchRanges[0][0]!=0)
{
searchRanges.insert(searchRanges.begin(),makeVector2i(0,0));
}
if(searchRanges[searchRanges.size()-1][1]!=curvature.size()-1)
{
searchRanges.push_back(makeVector2i(curvature.size()-1,curvature.size()-1));
}
for(int i=0; i<searchRanges.size()-1; i++)
{
bool old_sign = curvature[searchRanges[i][1]]>0;
for(int j=searchRanges[i][1]+1; j<searchRanges[i+1][0]; j++)
{
bool new_sign = curvature[j]>0;
if(new_sign!=old_sign)
{
inflection_idx.push_back(j);
}
old_sign=new_sign;
}
}
#if 1
for(int i=0; i<inflection_idx.size()-1; i++)
{
if(i < inflection_idx.size()-2)
{
if(( inflection_idx[i+1]-inflection_idx[i]) < inflection_filter)
{
inflection_idx.erase(inflection_idx.begin()+i+1);
}
}
else
{
if(( inflection_idx[i+1]-inflection_idx[i]) < inflection_filter)
{
inflection_idx.erase(inflection_idx.begin()+i);
}
}
}
#endif
}
sort(inflection_idx.begin(),inflection_idx.end());
vector<int> knots;
knots = inflection_idx;
//knots.push_back(inflection_idx[0]);
//knots.push_back(inflection_idx[inflection_idx.size()-1]);
//for(int i=1; i<inflection_idx.size()-1; i+=2)
//{ knots.push_back(inflection_idx[i]);
//}
//Do the tangent check between inflection points and add to knots
for(int j=0; j<inflection_idx.size()-1; j++)
{
tangent_check(inflection_idx[j],inflection_idx[j+1],knots,smoothData);
}
std::sort(knots.begin(), knots.end());
#if 1
for(int i=0; i<knots.size()-1; i++)
{
if(i < knots.size()-2)
{
if(( knots[i+1]-knots[i]) < knot_filter)
{
knots.erase(knots.begin()+i+1);
}
}
else
{
if(( knots[i+1]-knots[i]) < knot_filter)
{
knots.erase(knots.begin()+i);
}
}
}
#endif
output<<"Total Number of control points (knots):"<< knots.size()<<"\n";
int start = 0;
int end = knots.size()-1;
if(segment_id != -1)
{
start = segment_id;
end = segment_id+1;
}
for(int j = start; j < end; ++j)
{
int i_start = knots[j];
int i_end = knots[j+1];
auto p0 = smoothData[i_start];
auto p3 = smoothData[i_end];
// estimate derivatives
vector<Vector2d> d(2);
d[0] = to(firstDerivative(smoothData, i_start));
d[0].normalize();
d[1] = -to(firstDerivative(smoothData, i_end));
d[1].normalize();
vector<Vector2f> points;
for(int k = knots[j]; k <= knots[j+1]; k++)
points.push_back(smoothData[k]);
// if(points.size() < 6)
// continue;
Vector2d s;
if(least_squares)
{
s = leastSquaresEstimate(points, p0, p3, d[0], d[1]);
}
else
{
s[0] = sqrt((points[1]-points[0]).getSqrNorm());
s[1] = sqrt((points[points.size()-1]-points[points.size()-2]).getSqrNorm());
}
auto ctrl = s2ctrl(to(p0), to(p3), d, s);
if(gradient_descent)
{
auto new_s = gradientDescent(ctrl, d, points, num_iterations);
ctrl = s2ctrl(to(p0), to(p3), d, new_s);
}
// display the curve!
auto piece = renderCubicBezier(ctrl, std::max(points.size(), size_t(20)));
displayCurve(viewer, piece, 2);
// display control points
int pt_indices[4];
if(display_control_points)
{
for(int k = 0; k < 4; ++k)
{
Point2D pt;
pt.position = to(ctrl[k]);
bool is_knot = k == 0 || k == 3;
pt.size = is_knot ? 8 : 6;
pt.color = is_knot ? makeVector4f(0, 1.0, 0, 1) :
makeVector4f(1, 0, 1, 1);
pt_indices[k] = viewer->addPoint(pt);
}
Line l;
l.thickness = 1;
l.color = makeVector4f(1,1,1,1);
l.vertices[0] = pt_indices[0];
l.vertices[1] = pt_indices[1];
viewer->addLine(l);
l.vertices[0] = pt_indices[2];
l.vertices[1] = pt_indices[3];
viewer->addLine(l);
}
}
viewer->refresh();
}
int main(int argc, char **argv){
int numProcs, procId;
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&numProcs);
MPI_Comm_rank(MPI_COMM_WORLD,&procId);
char* filename = argv[1];
/*This is the number of features provided on the command line*/
/*The csv should contain values for each x value and the result, per example*/
int features = atoi(argv[2]);
int examples = atoi(argv[3]); /*Number of training examples provided*/
int itsOver = 0;
double cost = DBL_MAX; /*Cost for the current hypothesis, set arbitratily high*/
/*Values for the coefficients in the hypothesis function*/
double *theta = malloc(features * sizeof(double));
double **X = malloc(examples * sizeof(double*));
for(int i = 0; i < examples; i++){
X[i] = malloc(features * sizeof(double));
}
double *Y = malloc(examples * sizeof(double));
parse(features, examples, X, Y, filename);
for(int i = 0; i < features; i++){
theta[i] = 0;
}
theta[0] = 1;
double **meanAndRange = malloc((features - 1) * sizeof(double*));
for(int i = 0; i < features - 1; i++){
meanAndRange[i] = malloc(2 * sizeof(double));
}
/*meanNormalization(X, Y, meanAndRange, features, examples);*/
double timeElapsed;
if(procId == 0){
timeElapsed = -MPI_Wtime();
}
double converged = gradientDescent(X, Y, theta, meanAndRange, features, examples, numProcs, procId);
if(converged == 0){
itsOver = 1;
for(int i = 0; i < numProcs; i++){
if(i != procId){
MPI_Send(&itsOver, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
}
}
printf("Proc %d learned function: %f", procId, theta[0]);
fflush(stdout);
for(int i = 1; i < features; i++){
printf(" + %f(x%d)",theta[i], i);
}
printf("\n");
fflush(stdout);
}
else{
if(converged == -1){
/*Skip to finalize*/
}
else{
printf("proc %d: %f\n", procId, converged);
MPI_Barrier(MPI_COMM_WORLD);
double min;
MPI_Allreduce(&converged, &min, 1, MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD);
if(converged == min){
/*Print the learned formula*/
printf("Proc %d learned function: %f", procId, theta[0]);
for(int i = 1; i < features; i++){
printf(" + %f(x%d)",theta[i], i);
}
printf("\n");
fflush(stdout);
}
}
}
/*
Obtain experimental values
int *values = malloc((features - 1) * sizeof(int));
values[0] = 1;
char val[5];
for(int i = 1; i < features ; i++){
printf("Value for x%d:", i);
scanf("%s", val);
values[i] = atoi(val);
}
Print out the estimate for given values
float output = 0;
for(int i = 0; i < features; i++){
output += values[i] * theta[i];
}
printf("\nOutput: %f\n", output);
*/
MPI_Barrier(MPI_COMM_WORLD);
if(procId == 0){
timeElapsed += MPI_Wtime();
printf("Elapsed time: %f\n", timeElapsed);
fflush(stdout);
}
MPI_Finalize();
}
