/* Krichevski-Trofimov estimated log probability accessor */
weight_t CTNode::logProbEstimated() const {
if (UseZeroRedundancy) {
if (m_count[0]+m_count[1] == 0) return 0.0;
double rval = log_point_five + m_log_prob_est;
if (m_count[0] == 0) rval = logAdd(log_quarter, rval);
if (m_count[1] == 0) rval = logAdd(log_quarter, rval);
return rval;
}
return m_log_prob_est;
}
/*
* CRF_StdSegStateNode::computeAlpha
*
* Read alpha vectors of previous nodes directly from prevNode and store the result of the alpha vector in alphaArray.
*
* Compute the alpha vector for the forward backward computation for this node.
*/
double CRF_StdSegStateNode::computeAlpha()
{
//QNUInt32 nLabs = this->crf_ptr->getNLabs();
this->alphaScale=0.0;
checkNumPrevNodes();
QNUInt32 clab = 0;
for (QNUInt32 dur = 1; dur <= this->numPrevNodes; dur++)
{
CRF_StateNode* prevAdjacentSeg = this->prevNodes[this->numPrevNodes - dur];
double* prev_adj_seg_alpha = prevAdjacentSeg->getAlpha();
for (QNUInt32 lab = 0; lab < this->nActualLabs; lab++)
{
this->logAddAcc[0]=prev_adj_seg_alpha[0]+this->transMatrix[0+clab];
double maxv=this->logAddAcc[0];
//TODO: for (QNUInt32 plab = 1; plab < prevAdjacentSeg->getNLabs(); plab++) { //full implementation. But not working now because logAdd() cannot do calculation on LOG0 yet.
for (QNUInt32 plab = 1; plab < prevAdjacentSeg->getNumAvailLabs(); plab++) { //faster implementation, not guaranteed to work for all classes of previous nodes.
//this->logAddAcc[plab]=prev_adj_seg_alpha[plab]+this->transMatrix[plab * prevAdjacentSeg->getNLabs() + clab];
this->logAddAcc[plab]=prev_adj_seg_alpha[plab]+this->transMatrix[plab * this->nLabs + clab];
if (this->logAddAcc[plab]>maxv) {
maxv=logAddAcc[plab];
}
}
try {
//TODO: this->alphaArray[clab]=logAdd(this->logAddAcc,maxv,prevAdjacentSeg->getNLabs()); //full implementation. But not working now because logAdd() cannot do calculation on LOG0 yet.
this->alphaArray[clab]=logAdd(this->logAddAcc,maxv,prevAdjacentSeg->getNumAvailLabs()); //faster implementation, not guaranteed to work for all classes of previous nodes.
}
catch (exception &e) {
string errstr="CRF_StdSegStateNode::computeAlpha() caught exception: "+string(e.what())+" while computing alpha";
throw runtime_error(errstr);
return(-1);
}
this->alphaArray[clab]+=this->stateArray[clab];
clab++;
}
}
// These are the cases when the current node serves as the beginning segment of the sequence, so there is no previous node.
for (QNUInt32 dur = this->numPrevNodes + 1; dur <= this->nodeLabMaxDur; dur++)
{
for (QNUInt32 lab = 0; lab < this->nActualLabs; lab++)
{
this->alphaArray[clab]=this->stateArray[clab];
clab++;
}
}
return this->alphaScale;
}
void LogSoftMax::frameForward(int t, real *f_inputs, real *f_outputs)
{
real sum = LOG_ZERO;
for(int i = 0; i < n_inputs; i++)
{
real z = f_inputs[i];
f_outputs[i] = z;
sum = logAdd(sum, z);
}
for(int i = 0; i < n_outputs; i++)
f_outputs[i] -= sum;
}
/*
* CRF_StdSegStateNode::computeAlphaAlignedSum
*
* Returns: Sum of the values in the alphaAligned vector of this node
*
* Used to compute the "soft" normalization constant.
*/
double CRF_StdSegStateNode::computeAlphaAlignedSum()
{
double Zx;
try {
//Zx=logAdd(&(this->alphaArrayAligned),this->nLabs); // full implementation. But not working now because logAdd() cannot do calculation on LOG0 yet.
Zx=logAdd(&(this->alphaArrayAligned),this->numAvailLabs);
}
catch (exception& e) {
//changed by Ryan
//string errstr="CRF_StdStateNodeLog::computeExpF() threw exception: "+string(e.what());
string errstr="CRF_StdStateNode::computeAlphaAlignedSum() threw exception: "+string(e.what());
throw runtime_error(errstr);
}
return Zx;
}
void IPFP::computePredictedPeakHeights( Message &out, spec_factor_t &log_spec_factor, Message &marginals){
out.reset( log_spec_factor.size() );
for( unsigned int i = 0; i < log_spec_factor.size(); i++ ){
Message::const_iterator it = marginals.begin();
double log_sum = NULL_PROB;
for( ; it != marginals.end(); ++it ){
double tmp = *it;
tmp += log_spec_factor[i][it.index()];
log_sum= logAdd( log_sum, tmp );
}
out.addToIdx(i,log_sum);
}
}
void IPFP::setDesiredMarginals( std::vector<Message> &desired_margs, std::vector<Message> &marginals ){
unsigned int num_spectra = cfg->dv_spectrum_depths.size();
//Initialise factor probabilities between peaks and fragments
//and peak targets
std::vector<Message> peak_targets( num_spectra );
std::vector<spec_factor_t> log_spec_factors( num_spectra );
for( unsigned int energy = 0; energy < num_spectra; energy++ ){
int spec_idx = cfg->dv_spectrum_indexes[energy];
initLogSpecFactor( log_spec_factors[energy], moldata->getSpectrum(spec_idx) );
initPeakTargets( peak_targets[energy], moldata->getSpectrum(spec_idx) );
}
//Set desired marginals
Message pk_heights;
desired_margs.resize(num_spectra);
for( unsigned int energy = 0; energy < num_spectra; energy++ ){
//Adjust the fragment marginals so that they sum to the correct target for each
//peak yet maintain the correct ratio between fