void PanoramaTracker::run() {
while (isRunning() && m_scaled.size() <= MAX_TRACKER_FRAMES) {
QScopedPointer<QtCamGstSample> sample(m_input->sample());
if (!sample) {
continue;
}
if (!Tracker::isInitialized()) {
QSize size = QSize(sample->width(), sample->height());
int m_width = size.width() > 720 ? size.width() / 8 : size.width() / 4;
int m_height = size.width() > 720 ? size.height() / 8 : size.height() / 4;
m_inputSize = size;
// TODO: This should be 5.0 but we fail to stitch sometimes if we set it to 5
if (!Tracker::initialize(m_width, m_height, 2.0f)) {
emit error(Panorama::ErrorTrackerInit);
return;
}
}
// Now we can process the sample:
const guint8 *src = sample->data();
QScopedArrayPointer<guint8>
dst(new guint8[m_inputSize.width() * m_inputSize.height() * 3 / 2]);
enum libyuv::FourCC fmt;
switch (sample->format()) {
case GST_VIDEO_FORMAT_UYVY:
fmt = libyuv::FOURCC_UYVY;
break;
default:
qCritical() << "Unsupported color format";
emit error(Panorama::ErrorTrackerFormat);
return;
}
guint8 *y = dst.data(),
*u = y + m_inputSize.width() * m_inputSize.height(),
*v = u + m_inputSize.width()/2 * m_inputSize.height()/2;
if (ConvertToI420(src, sample->size(),
y, m_inputSize.width(),
u, m_inputSize.width() / 2,
v, m_inputSize.width() / 2,
0, 0,
m_inputSize.width(), m_inputSize.height(),
m_inputSize.width(), m_inputSize.height(),
libyuv::kRotate0, fmt) != 0) {
emit error(Panorama::ErrorTrackerConvert);
return;
}
QScopedArrayPointer<guint8> scaled(new guint8[m_width * m_height * 3 / 2]);
guint8 *ys = scaled.data(),
*us = ys + m_width * m_height,
*vs = us + m_width/2 * m_height/2;
// Now scale:
// No need for error checking because the function always returns 0
libyuv::I420Scale(y, m_inputSize.width(),
u, m_inputSize.width()/2,
v, m_inputSize.width()/2,
m_inputSize.width(), m_inputSize.height(),
ys, m_width,
us, m_width/2,
vs, m_width/2,
m_width, m_height,
libyuv::kFilterBilinear);
int err = addFrame(scaled.data());
if (err >= 0) {
m_scaled.push_back(scaled.take());
m_frames.push_back(dst.take());
emit frameCountChanged();
}
}
}
kernel void sample_z(global int *cur_y,
global int *cur_z,
global int *cur_r,
global int *z_by_ry,
global int *z_col_sum,
global int *obs,
global float *rand,
uint N, uint D, uint K, uint f_img_width,
float lambda, float epislon, float theta) {
const uint V_SCALE = 0, H_SCALE = 1, V_TRANS = 2, H_TRANS = 3, NUM_TRANS = 4;
uint h, w, new_index; // variables used in the for loop
uint nth = get_global_id(0); // n is the index of data
uint kth = get_global_id(1); // k is the index of features
uint f_img_height = D / f_img_width;
// calculate the prior probability of each cell is 1
float on_prob_temp = (z_col_sum[kth] - cur_z[nth * K + kth]) / (float)N;
float off_prob_temp = 1 - (z_col_sum[kth] - cur_z[nth * K + kth]) / (float)N;
// retrieve the transformation applied to this feature by this object
int v_scale = cur_r[nth * (K * NUM_TRANS) + kth * NUM_TRANS + V_SCALE];
int h_scale = cur_r[nth * (K * NUM_TRANS) + kth * NUM_TRANS + H_SCALE];
int v_dist = cur_r[nth * (K * NUM_TRANS) + kth * NUM_TRANS + V_TRANS];
int h_dist = cur_r[nth * (K * NUM_TRANS) + kth * NUM_TRANS + H_TRANS];
int new_height = f_img_height + v_scale, new_width = f_img_width + h_scale;
uint d, hh, ww;
// extremely hackish way to calculate the likelihood
for (d = 0; d < D; d++) {
// if the kth feature can turn on a pixel at d
if (cur_y[kth * D + d] == 1) {
// unpack d into h and w and get new index
h = d / f_img_width;
w = d % f_img_width;
for (hh = 0; hh < f_img_height; hh++) {
for (ww = 0; ww < f_img_width; ww++) {
if ((int)round((float)hh / new_height * f_img_height) == h &
(int)round((float)ww / new_width * f_img_width) == w) {
new_index = ((v_dist + hh) % f_img_height) * f_img_width + (h_dist + ww) % f_img_width;
// then the corresponding observed pixel is at new_index
// so, if the observed pixel at new_index is on
if (obs[nth * D + new_index] == 1) {
// if the nth object previously has the kth feature
if (cur_z[nth * K + kth] == 1) {
on_prob_temp *= 1 - pow(1 - lambda, z_by_ry[nth * D + new_index]) * (1 - epislon);
off_prob_temp *= 1 - pow(1 - lambda, z_by_ry[nth * D + new_index] - 1) * (1 - epislon);
} else {
on_prob_temp *= 1 - pow(1 - lambda, z_by_ry[nth * D + new_index] + 1) * (1 - epislon);
off_prob_temp *= 1 - pow(1 - lambda, z_by_ry[nth * D + new_index]) * (1 - epislon);
}
} else {
on_prob_temp *= 1 - lambda;
off_prob_temp *= 1.0f;
}
}
}
}
}
}
//printf("index: %d post_on: %f post_off: %f\n", nth * K + kth, on_prob_temp, off_prob_temp);
float post[2] = {on_prob_temp, off_prob_temp};
uint labels[2] = {1, 0};
pnormalize(post, 0, 2);
//printf("before index: %d %f %f %d \n", nth * K + kth, post[0], post[1], cur_z[nth * K + kth]);
cur_z[nth * K + kth] = sample(2, labels, post, 0, rand[nth * K + kth]);
//printf("after index: %d %f %f %d \n", nth * K + kth, post[0], post[1], cur_z[nth * K + kth]);
}
void Mesh2Cloud::addProperties() {
auto group = gui()->properties()->add<Section>("Sampling", "group");
auto samples = group->add<Number>("Samples Per Square Unit", "samples");
samples->setDigits(0);
samples->setMin(1);
samples->setMax(100000);
samples->setValue(100);
group->add<Button>("Sample", "sample")->setCallback([&] () { auto ns = gui()->properties()->get<Number>({"group", "samples"})->value(); sample(ns); });;
auto iogroup = gui()->properties()->add<Section>("Input/Output", "iogroup");
auto outFile = iogroup->add<File>("Save to: ", "outFile");
outFile->setMode(File::SAVE);
outFile->setCallback([&] (fs::path p) {
pcl::io::savePCDFileBinary(p.string(), *m_cloud);
gui()->log()->info("Saved pointcloud to: \""+p.string()+"\"");
});
outFile->disable();
}
nsresult
EMEH264Decoder::GmpInput(MP4Sample* aSample)
{
MOZ_ASSERT(IsOnGMPThread());
nsAutoPtr<MP4Sample> sample(aSample);
if (!mGMP) {
mCallback->Error();
return NS_ERROR_FAILURE;
}
if (sample->crypto.valid) {
CDMCaps::AutoLock caps(mProxy->Capabilites());
MOZ_ASSERT(caps.CanDecryptAndDecodeVideo());
const auto& keyid = sample->crypto.key;
if (!caps.IsKeyUsable(keyid)) {
nsRefPtr<nsIRunnable> task(new DeliverSample(this, sample.forget()));
caps.CallWhenKeyUsable(keyid, task, mGMPThread);
return NS_OK;
}
}
mLastStreamOffset = sample->byte_offset;
GMPVideoFrame* ftmp = nullptr;
GMPErr err = mHost->CreateFrame(kGMPEncodedVideoFrame, &ftmp);
if (GMP_FAILED(err)) {
mCallback->Error();
return NS_ERROR_FAILURE;
}
gmp::GMPVideoEncodedFrameImpl* frame = static_cast<gmp::GMPVideoEncodedFrameImpl*>(ftmp);
err = frame->CreateEmptyFrame(sample->size);
if (GMP_FAILED(err)) {
mCallback->Error();
return NS_ERROR_FAILURE;
}
memcpy(frame->Buffer(), sample->data, frame->Size());
frame->SetEncodedWidth(mConfig.display_width);
frame->SetEncodedHeight(mConfig.display_height);
frame->SetTimeStamp(sample->composition_timestamp);
frame->SetCompleteFrame(true);
frame->SetDuration(sample->duration);
if (sample->crypto.valid) {
frame->InitCrypto(sample->crypto);
}
frame->SetFrameType(sample->is_sync_point ? kGMPKeyFrame : kGMPDeltaFrame);
frame->SetBufferType(GMP_BufferLength32);
nsTArray<uint8_t> info; // No codec specific per-frame info to pass.
nsresult rv = mGMP->Decode(frame, false, info, 0);
if (NS_FAILED(rv)) {
mCallback->Error();
return rv;
}
return NS_OK;
}
// Sample a hidden neuron, given a visible neuron vector
double sample_hidden(rbm* r, unsigned int j, std::vector<int> visible) {
return sample(hidden_probability(r, j, visible));
}
void Shape::sample(const core::Vec3 &ps, float u1, float u2, float u3, int *primID, core::Vec3 *p, core::Vec3 *n) const
{
return sample(u1, u2, u3, primID, p, n);
}
bool ClassificationData::loadDatasetFromFile(const std::string &filename){
std::fstream file;
file.open(filename.c_str(), std::ios::in);
UINT numClasses = 0;
clear();
if( !file.is_open() ){
errorLog << "loadDatasetFromFile(const std::string &filename) - could not open file!" << std::endl;
return false;
}
std::string word;
//Check to make sure this is a file with the Training File Format
file >> word;
if(word != "GRT_LABELLED_CLASSIFICATION_DATA_FILE_V1.0"){
errorLog << "loadDatasetFromFile(const std::string &filename) - could not find file header!" << std::endl;
file.close();
return false;
}
//Get the name of the dataset
file >> word;
if(word != "DatasetName:"){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find DatasetName header!" << std::endl;
errorLog << word << std::endl;
file.close();
return false;
}
file >> datasetName;
file >> word;
if(word != "InfoText:"){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find InfoText header!" << std::endl;
file.close();
return false;
}
//Load the info text
file >> word;
infoText = "";
while( word != "NumDimensions:" ){
infoText += word + " ";
file >> word;
}
//Get the number of dimensions in the training data
if( word != "NumDimensions:" ){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find NumDimensions header!" << std::endl;
file.close();
return false;
}
file >> numDimensions;
//Get the total number of training examples in the training data
file >> word;
if( word != "TotalNumTrainingExamples:" && word != "TotalNumExamples:" ){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find TotalNumTrainingExamples header!" << std::endl;
file.close();
return false;
}
file >> totalNumSamples;
//Get the total number of classes in the training data
file >> word;
if(word != "NumberOfClasses:"){
errorLog << "loadDatasetFromFile(string filename) - failed to find NumberOfClasses header!" << std::endl;
file.close();
return false;
}
file >> numClasses;
//Resize the class counter buffer and load the counters
classTracker.resize(numClasses);
//Get the total number of classes in the training data
file >> word;
if(word != "ClassIDsAndCounters:"){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find ClassIDsAndCounters header!" << std::endl;
file.close();
return false;
}
for(UINT i=0; i<classTracker.getSize(); i++){
file >> classTracker[i].classLabel;
file >> classTracker[i].counter;
file >> classTracker[i].className;
}
//Check if the dataset should be scaled using external ranges
file >> word;
if(word != "UseExternalRanges:"){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find UseExternalRanges header!" << std::endl;
file.close();
return false;
}
file >> useExternalRanges;
//If we are using external ranges then load them
if( useExternalRanges ){
externalRanges.resize(numDimensions);
for(UINT i=0; i<externalRanges.getSize(); i++){
file >> externalRanges[i].minValue;
file >> externalRanges[i].maxValue;
}
}
//Get the main training data
file >> word;
if( word != "LabelledTrainingData:" && word != "Data:"){
errorLog << "loadDatasetFromFile(const std::string &filename) - failed to find LabelledTrainingData header!" << std::endl;
file.close();
return false;
}
ClassificationSample tempSample( numDimensions );
data.resize( totalNumSamples, tempSample );
for(UINT i=0; i<totalNumSamples; i++){
UINT classLabel = 0;
VectorFloat sample(numDimensions,0);
file >> classLabel;
for(UINT j=0; j<numDimensions; j++){
file >> sample[j];
}
data[i].set(classLabel, sample);
}
file.close();
//Sort the class labels
sortClassLabels();
return true;
}
int main() {
srand(time(NULL));
sample(10000000, 1000000);
}
virtual void process()
{
ActPrintLog("GlobalMergeActivityMaster::process");
CMasterActivity::process();
IHThorMergeArg *helper = (IHThorMergeArg *)queryHelper();
Owned<IThorRowInterfaces> rowif = createRowInterfaces(helper->queryOutputMeta());
CThorKeyArray sample(*this, rowif,helper->querySerialize(),helper->queryCompare(),helper->queryCompareKey(),helper->queryCompareRowKey());
unsigned n = container.queryJob().querySlaves();
mptag_t *replytags = new mptag_t[n];
mptag_t *intertags = new mptag_t[n];
unsigned i;
for (i=0;i<n;i++) {
replytags[i] = TAG_NULL;
intertags[i] = TAG_NULL;
}
try {
for (i=0;i<n;i++) {
if (abortSoon)
return;
CMessageBuffer mb;
#ifdef _TRACE
ActPrintLog("Merge process, Receiving on tag %d",replyTag);
#endif
rank_t sender;
if (!receiveMsg(mb, RANK_ALL, replyTag, &sender)||abortSoon)
return;
#ifdef _TRACE
ActPrintLog("Merge process, Received sample from %d",sender);
#endif
sender--;
assertex((unsigned)sender<n);
assertex(replytags[(unsigned)sender]==TAG_NULL);
deserializeMPtag(mb,replytags[(unsigned)sender]);
deserializeMPtag(mb,intertags[(unsigned)sender]);
sample.deserialize(mb,true);
}
ActPrintLog("GlobalMergeActivityMaster::process samples merged");
sample.createSortedPartition(n);
ActPrintLog("GlobalMergeActivityMaster::process partition generated");
for (i=0;i<n;i++) {
if (abortSoon)
break;
CMessageBuffer mb;
mb.append(n);
for (unsigned j = 0;j<n;j++)
serializeMPtag(mb,intertags[j]);
sample.serialize(mb);
#ifdef _TRACE
ActPrintLog("Merge process, Replying to node %d tag %d",i+1,replytags[i]);
#endif
if (!queryJobChannel().queryJobComm().send(mb, (rank_t)i+1, replytags[i]))
break;
}
}
catch (IException *e) {
delete [] replytags;
delete [] intertags;
ActPrintLog(e, "MERGE");
throw;
}
delete [] replytags;
delete [] intertags;
ActPrintLog("GlobalMergeActivityMaster::process exit");
}
double UniformSample(double max) {
boost::uniform_real<> dist(0, max);
boost::variate_generator<boost::mt19937&, boost::uniform_real<> > sample(
gen, dist);
return sample();
}
double DeltaStarGibbs(const vector<int> &oldClique,
const vector<vector<int> > &oldComponents,
int Q, int G, const int *S,
double *Delta,const double *r,
const double *sigma2,const double *phi,
const double *tau2R,const double *b,
const double *nu,const int *delta,
const int *psi,const double *x,
const vector<vector<vector<double> > > &Omega,
Random &ran,int draw) {
// compute prior presicion matrix
// cout << "start computing prior precision matrix" << endl;
vector<vector<vector<double> > > OmegaInv(Omega);
vector<double> OmegaDet(Omega.size(),0.0);
int k;
for (k = 0; k < OmegaInv.size(); k++)
OmegaDet[k] = inverse(Omega[k],OmegaInv[k]);
vector<vector<vector<double> > > OmegaSep;
OmegaSep.resize(Omega.size());
for (k = 0; k < OmegaSep.size(); k++) {
OmegaSep[k].resize(oldComponents[k].size());
int i;
for (i = 0; i < oldComponents[k].size(); i++) {
OmegaSep[k][i].resize(oldComponents[k].size());
int j;
for (j = 0; j < oldComponents[k].size(); j++) {
OmegaSep[k][i][j] = Omega[oldClique[k]][oldComponents[k][i]][oldComponents[k][j]];
}
}
}
vector<vector<vector<double> > > OmegaSepInv(OmegaSep);
vector<double> OmegaSepDet(Omega.size(),0.0);
for (k = 0; k < OmegaSep.size(); k++) {
if (OmegaSep[k].size() > 0)
OmegaSepDet[k] = inverse(OmegaSep[k],OmegaSepInv[k]);
}
vector<map<int,double> > OmegaInvSparse;
OmegaInvSparse.resize(G);
int g;
for (g = 0; g < G; g++)
OmegaInvSparse[g].clear();
vector<vector<int> > nr;
nr.resize(Omega.size());
g = 0;
for (k = 0; k < Omega.size(); k++) {
nr[k].resize(Omega[k].size());
int gg;
for (gg = 0; gg < oldComponents[k].size(); gg++)
nr[k][gg] = nr[oldClique[k]][oldComponents[k][gg]];
for (gg = oldComponents[k].size(); gg < Omega[k].size(); gg++) {
nr[k][gg] = g;
g++;
}
}
// print to files
/*
for (k = 0; k < Omega.size(); k++) {
char filename[120];
sprintf(filename,"Omega-%d.txt",k);
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < Omega[k].size(); i++) {
for (j = 0; j < Omega[k][i].size(); j++)
fprintf(out,"%20.18e ",Omega[k][i][j]);
fprintf(out,"\n");
}
fclose(out);
}
for (k = 0; k < OmegaInv.size(); k++) {
char filename[120];
sprintf(filename,"OmegaInv-%d.txt",k);
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < OmegaInv[k].size(); i++) {
for (j = 0; j < OmegaInv[k][i].size(); j++)
fprintf(out,"%20.18e ",OmegaInv[k][i][j]);
fprintf(out,"\n");
}
fclose(out);
}
for (k = 1; k < OmegaSep.size(); k++) {
char filename[120];
sprintf(filename,"OmegaSep-%d.txt",k);
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < OmegaSep[k].size(); i++) {
for (j = 0; j < OmegaSep[k][i].size(); j++)
fprintf(out,"%20.18e ",OmegaSep[k][i][j]);
fprintf(out,"\n");
}
fclose(out);
}
for (k = 1; k < OmegaSepInv.size(); k++) {
char filename[120];
sprintf(filename,"OmegaSepInv-%d.txt",k);
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < OmegaSepInv[k].size(); i++) {
for (j = 0; j < OmegaSepInv[k][i].size(); j++)
fprintf(out,"%20.18e ",OmegaSepInv[k][i][j]);
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing to files
// establish and print large Omega
/*
vector<vector<double> > OmegaTotal;
OmegaTotal.resize(G);
for (g = 0; g < G; g++) {
OmegaTotal[g].resize(G);
int gg;
for (gg = 0; gg < G; gg++)
OmegaTotal[g][gg] = 0.0;
}
for (k = 0; k < OmegaInv.size(); k++) {
int r;
for (r = 0; r < nr[k].size(); r++) {
int g = nr[k][r];
int s;
for (s = 0; s <= r; s++) {
int gg = nr[k][s];
double value = Omega[k][r][s];
OmegaTotal[g][gg] = value;
OmegaTotal[gg][g] = value;
}
}
}
{
char filename[120];
sprintf(filename,"OmegaTotal.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < OmegaTotal.size(); i++) {
for (j = 0; j < OmegaTotal[i].size(); j++)
fprintf(out,"%20.18e ",OmegaTotal[i][j]);
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing files
for (k = 0; k < OmegaInv.size(); k++) {
int r;
for (r = 0; r < nr[k].size(); r++) {
int s;
for (s = 0; s < nr[k].size(); s++) {
int g = nr[k][r];
int gg = nr[k][s];
double value = OmegaInv[k][r][s];
map<int,double>::iterator it;
it = OmegaInvSparse[g].find(gg);
if (it == OmegaInvSparse[g].end())
OmegaInvSparse[g].insert(pair<int,double>(gg,value));
else {
OmegaInvSparse[g][gg] += value;
}
}
}
}
for (k = 0; k < OmegaSepInv.size(); k++) {
int r;
for (r = 0; r < oldComponents[k].size(); r++) {
int s;
for (s = 0; s < oldComponents[k].size(); s++) {
int g = nr[k][r];
int gg = nr[k][s];
double value = OmegaSepInv[k][r][s];
OmegaInvSparse[g][gg] -= value;
}
}
}
// print OmegaInv to files
/*
{
char filename[120];
sprintf(filename,"OmegaTotalInv.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < OmegaTotal.size(); i++) {
for (j = 0; j < OmegaTotal[i].size(); j++) {
double value = 0.0;
if (OmegaInvSparse[i].find(j) != OmegaInvSparse[i].end())
value = OmegaInvSparse[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// establish covariance matrix R and its inverse
vector<vector<double> > R;
R.resize(Q);
int p;
for (p = 0; p < Q; p++) {
R[p].resize(Q);
}
int q;
for (p = 0; p < Q; p++) {
R[p][p] = tau2R[p];
for (q = p + 1; q < Q; q++) {
R[p][q] = sqrt(tau2R[p] * tau2R[q]) * r[qq2index(p,q,Q)];
R[q][p] = R[p][q];
}
}
vector<vector<double> > RInverse;
inverse(R,RInverse);
// print RInverse to file
/*
{
char filename[120];
sprintf(filename,"RInverse.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < RInverse.size(); i++) {
for (j = 0; j < RInverse[i].size(); j++)
fprintf(out,"%20.18e ",RInverse[i][j]);
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing RInverse
// compute precision matrix of full conditional
// cout << "start computing precision matrix of full conditional" << endl;
vector<map<int,double> > VinvSparse;
VinvSparse.resize(G*Q);
for (k = 0; k < G*Q; k++)
VinvSparse[k].clear();
for (g = 0; g < G; g++) {
map<int,double>::iterator it;
for (it = OmegaInvSparse[g].begin(); it != OmegaInvSparse[g].end(); it++) {
int gg = it->first;
double value = it->second;
int qq;
for (q = 0; q < Q; q++) {
for (qq = 0; qq < Q; qq++) {
int index = g * Q + q;
int indexindex = gg * Q + qq;
double cov = value * RInverse[q][qq];
VinvSparse[index].insert(pair<int,double>(indexindex,cov));
}
}
}
}
// print precision matrix of full conditional
/*
{
char filename[120];
sprintf(filename,"VinvReduced.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < VinvSparse.size(); i++) {
for (j = 0; j < VinvSparse.size(); j++) {
double value = 0.0;
if (VinvSparse[i].find(j) != VinvSparse[i].end())
value = VinvSparse[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// add effect of data
vector<double> l(VinvSparse.size(),0.0);
vector<double> L(VinvSparse.size(),0.0);
for (g = 0; g < G; g++)
for (q = 0; q < Q; q++) {
if (delta[qg2index(q,g,Q,G)] == 1) {
int index = g * Q + q;
double v0 = sigma2[qg2index(q,g,Q,G)] * phi[qg2index(q,g,Q,G)];
double v1 = sigma2[qg2index(q,g,Q,G)] / phi[qg2index(q,g,Q,G)];
double diag = 0.0;
double ll = 0.0;
int s;
for (s = 0; s < S[q]; s++) {
double variance = psi[sq2index(s,q,S,Q)] == 0 ? v0 : v1;
diag += exp(b[q] * log(sigma2[qg2index(q,g,Q,G)])) / variance;
ll += (2.0 * psi[sq2index(s,q,S,Q)] - 1.0) * (x[sqg2index(s,q,g,S,Q,G)] - nu[qg2index(q,g,Q,G)]) / variance;
}
ll *= exp(0.5 * b[q] * log(sigma2[qg2index(q,g,Q,G)]));
L[index] = diag;
l[index] = ll;
}
}
int index;
for (index = 0; index < L.size(); index++)
VinvSparse[index][index] += L[index];
// print L and l
/*
{
char filename[120];
sprintf(filename,"l.txt");
FILE *out = fopen(filename,"w");
int i;
for (i = 0; i < l.size(); i++) {
fprintf(out,"%20.18e\n",l[i]);
}
fclose(out);
sprintf(filename,"L.txt");
out = fopen(filename,"w");
for (i = 0; i < L.size(); i++) {
fprintf(out,"%20.18e\n",L[i]);
}
fclose(out);
}
*/
// print precision matrix of full conditional
/*
{
char filename[120];
sprintf(filename,"Vinv.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < VinvSparse.size(); i++) {
for (j = 0; j < VinvSparse.size(); j++) {
double value = 0.0;
if (VinvSparse[i].find(j) != VinvSparse[i].end())
value = VinvSparse[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// establish a version of VinvSparse with the indices reversed
vector<map<int,double> > VinvSparseReversed;
VinvSparseReversed.resize(VinvSparse.size());
for (k = 0; k < VinvSparseReversed.size(); k++)
VinvSparseReversed[k].clear();
for (k = 0; k < VinvSparseReversed.size(); k++) {
map<int,double>::iterator it;
for (it = VinvSparse[k].begin(); it != VinvSparse[k].end(); it++) {
map<int,double>::iterator itextra = VinvSparse[k].end();
int r = it->first;
double value = it->second;
int kk = VinvSparseReversed.size() - k - 1;
int rr = VinvSparseReversed.size() - r - 1;
VinvSparseReversed[rr].insert(pair<int,double>(kk,value));
}
}
// print VinvSparseReversed
/*
{
char filename[120];
sprintf(filename,"VinvReversed.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < VinvSparseReversed.size(); i++) {
for (j = 0; j < VinvSparseReversed.size(); j++) {
double value = 0.0;
if (VinvSparseReversed[i].find(j) != VinvSparseReversed[i].end())
value = VinvSparseReversed[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// perform Cholesky decomposition for (the sparse matrix) VinvSparseReversed
// cout << "start Cholesky factorization" << endl;
vector<map<int,double> > cholReversed;
cholReversed.resize(Q * G);
for (index = 0; index < cholReversed.size(); index++)
cholReversed[index].clear();
int N = cholReversed.size();
int i;
for (i = 0; i < N; i++) {
map<int,double>::iterator it;
for (it = VinvSparseReversed[i].find(i); it != VinvSparseReversed[i].end(); it++) {
int j = it->first;
double value = it->second;
double sum = value;
map<int,double>::iterator it2;
for (it2 = cholReversed[i].begin(); it2 != cholReversed[i].end(); it2++) {
if (it2->first < i) {
int k = it2->first;
double valuei = it2->second;
map<int,double>::iterator it3 = cholReversed[j].find(k);
if (it3 != cholReversed[j].end()) {
double valuej = it3->second;
sum -= valuei * valuej;
}
}
}
if (i == j && sum <= 0.0) {
fprintf(stderr,"DeltaStarGibbs: Matrix is not positive definite!\n");
exit(-1);
}
if (i == j)
cholReversed[j].insert(pair<int,double>(i,sqrt(sum)));
else
cholReversed[j].insert(pair<int,double>(i,sum / cholReversed[i][i]));
}
}
/*
// print (reversed) cholesky matrix
{
char filename[120];
sprintf(filename,"cholReversed.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < cholReversed.size(); i++) {
for (j = 0; j < cholReversed.size(); j++) {
double value = 0.0;
if (cholReversed[i].find(j) != cholReversed[i].end())
value = cholReversed[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// establish er version of chol with the indices reversed back
vector<map<int,double> > chol;
chol.resize(cholReversed.size());
for (k = 0; k < chol.size(); k++)
chol[k].clear();
for (k = 0; k < cholReversed.size(); k++) {
map<int,double>::iterator it;
for (it = cholReversed[k].begin(); it != cholReversed[k].end(); it++) {
int r = it->first;
double value = it->second;
int kk = chol.size() - k - 1;
int rr = chol.size() - r - 1;
chol[rr].insert(pair<int,double>(kk,value));
}
}
// print cholesky matrix
/*
{
char filename[120];
sprintf(filename,"chol.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < chol.size(); i++) {
for (j = 0; j < chol.size(); j++) {
double value = 0.0;
if (chol[i].find(j) != chol[i].end())
value = chol[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// establish representation of the transpose of the cholesky matrix
vector<map<int,double> > cholT;
cholT.resize(chol.size());
for (k = 0; k < cholT.size(); k++)
cholT[k].clear();
for (i = 0; i < chol.size(); i++) {
map<int,double>::iterator it;
for (it = chol[i].begin(); it != chol[i].end(); it++) {
int j = it->first;
double value = it->second;
cholT[j].insert(pair<int,double>(i,value));
}
}
// print cholT
/*
{
char filename[120];
sprintf(filename,"cholT.txt");
FILE *out = fopen(filename,"w");
int i,j;
for (i = 0; i < cholT.size(); i++) {
for (j = 0; j < cholT.size(); j++) {
double value = 0.0;
if (cholT[i].find(j) != cholT[i].end())
value = cholT[i][j];
fprintf(out,"%20.18e ",value);
}
fprintf(out,"\n");
}
fclose(out);
}
*/
// finished printing
// cout << "start computing mean value" << endl;
vector<double> u(l.size(),0.0);
for (i = u.size() - 1; i >= 0; i--) {
double diag = 0.0;
double sum = 0.0;
map<int,double>::iterator it;
for (it = cholT[i].begin(); it != cholT[i].end(); it++) {
int j= it->first;
double value = it->second;
if (i == j)
diag = value;
else
sum += value * u[j];
}
u[i] = (l[i] - sum) / diag;
}
vector<double> mean(u.size(),0.0);
for (i = 0; i < mean.size(); i++) {
double diag = 0.0;
double sum = 0.0;
map<int,double>::iterator it;
for (it = chol[i].begin(); it != chol[i].end(); it++) {
int j = it->first;
double value = it->second;
if (j == i)
diag = value;
else
sum += value * mean[j];
}
mean[i] = (u[i] - sum) / diag;
}
// print mean value
/*
{
char filename[120];
sprintf(filename,"mean.txt");
FILE *out = fopen(filename,"w");
int i;
for (i = 0; i < mean.size(); i++) {
fprintf(out,"%20.18e\n",mean[i]);
}
fclose(out);
}
*/
// finished printing
// generate a sample with zero mean, or compute the sample that should have been sampled
// cout << "start sampling" << endl;
vector<double> sample(chol.size(),0.0);
if (draw == 1) {
vector<double> z(chol.size(),0.0);
for (k = 0; k < z.size(); k++)
z[k] = ran.Norm01();
// print z
/*
{
char filename[120];
sprintf(filename,"z.txt");
FILE *out = fopen(filename,"w");
int i;
for (i = 0; i < z.size(); i++) {
fprintf(out,"%20.18e\n",z[i]);
}
fclose(out);
}
*/
// finished printing
for (i = 0; i < sample.size(); i++) {
double diag = 0.0;
double sum = 0.0;
map<int,double>::iterator it;
for (it = chol[i].begin(); it != chol[i].end(); it++) {
int j = it->first;
double value = it->second;
if (j == i)
diag = value;
else
sum += value * sample[j];
}
sample[i] = (z[i] - sum) / diag;
}
// print sample
/*
{
char filename[120];
sprintf(filename,"sample.txt");
FILE *out = fopen(filename,"w");
int i;
for (i = 0; i < sample.size(); i++) {
fprintf(out,"%20.18e\n",sample[i]);
}
fclose(out);
}
*/
// finished printing
}
else { // compute sample[i] necessary to generate the current values
// compute DeltaStar
vector<vector<double> > DeltaStar;
DeltaStar.resize(G);
for (g = 0; g < G; g++) {
DeltaStar[g].resize(Q);
for (q = 0; q < Q; q++) {
DeltaStar[g][q] = Delta[qg2index(q,g,Q,G)] / exp(0.5 * b[q] * log(sigma2[qg2index(q,g,Q,G)]));
}
}
// subtract mean value
for (g = 0; g < G; g++)
for (q = 0; q < Q; q++) {
int index = g * Q + q;
DeltaStar[g][q] -= mean[index];
}
// insert value in sample[]
for (g = 0; g < G; g++)
for (q = 0; q < Q; q++) {
int index = g * Q + q;
sample[index] = DeltaStar[g][q];
}
}
// compute potential for sample
double pot = 0.0;
for (i = 0; i < sample.size(); i++) {
double diag = 0.0;
double sum = 0.0;
map<int,double>::iterator it;
for (it = chol[i].begin(); it != chol[i].end(); it++) {
int j = it->first;
double value = it->second;
if (j == i)
diag = value;
else
sum += value * sample[j];
}
double mean = - sum / diag;
double variance = 1.0 / (diag * diag);
pot += ran.PotentialGaussian(variance,mean,sample[i]);
}
if (draw == 1) { // add mean value and insert in data structure
for (k = 0; k < Q * G; k++)
sample[k] += mean[k];
for (g = 0; g < G; g++)
for (q = 0; q < Q; q++) {
int index = g * Q + q;
double DeltaStar = sample[index];
Delta[qg2index(q,g,Q,G)] = DeltaStar * exp(0.5 * b[q] * log(sigma2[qg2index(q,g,Q,G)]));
}
}
// cout << "finished" << endl;
return pot;
}
void GaussianMean1DRegressionCompute(const QUESO::BaseEnvironment& env,
double priorMean, double priorVar, const likelihoodData& dat)
{
// parameter space: 1-D on (-infinity, infinity)
QUESO::VectorSpace<P_V, P_M> paramSpace(
env, // queso environment
"param_", // name prefix
1, // dimensions
NULL); // names
P_V paramMin(paramSpace.zeroVector());
P_V paramMax(paramSpace.zeroVector());
paramMin[0] = -INFINITY;
paramMax[0] = INFINITY;
QUESO::BoxSubset<P_V, P_M> paramDomain(
"paramBox_", // name prefix
paramSpace, // vector space
paramMin, // min values
paramMax); // max values
// gaussian prior with user supplied mean and variance
P_V priorMeanVec(paramSpace.zeroVector());
P_V priorVarVec(paramSpace.zeroVector());
priorMeanVec[0] = priorMean;
priorVarVec[0] = priorVar;
QUESO::GaussianVectorRV<P_V, P_M> priorRv("prior_", paramDomain, priorMeanVec,
priorVarVec);
// likelihood is important
QUESO::GenericScalarFunction<P_V, P_M> likelihoodFunctionObj(
"like_", // name prefix
paramDomain, // image set
LikelihoodFunc<P_V, P_M>, // routine
(void *) &dat, // routine data ptr
true); // routineIsForLn
QUESO::GenericVectorRV<P_V, P_M> postRv(
"post_", // name prefix
paramSpace); // image set
// Initialize and solve the Inverse Problem with Bayes multi-level sampling
QUESO::StatisticalInverseProblem<P_V, P_M> invProb(
"", // name prefix
NULL, // alt options
priorRv, // prior RV
likelihoodFunctionObj, // likelihood fcn
postRv); // posterior RV
invProb.solveWithBayesMLSampling();
// compute mean and second moment of samples on each proc via Knuth online mean/variance algorithm
int N = invProb.postRv().realizer().subPeriod();
double subMean = 0.0;
double subM2 = 0.0;
double delta;
P_V sample(paramSpace.zeroVector());
for (int n = 1; n <= N; n++) {
invProb.postRv().realizer().realization(sample);
delta = sample[0] - subMean;
subMean += delta / n;
subM2 += delta * (sample[0] - subMean);
}
// gather all Ns, means, and M2s to proc 0
std::vector<int> unifiedNs(env.inter0Comm().NumProc());
std::vector<double> unifiedMeans(env.inter0Comm().NumProc());
std::vector<double> unifiedM2s(env.inter0Comm().NumProc());
MPI_Gather(&N, 1, MPI_INT, &(unifiedNs[0]), 1, MPI_INT, 0,
env.inter0Comm().Comm());
MPI_Gather(&subMean, 1, MPI_DOUBLE, &(unifiedMeans[0]), 1, MPI_DOUBLE, 0,
env.inter0Comm().Comm());
MPI_Gather(&subM2, 1, MPI_DOUBLE, &(unifiedM2s[0]), 1, MPI_DOUBLE, 0,
env.inter0Comm().Comm());
// get the total number of likelihood calls at proc 0
unsigned long totalLikelihoodCalls = 0;
MPI_Reduce(&likelihoodCalls, &totalLikelihoodCalls, 1, MPI_UNSIGNED_LONG,
MPI_SUM, 0, env.inter0Comm().Comm());
// compute global posterior mean and std via Chan algorithm, output results on proc 0
if (env.inter0Rank() == 0) {
int postN = unifiedNs[0];
double postMean = unifiedMeans[0];
double postVar = unifiedM2s[0];
for (unsigned int i = 1; i < unifiedNs.size(); i++) {
delta = unifiedMeans[i] - postMean;
postMean = (postN * postMean + unifiedNs[i] * unifiedMeans[i]) /
(postN + unifiedNs[i]);
postVar += unifiedM2s[i] + delta * delta *
(((double)postN * unifiedNs[i]) / (postN + unifiedNs[i]));
postN += unifiedNs[i];
}
postVar /= postN;
//compute exact answer - available in this case since the exact posterior is a gaussian
N = dat.dataSet.size();
double dataSum = 0.0;
for (int i = 0; i < N; i++)
dataSum += dat.dataSet[i];
double datMean = dataSum / N;
double postMeanExact = (N * priorVar / (N * priorVar + dat.samplingVar)) *
datMean + (dat.samplingVar / (N * priorVar + dat.samplingVar)) * priorMean;
double postVarExact = 1.0 / (N / dat.samplingVar + 1.0 / priorVar);
std::cout << "Number of posterior samples: " << postN << std::endl;
std::cout << "Estimated posterior mean: " << postMean << " +/- "
<< std::sqrt(postVar) << std::endl;
std::cout << "Likelihood function calls: " << totalLikelihoodCalls
<< std::endl;
std::cout << "\nExact posterior: Gaussian with mean " << postMeanExact
<< ", standard deviation " << std::sqrt(postVarExact) << std::endl;
}
}
Point Geometry::sample(const Point&, const GeomSample &gs, Normal &normal) const {
return sample(gs, normal);
}
int main (int argc, char **argv)
{
char c;
unsigned int flag = 0;
int interval = 1, count = 0, max_count = 1;
struct vg_data vg_now, vg_prev;
VMGuestLibError ret;
while ((c = getopt(argc, argv, "i:c:hvru")) != -1) {
switch(c) {
case 'i':
interval = atoi(optarg);
break;
case 'c':
max_count = atoi(optarg);
break;
case 'h':
usage();
return 0;
break;
case 'r': /* raw output */
flag |= FLAG_RAWOUTPUT;
break;
case 'v': /* verbose mode */
flag |= FLAG_VERBOSE;
break;
case 'u':
flag |= FLAG_UNIXTIME;
break;
default:
printf("Unkown option '%c'\n", c);
}
}
memset(&vg_now, 0x0, sizeof(struct vg_data));
ret = VMGuestLib_OpenHandle(&vg_now.handle);
if (ret != VMGUESTLIB_ERROR_SUCCESS) {
if (IS_VERBOSE(flag)) {
printf("VMGuestLib_OpenHandle: %d (%s)\n",
ret, VMGuestLib_GetErrorText(ret));
}
return 1;
}
if (sample(&vg_now, flag) != 0) {
goto bailout;
}
if (IS_RAWOUTPUT(flag)) {
printf("Timestamp "
"SessionId "
"HostProcessorSpeed "
"CpuReservationMHz CpuLimitMHz CpuShares "
"ElapsedMs CpuUsedMs CpuStolenMs "
"MemReservationMB MemLimitMB MemShares MemMappedMB "
"MemActiveMB MemOverheadMB MemBalloonedMB MemSwappedMB "
"MemSharedMB MemSharedSavedMB MemUsedMB\n"
);
} else {
printf("%-24s %-8s %-8s %8s %8s %8s %8s\n",
"Timestamp", "intvl(g)", "intvl(h)",
"used", "stolen", "%used", "%ready");
}
for (count = 0; count < max_count; count++) {
vg_prev = vg_now;
sleep(interval);
if (sample(&vg_now, flag) != 0) {
goto bailout;
}
output(&vg_now, &vg_prev, flag);
}
bailout:
ret = VMGuestLib_CloseHandle(vg_now.handle);
if (ret != VMGUESTLIB_ERROR_SUCCESS) {
if (IS_VERBOSE(flag)) {
printf("VMGuestLib_CloseHandle: %d (%s)\n",
ret, VMGuestLib_GetErrorText(ret));
}
return 1;
}
return 0;
}
int main ( int argc, char *argv[] )
{
//Variables for parsing the data file
std::string filename = "SPECT.train";
std::string line;
std::stringstream parse;
int ssize = 100; //establish a buffer size to store attribute values,
//which for binary classification string are no bigger than 1
char c[ssize];
char delimiter = ',';
//Variables to store the values in the data file
std::vector<int> tmpcase;
std::vector< std::vector<int> > training_set;
cv::Mat sample(0, 1, CV_32FC1);
cv::Mat labels(0, 1 , CV_16SC1);
cv::Mat train_set;
std::ifstream dataset_file(filename.c_str(), std::ios::in);
if(!dataset_file)
{
std::cerr << "Cannot load training set file" << std::endl;
}
else
{
while( (getline(dataset_file, line))!= NULL )
{
parse << line;
while( parse.getline(c,ssize,delimiter) )
{
tmpcase.push_back( (*c-'0') );
sample.push_back( (float)(*c-'0') );
}
parse.str(""); //safety measure to erase previous contents
parse.clear(); //clear flags to be able to read from it again
training_set.push_back(tmpcase);
tmpcase.clear();
train_set.push_back(sample.reshape(0,1));
labels.push_back((int)(sample.at<float>(0)));
sample = cv::Mat();
}
}
std::cout << train_set << std::endl;
cv::FileStorage fstore_traindata("spect_train.yml",cv::FileStorage::WRITE);
cv::Mat train_samples(train_set.colRange(1,train_set.cols));
fstore_traindata << "train_samples" << train_samples;
fstore_traindata << "train_labels" << labels;
fstore_traindata.release();
std::cout << train_samples << std::endl;
std::cout << labels << std::endl;
std::vector<int> tmp;
for(std::vector< std::vector<int> >::iterator it = training_set.begin(); it != training_set.end(); ++it)
{
tmp = *it;
for(std::vector<int>::iterator it2 = tmp.begin(); it2 != tmp.end(); ++it2)
{
std::cout << *it2 << " ";
}
std::cout << std::endl;
tmp.clear();
}
}
static int compute_tree_bagging(ETree *etree,int n,int d,double *x[],
int y[], int nmodels,int stumps, int minsize)
{
int i,b;
int *samples;
double **trx;
int *try;
if(nmodels<1){
fprintf(stderr,"compute_tree_bagging: nmodels must be greater than 0\n");
return 1;
}
if(stumps != 0 && stumps != 1){
fprintf(stderr,"compute_tree_bagging: parameter stumps must be 0 or 1\n");
return 1;
}
if(minsize < 0){
fprintf(stderr,"compute_tree_bagging: parameter minsize must be >= 0\n");
return 1;
}
etree->nclasses=iunique(y,n, &(etree->classes));
if(etree->nclasses<=0){
fprintf(stderr,"compute_tree_bagging: iunique error\n");
return 1;
}
if(etree->nclasses==1){
fprintf(stderr,"compute_tree_bagging: only 1 class recognized\n");
return 1;
}
if(etree->nclasses==2)
if(etree->classes[0] != -1 || etree->classes[1] != 1){
fprintf(stderr,"compute_tree_bagging: for binary classification classes must be -1,1\n");
return 1;
}
if(etree->nclasses>2)
for(i=0;i<etree->nclasses;i++)
if(etree->classes[i] != i+1){
fprintf(stderr,"compute_tree_bagging: for %d-class classification classes must be 1,...,%d\n",etree->nclasses,etree->nclasses);
return 1;
}
if(!(etree->tree=(Tree *)calloc(nmodels,sizeof(Tree)))){
fprintf(stderr,"compute_tree_bagging: out of memory\n");
return 1;
}
etree->nmodels=nmodels;
if(!(etree->weights=dvector(nmodels))){
fprintf(stderr,"compute_tree_bagging: out of memory\n");
return 1;
}
for(b=0;b<nmodels;b++)
etree->weights[b]=1.0 / (double) nmodels;
if(!(trx=(double **)calloc(n,sizeof(double*)))){
fprintf(stderr,"compute_tree_bagging: out of memory\n");
return 1;
}
if(!(try=ivector(n))){
fprintf(stderr,"compute_tree_bagging: out of memory\n");
return 1;
}
for(b=0;b<nmodels;b++){
if(sample(n, NULL, n, &samples, TRUE,b)!=0){
fprintf(stderr,"compute_tree_bagging: sample error\n");
return 1;
}
for(i =0;i<n;i++){
trx[i] = x[samples[i]];
try[i] = y[samples[i]];
}
if(compute_tree(&(etree->tree[b]),n,d,trx,try,stumps,minsize)!=0){
fprintf(stderr,"compute_tree_bagging: compute_tree error\n");
return 1;
}
free_ivector(samples);
}
free(trx);
free_ivector(try);
return 0;
}
static int compute_tree_aggregate(ETree *etree,int n,int d,double *x[],int y[],
int nmodels,int stumps, int minsize)
{
int i,b;
int *samples;
double **trx;
int *try;
int indx;
if(nmodels<1){
fprintf(stderr,"compute_tree_aggregate: nmodels must be greater than 0\n");
return 1;
}
if(nmodels > n){
fprintf(stderr,"compute_tree_aggregate: nmodels must be less than n\n");
return 1;
}
if(stumps != 0 && stumps != 1){
fprintf(stderr,"compute_tree_bagging: parameter stumps must be 0 or 1\n");
return 1;
}
if(minsize < 0){
fprintf(stderr,"compute_tree_bagging: parameter minsize must be >= 0\n");
return 1;
}
etree->nclasses=iunique(y,n, &(etree->classes));
if(etree->nclasses<=0){
fprintf(stderr,"compute_tree_aggregate: iunique error\n");
return 1;
}
if(etree->nclasses==1){
fprintf(stderr,"compute_tree_aggregate: only 1 class recognized\n");
return 1;
}
if(etree->nclasses==2)
if(etree->classes[0] != -1 || etree->classes[1] != 1){
fprintf(stderr,"compute_tree_aggregate: for binary classification classes must be -1,1\n");
return 1;
}
if(etree->nclasses>2)
for(i=0;i<etree->nclasses;i++)
if(etree->classes[i] != i+1){
fprintf(stderr,"compute_tree_aggregate: for %d-class classification classes must be 1,...,%d\n",etree->nclasses,etree->nclasses);
return 1;
}
if(!(etree->tree=(Tree *)calloc(nmodels,sizeof(Tree)))){
fprintf(stderr,"compute_tree_aggregate: out of memory\n");
return 1;
}
etree->nmodels=nmodels;
if(!(etree->weights=dvector(nmodels))){
fprintf(stderr,"compute_tree_aggregate: out of memory\n");
return 1;
}
for(b=0;b<nmodels;b++)
etree->weights[b]=1.0 / (double) nmodels;
if(!(trx=(double **)calloc(n,sizeof(double*)))){
fprintf(stderr,"compute_tree_aggregate: out of memory\n");
return 1;
}
if(!(try=ivector(n))){
fprintf(stderr,"compute_tree_aggregate: out of memory\n");
return 1;
}
if(sample(nmodels, NULL, n, &samples, TRUE,0)!=0){
fprintf(stderr,"compute_tree_aggregate: sample error\n");
return 1;
}
for(b=0;b<nmodels;b++){
indx=0;
for(i=0;i<n;i++)
if(samples[i] == b){
trx[indx] = x[i];
try[indx++] = y[i];
}
if(compute_tree(&(etree->tree[b]),indx,d,trx,try,stumps,minsize)!=0){
fprintf(stderr,"compute_tree_aggregate: compute_tree error\n");
return 1;
}
}
free_ivector(samples);
free(trx);
free_ivector(try);
return 0;
}
void MetropolisRenderer::Render(const Scene *scene) {
PBRT_MLT_STARTED_RENDERING();
if (scene->lights.size() > 0) {
int x0, x1, y0, y1;
camera->film->GetPixelExtent(&x0, &x1, &y0, &y1);
float t0 = camera->shutterOpen, t1 = camera->shutterClose;
Distribution1D *lightDistribution = ComputeLightSamplingCDF(scene);
if (directLighting != NULL) {
PBRT_MLT_STARTED_DIRECTLIGHTING();
// Compute direct lighting before Metropolis light transport
if (nDirectPixelSamples > 0) {
LDSampler sampler(x0, x1, y0, y1, nDirectPixelSamples, t0, t1);
Sample *sample = new Sample(&sampler, directLighting, NULL, scene);
vector<Task *> directTasks;
int nDirectTasks = max(32 * NumSystemCores(),
(camera->film->xResolution * camera->film->yResolution) / (16*16));
nDirectTasks = RoundUpPow2(nDirectTasks);
ProgressReporter directProgress(nDirectTasks, "Direct Lighting");
for (int i = 0; i < nDirectTasks; ++i)
directTasks.push_back(new SamplerRendererTask(scene, this, camera, directProgress,
&sampler, sample, false, i, nDirectTasks));
std::reverse(directTasks.begin(), directTasks.end());
EnqueueTasks(directTasks);
WaitForAllTasks();
for (uint32_t i = 0; i < directTasks.size(); ++i)
delete directTasks[i];
delete sample;
directProgress.Done();
}
camera->film->WriteImage();
PBRT_MLT_FINISHED_DIRECTLIGHTING();
}
// Take initial set of samples to compute $b$
PBRT_MLT_STARTED_BOOTSTRAPPING(nBootstrap);
RNG rng(0);
MemoryArena arena;
vector<float> bootstrapI;
vector<PathVertex> cameraPath(maxDepth, PathVertex());
vector<PathVertex> lightPath(maxDepth, PathVertex());
float sumI = 0.f;
bootstrapI.reserve(nBootstrap);
MLTSample sample(maxDepth);
for (uint32_t i = 0; i < nBootstrap; ++i) {
// Generate random sample and path radiance for MLT bootstrapping
float x = Lerp(rng.RandomFloat(), x0, x1);
float y = Lerp(rng.RandomFloat(), y0, y1);
LargeStep(rng, &sample, maxDepth, x, y, t0, t1, bidirectional);
Spectrum L = PathL(sample, scene, arena, camera, lightDistribution,
&cameraPath[0], &lightPath[0], rng);
// Compute contribution for random sample for MLT bootstrapping
float I = ::I(L);
sumI += I;
bootstrapI.push_back(I);
arena.FreeAll();
}
float b = sumI / nBootstrap;
PBRT_MLT_FINISHED_BOOTSTRAPPING(b);
Info("MLT computed b = %f", b);
// Select initial sample from bootstrap samples
float contribOffset = rng.RandomFloat() * sumI;
rng.Seed(0);
sumI = 0.f;
MLTSample initialSample(maxDepth);
for (uint32_t i = 0; i < nBootstrap; ++i) {
float x = Lerp(rng.RandomFloat(), x0, x1);
float y = Lerp(rng.RandomFloat(), y0, y1);
LargeStep(rng, &initialSample, maxDepth, x, y, t0, t1,
bidirectional);
sumI += bootstrapI[i];
if (sumI > contribOffset)
break;
}
// Launch tasks to generate Metropolis samples
uint32_t nTasks = largeStepsPerPixel;
uint32_t largeStepRate = nPixelSamples / largeStepsPerPixel;
Info("MLT running %d tasks, large step rate %d", nTasks, largeStepRate);
ProgressReporter progress(nTasks * largeStepRate, "Metropolis");
vector<Task *> tasks;
Mutex *filmMutex = Mutex::Create();
Assert(IsPowerOf2(nTasks));
uint32_t scramble[2] = { rng.RandomUInt(), rng.RandomUInt() };
uint32_t pfreq = (x1-x0) * (y1-y0);
for (uint32_t i = 0; i < nTasks; ++i) {
float d[2];
Sample02(i, scramble, d);
tasks.push_back(new MLTTask(progress, pfreq, i,
d[0], d[1], x0, x1, y0, y1, t0, t1, b, initialSample,
scene, camera, this, filmMutex, lightDistribution));
}
EnqueueTasks(tasks);
WaitForAllTasks();
for (uint32_t i = 0; i < tasks.size(); ++i)
delete tasks[i];
progress.Done();
Mutex::Destroy(filmMutex);
delete lightDistribution;
}
camera->film->WriteImage();
PBRT_MLT_FINISHED_RENDERING();
}
static int compute_tree_adaboost(ETree *etree,int n,int d,double *x[],int y[],
int nmodels,int stumps, int minsize)
{
int i,b;
int *samples;
double **trx;
int *try;
double *prob;
double *prob_copy;
double sumalpha;
double eps;
int *pred;
double *margin;
double sumprob;
if(nmodels<1){
fprintf(stderr,"compute_tree_adaboost: nmodels must be greater than 0\n");
return 1;
}
if(stumps != 0 && stumps != 1){
fprintf(stderr,"compute_tree_bagging: parameter stumps must be 0 or 1\n");
return 1;
}
if(minsize < 0){
fprintf(stderr,"compute_tree_bagging: parameter minsize must be >= 0\n");
return 1;
}
etree->nclasses=iunique(y,n, &(etree->classes));
if(etree->nclasses<=0){
fprintf(stderr,"compute_tree_adaboost: iunique error\n");
return 1;
}
if(etree->nclasses==1){
fprintf(stderr,"compute_tree_adaboost: only 1 class recognized\n");
return 1;
}
if(etree->nclasses==2)
if(etree->classes[0] != -1 || etree->classes[1] != 1){
fprintf(stderr,"compute_tree_adaboost: for binary classification classes must be -1,1\n");
return 1;
}
if(etree->nclasses>2){
fprintf(stderr,"compute_tree_adaboost: multiclass classification not allowed\n");
return 1;
}
if(!(etree->tree=(Tree *)calloc(nmodels,sizeof(Tree)))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(etree->weights=dvector(nmodels))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(trx=(double **)calloc(n,sizeof(double*)))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(try=ivector(n))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(prob_copy=dvector(n))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(prob=dvector(n))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
if(!(pred=ivector(n))){
fprintf(stderr,"compute_tree_adaboost: out of memory\n");
return 1;
}
for(i =0;i<n;i++)
prob[i]=1.0/(double)n;
etree->nmodels=nmodels;
sumalpha=0.0;
for(b=0;b<nmodels;b++){
for(i =0;i<n;i++)
prob_copy[i]=prob[i];
if(sample(n, prob_copy, n, &samples, TRUE,b)!=0){
fprintf(stderr,"compute_tree_adaboost: sample error\n");
return 1;
}
for(i=0;i<n;i++){
trx[i] = x[samples[i]];
try[i] = y[samples[i]];
}
if(compute_tree(&(etree->tree[b]),n,d,trx,try,stumps,minsize)!=0){
fprintf(stderr,"compute_tree_adaboost: compute_tree error\n");
return 1;
}
free_ivector(samples);
eps=0.0;
for(i=0;i<n;i++){
pred[i]=predict_tree(&(etree->tree[b]),x[i],&margin);
if(pred[i] < -1 ){
fprintf(stderr,"compute_tree_adaboost: predict_tree error\n");
return 1;
}
if(pred[i]==0 || pred[i] != y[i])
eps += prob[i];
free_dvector(margin);
}
if(eps > 0.0 && eps < 0.5){
etree->weights[b]=0.5 *log((1.0-eps)/eps);
sumalpha+=etree->weights[b];
}else{
etree->nmodels=b;
break;
}
sumprob=0.0;
for(i=0;i<n;i++){
prob[i]=prob[i]*exp(-etree->weights[b]*y[i]*pred[i]);
sumprob+=prob[i];
}
if(sumprob <=0.0){
fprintf(stderr,"compute_tree_adaboost: sumprob = 0\n");
return 1;
}
for(i=0;i<n;i++)
prob[i] /= sumprob;
}
if(etree->nmodels<=0){
fprintf(stderr,"compute_tree_adaboost: no models produced\n");
return 1;
}
if(sumalpha <=0){
fprintf(stderr,"compute_tree_adaboost: sumalpha = 0\n");
return 1;
}
for(b=0;b<etree->nmodels;b++)
etree->weights[b] /= sumalpha;
free(trx);
free_ivector(try);
free_ivector(pred);
free_dvector(prob);
free_dvector(prob_copy);
return 0;
}
static void split_node(Node *node,Node *nodeL,Node *nodeR,int classes[],
int nclasses)
{
int **indx;
double *tmpvar;
int i,j,k;
int **npL , **npR;
double **prL , **prR;
int totL,totR;
double a,b;
double *decrease_in_inpurity;
double max_decrease=0;
int splitvar;
int splitvalue;
int morenumerous;
nodeL->priors=dvector(nclasses);
nodeR->priors=dvector(nclasses);
nodeL->npoints_for_class=ivector(nclasses);
nodeR->npoints_for_class=ivector(nclasses);
indx=imatrix(node->nvar,node->npoints);
tmpvar=dvector(node->npoints);
decrease_in_inpurity=dvector(node->npoints-1);
npL=imatrix(node->npoints,nclasses);
npR=imatrix(node->npoints,nclasses);
prL=dmatrix(node->npoints,nclasses);
prR=dmatrix(node->npoints,nclasses);
splitvar=0;
splitvalue=0;
max_decrease=0;
for(i=0;i<node->nvar;i++){
for(j=0;j<node->npoints;j++)
tmpvar[j]=node->data[j][i];
for(j=0;j<node->npoints;j++)
indx[i][j]=j;
dsort(tmpvar,indx[i],node->npoints,SORT_ASCENDING);
for(k=0;k<nclasses;k++)
if(node->classes[indx[i][0]]==classes[k]){
npL[0][k] = 1;
npR[0][k] = node->npoints_for_class[k]-npL[0][k];
} else{
npL[0][k] = 0;
npR[0][k] = node->npoints_for_class[k];
}
for(j=1;j<node->npoints-1;j++)
for(k=0;k<nclasses;k++)
if(node->classes[indx[i][j]]==classes[k]){
npL[j][k] = npL[j-1][k] +1;
npR[j][k] = node->npoints_for_class[k] - npL[j][k];
}
else {
npL[j][k] = npL[j-1][k];
npR[j][k] = node->npoints_for_class[k] - npL[j][k];
}
for(j=0;j<node->npoints-1;j++){
if(node->data[indx[i][j]][i] != node->data[indx[i][j+1]][i]){
totL = totR = 0;
for(k=0;k<nclasses;k++)
totL += npL[j][k];
for(k=0;k<nclasses;k++)
prL[j][k] = (double) npL[j][k] / (double) totL;
for(k=0;k<nclasses;k++)
totR += npR[j][k];
for(k=0;k<nclasses;k++)
prR[j][k] = (double) npR[j][k] /(double) totR;
a = (double) totL / (double) node->npoints;
b = (double) totR / (double) node->npoints ;
decrease_in_inpurity[j] = gini_index(node->priors,nclasses) -
a * gini_index(prL[j],nclasses) - b * gini_index(prR[j],nclasses);
}
}
for(j=0;j<node->npoints-1;j++)
if(decrease_in_inpurity[j] > max_decrease){
max_decrease = decrease_in_inpurity[j];
splitvar=i;
splitvalue=j;
for(k=0;k<nclasses;k++){
nodeL->priors[k]=prL[splitvalue][k];
nodeR->priors[k]=prR[splitvalue][k];
nodeL->npoints_for_class[k]=npL[splitvalue][k];
nodeR->npoints_for_class[k]=npR[splitvalue][k];
}
}
}
node->var=splitvar;
node->value=(node->data[indx[splitvar][splitvalue]][node->var]+
node->data[indx[splitvar][splitvalue+1]][node->var])/2.;
nodeL->nvar=node->nvar;
nodeL->nclasses=node->nclasses;
nodeL->npoints=splitvalue+1;
nodeL->terminal=TRUE;
if(gini_index(nodeL->priors,nclasses) >0)
nodeL->terminal=FALSE;
nodeL->data=(double **) calloc(nodeL->npoints,sizeof(double *));
nodeL->classes=ivector(nodeL->npoints);
for(i=0;i<nodeL->npoints;i++){
nodeL->data[i] = node->data[indx[splitvar][i]];
nodeL->classes[i] = node->classes[indx[splitvar][i]];
}
morenumerous=0;
for(k=0;k<nclasses;k++)
if(nodeL->npoints_for_class[k] > morenumerous){
morenumerous = nodeL->npoints_for_class[k];
nodeL->node_class=classes[k];
}
nodeR->nvar=node->nvar;
nodeR->nclasses=node->nclasses;
nodeR->npoints=node->npoints-nodeL->npoints;
nodeR->terminal=TRUE;
if(gini_index(nodeR->priors,nclasses) >0)
nodeR->terminal=FALSE;
nodeR->data=(double **) calloc(nodeR->npoints,sizeof(double *));
nodeR->classes=ivector(nodeR->npoints);
for(i=0;i<nodeR->npoints;i++){
nodeR->data[i] = node->data[indx[splitvar][nodeL->npoints+i]];
nodeR->classes[i] = node->classes[indx[splitvar][nodeL->npoints+i]];
}
morenumerous=0;
for(k=0;k<nclasses;k++)
if(nodeR->npoints_for_class[k] > morenumerous){
morenumerous = nodeR->npoints_for_class[k];
nodeR->node_class=classes[k];
}
free_imatrix(indx, node->nvar,node->npoints);
free_imatrix(npL, node->npoints,nclasses);
free_imatrix(npR, node->npoints,nclasses);
free_dmatrix(prL, node->npoints,nclasses);
free_dmatrix(prR, node->npoints,nclasses);
free_dvector(tmpvar);
free_dvector(decrease_in_inpurity);
}
_Use_decl_annotations_
int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE, LPSTR, int nCmdShow)
{
D3D12HelloConstBuffers sample(1280, 720, L"D3D12 Raymarcher");
return sample.Run(hInstance, nCmdShow);
}
void compute_sift_keypoints(float *input, keypointslist& keypoints, int width, int height, siftPar &par)
{
flimage image;
/// Make zoom of image if necessary
float octSize = 1.0;
if (par.DoubleImSize){
//printf("... compute_sift_keypoints :: applying zoom\n");
// image.create(2*width, 2*height);
// apply_zoom(input,image.getPlane(),2.0,par.order,width,height);
// octSize *= 0.5;
printf("Doulbe image size not allowed. Guoshen Yu\n");
exit(-1);
} else
{
image.create(width,height,input);
}
// printf("Using initial Dog value: %f\n", par.PeakThresh);
// printf("Double image size: %d\n", par.DoubleImSize);
// printf("Interpolation order: %d\n", par.order);
/// Apply initial smoothing to input image to raise its smoothing to par.InitSigma.
/// We assume image from camera has smoothing of sigma = 0.5, which becomes sigma = 1.0 if image has been doubled.
/// increase = sqrt(Init^2 - Current^2)
float curSigma;
if (par.DoubleImSize) curSigma = 1.0; else curSigma = 0.5;
if (par.InitSigma > curSigma ) {
if (DEBUG) printf("Convolving initial image to achieve std: %f \n", par.InitSigma);
float sigma = (float) sqrt((double)(par.InitSigma * par.InitSigma - curSigma * curSigma));
gaussian_convolution( image.getPlane(), image.getPlane(), image.nwidth(), image.nheight(), sigma);
}
/// Convolve by par.InitSigma at each step inside OctaveKeypoints by steps of
/// Subsample of factor 2 while reasonable image size
/// Keep reducing image by factors of 2 until one dimension is
/// smaller than minimum size at which a feature could be detected.
int minsize = 2 * par.BorderDist + 2;
int OctaveCounter = 0;
//printf("... compute_sift_keypoints :: maximum number of scales : %d\n", par.OctaveMax);
while (image.nwidth() > minsize && image.nheight() > minsize && OctaveCounter < par.OctaveMax) {
if (DEBUG) printf("Calling OctaveKeypoints \n");
OctaveKeypoints(image, octSize, keypoints,par);
// image is blurred inside OctaveKeypoints and therefore can be sampled
flimage aux( (int)((float) image.nwidth() / 2.0f) , (int)((float) image.nheight() / 2.0f));
if (DEBUG) printf("Sampling initial image \n");
sample(image.getPlane(), aux.getPlane(), 2.0f, image.nwidth(), image.nheight());
image = aux;
octSize *= 2.0;
OctaveCounter++;
}
/* printf("sift:: %d keypoints\n", keypoints.size());
printf("sift:: plus non correctly localized: %d \n", par.noncorrectlylocalized);*/
}
void Sighter::procSight(cv::Mat &frame, cv::Mat &lfeat, cv::Mat &rfeat) {
cv::cvtColor(frame, drawMat, CV_BGRA2BGR);
cv::flip(drawMat, drawMat, 1);
cv::cvtColor(drawMat, grayMat, CV_BGR2GRAY);
cv::Rect frect;
cv::Mat& draw = drawMat;
cv::Mat& gray = grayMat;
//Detect face
cv::vector<cv::Rect> faces;
faceDetector.detectMultiScale(gray, faces, 1.1, 20, CV_HAAR_DO_CANNY_PRUNING|CV_HAAR_FIND_BIGGEST_OBJECT, cv::Size(gray.cols/4,gray.rows/4));
if(faces.empty()) {
return;
}
frect = faces[0];
cv::Mat face = gray(frect);
//Detect eye
cv::vector<cv::Rect> leyes,reyes;
cv::Size max_size = cv::Size(face.cols/2,face.rows/4);
cv::Size min_size = cv::Size(face.cols/10, 10);
cv::Mat lface = face(cv::Rect(0,face.rows/4,face.cols/2,face.rows/3));
eyeDetector.detectMultiScale(lface, leyes, 1.1, 20, CV_HAAR_DO_CANNY_PRUNING, min_size, max_size);
cv::Mat rface = face(cv::Rect(face.cols/2,face.rows/4,face.cols/2,face.rows/3));
eyeDetector.detectMultiScale(rface, reyes, 1.1, 20, CV_HAAR_DO_CANNY_PRUNING, min_size, max_size);
int szl = (int)leyes.size();
int szr = (int)reyes.size();
if(szl < 1 || szr < 1) {
return;
}
//have min vertical difference
int min_dy = INT_MAX;
cv::Rect lerect, rerect;
for(int i=0;i<szl;i++) {
int cy = leyes[i].y + leyes[i].height/2;
for(int j=0;j<szr;j++) {
int cyr = reyes[j].y + reyes[j].height/2;
int d = abs(cy-cyr);
if( d < min_dy) {
min_dy = d;
lerect = leyes[i];
rerect = reyes[j];
}
}
}
//transform to worldwide coordinate
lerect.x += frect.x;
lerect.y += frect.y + frect.height/4;
rerect.x += frect.x + frect.width/2;
rerect.y += frect.y + frect.height/4;
//classifing eye's action
cv::Mat legray = gray(lerect);
cv::Mat regray = gray(rerect);
cv::flip(regray, regray, 1);
lfeat = legray;
rfeat = regray;
cv::Mat sample(50,50,CV_8UC1);
cv::resize(legray, sample, sample.size());
int ldir = -1;
if(eyeActM)
ldir = eyeActM->predict(sample);
cv::resize(regray, sample, sample.size());
int rdir = -1;
if(eyeActM)
rdir = eyeActM->predict(sample);
SightState sstate = STARE_NONE;
if( ldir == 0 && rdir == 0 ) {
sstate = STARE_CENTER;
}
if( ldir == 2 && rdir == 1) {
sstate = STARE_LEFT;
}
if( ldir == 1 && rdir == 2) {
sstate = STARE_RIGHT;
}
{
#define ACT_STAT_TOTAL 10
static SightState sstates[ACT_STAT_TOTAL] = {STARE_NONE};
static int ssi = 0;
static int ssleft = 0;
static int ssright = 0;
static int sscenter = 0;
static int ssnone = ACT_STAT_TOTAL;
switch(sstates[ssi]) {
case STARE_NONE:
ssnone--;
break;
case STARE_LEFT:
ssleft--;
break;
case STARE_RIGHT:
ssright--;
break;
case STARE_CENTER:
sscenter--;
break;
default:
break;
}
switch(sstate) {
case STARE_NONE:
ssnone++;
if(ssnone >= ACT_STAT_TOTAL*0.6) {
sightState = STARE_NONE;
}
break;
case STARE_LEFT:
ssleft++;
if(ssleft >= ACT_STAT_TOTAL*0.6) {
sightState = STARE_LEFT;
}
break;
case STARE_RIGHT:
ssright++;
if(ssright >= ACT_STAT_TOTAL*0.6) {
sightState = STARE_RIGHT;
}
break;
case STARE_CENTER:
sscenter++;
if(sscenter >= ACT_STAT_TOTAL*0.6) {
sightState = STARE_CENTER;
}
break;
default:
break;
}
sstates[ssi] = sstate;
ssi = (ssi+1)%ACT_STAT_TOTAL;
}
//tracking the pupil
pulLeft.procPupil(legray);
pulRight.procPupil(regray);
pulLeft.pulCenter += lerect.tl();
pulRight.pulCenter += rerect.tl();
//Drawing
int r = MAX(pulLeft.pulRect.height, pulRight.pulRect.height)/3;
static int ani_r = 100;
static int ani_speed = 5;
if(ani_r > 0) {
cv::circle(draw, pulLeft.pulCenter, ani_r, CV_RGB(255,0,0),1.5);
cv::circle(draw, pulRight.pulCenter, ani_r, CV_RGB(255,0,0),1.5);
ani_r -= ani_speed;
ani_speed += 5;
}
else if( r > 0 ) {
cv::circle(draw, pulLeft.pulCenter, r, CV_RGB(255,0,0), CV_FILLED);
cv::circle(draw, pulLeft.pulCenter, r*3, CV_RGB(255,0,0), 1);
cv::circle(draw, pulRight.pulCenter, r, CV_RGB(255,0,0), CV_FILLED);
cv::circle(draw, pulRight.pulCenter, r*3, CV_RGB(255,0,0), 1);
}
cv::rectangle(draw, frect, CV_RGB(255,255,0));
cv::rectangle(draw, lerect, CV_RGB(0,255,0));
cv::rectangle(draw, rerect, CV_RGB(0,255,0));
}
_Use_decl_annotations_
int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE, LPSTR, int nCmdShow)
{
D3D12nBodyGravity sample(1280, 720, L"D3D12 n-Body Gravity Simulation");
return sample.Run(hInstance, nCmdShow);
}
// Sample a visible neuron, given a hidden neuron vector
double sample_visible(rbm* r, unsigned int i, std::vector<int> hidden) {
return sample(visible_probability(r, i, hidden));
}
void PropertiesDialog::setFontSample(const QFont & f)
{
fontSampleLabel->setFont(f);
QString sample("%1 %2 pt");
fontSampleLabel->setText(sample.arg(f.family()).arg(f.pointSize()));
}
kernel void sample_y(global int *cur_y,
global int *cur_z,
global int *z_by_ry,
global int *cur_r,
global int *obs,
global float *rand,
uint N, uint D, uint K, uint f_img_width,
float lambda, float epislon, float theta) {
const uint V_SCALE = 0, H_SCALE = 1, V_TRANS = 2, H_TRANS = 3, NUM_TRANS = 4;
uint kth = get_global_id(0); // k is the index of features
uint dth = get_global_id(1); // d is the index of pixels
uint f_img_height = D / f_img_width;
// unpack dth into h and w
uint h = dth / f_img_width;
uint w = dth % f_img_width;
// calculate the prior probability of each cell is 1
float on_loglik_temp = log(theta);
float off_loglik_temp = log(1 - theta);
int v_scale, h_scale, v_dist, h_dist, new_height, new_width, new_index, n, hh, ww;
// extremely hackish way to calculate the loglikelihood
for (n = 0; n < N; n++) {
// if the nth object has the kth feature
if (cur_z[n * K + kth] == 1) {
// retrieve the transformation applied to this feature by this object
v_scale = cur_r[n * (K * NUM_TRANS) + kth * NUM_TRANS + V_SCALE];
h_scale = cur_r[n * (K * NUM_TRANS) + kth * NUM_TRANS + H_SCALE];
v_dist = cur_r[n * (K * NUM_TRANS) + kth * NUM_TRANS + V_TRANS];
h_dist = cur_r[n * (K * NUM_TRANS) + kth * NUM_TRANS + H_TRANS];
new_height = f_img_height + v_scale;
new_width = f_img_width + h_scale;
// loop over all pixels
for (hh = 0; hh < f_img_height; hh++) {
for (ww = 0; ww < f_img_width; ww++) {
if ((int)round((float)hh / new_height * f_img_height) == h &
(int)round((float)ww / new_width * f_img_width) == w) {
new_index = ((v_dist + hh) % f_img_height) * f_img_width + (h_dist + ww) % f_img_width;
// if the observed pixel at dth is on
if (obs[n * D + new_index] == 1) { // transformed feature affects the pixel at new_index not dth, cf., ibp
// if the feature image previously has this pixel on
if (cur_y[kth * D + dth] == 1) { // this is dth instead of new_index because we are referring to the original y
on_loglik_temp += log(1 - pow(1 - lambda, z_by_ry[n * D + new_index]) * (1 - epislon));
off_loglik_temp += log(1 - pow(1 - lambda, z_by_ry[n * D + new_index] - 1) * (1 - epislon));
} else {
on_loglik_temp += log(1 - pow(1 - lambda, z_by_ry[n * D + new_index] + 1) * (1 - epislon));
off_loglik_temp += log(1 - pow(1 - lambda, z_by_ry[n * D + new_index]) * (1 - epislon));
}
} else { // else obs[n * D + new_index] == 0
on_loglik_temp += log(1 - lambda);
off_loglik_temp += log(1.0f);
}
}
}
}
}
}
float logpost[2] = {on_loglik_temp, off_loglik_temp};
//printf("%f %f %d \n", logpost[0], logpost[1], cur_y[kth * D + dth]);
uint labels[2] = {1, 0};
lognormalize(logpost, 0, 2);
cur_y[kth * D + dth] = sample(2, labels, logpost, 0, rand[kth * D + dth]);
//printf("%f %f %d \n", logpost[0], logpost[1], cur_y[kth * D + dth]);
}
dvec4 VolumeVectorSampler::sample(double x, double y, double z, double t) const {
return sample(dvec4(x, y, z, t));
}
_Use_decl_annotations_
int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE, LPSTR, int nCmdShow)
{
D3D12HelloBundles sample(1280, 720, L"D3D12 Hello Bundles");
return Win32Application::Run(&sample, hInstance, nCmdShow);
}
dvec4 VolumeVectorSampler::sample(const vec4 &pos) const { return sample(dvec4(pos)); }
vectord DiscreteModel::samplePoint()
{
randInt sample(mEngine, intUniformDist(0,mInputSet.size()-1));
return mInputSet[sample()];
};
pixel_t bilinear_sampler_t::sample( const Imath::V2d& p) const
{
int x = std::max( std::min( IECore::fastFloatFloor( p.x), src_area_.max.x - 1), src_area_.min.x);
int y = std::max( std::min( IECore::fastFloatFloor( p.y), src_area_.max.y - 1), src_area_.min.y);
return sample( x, y, p.x - x, p.y - y);
}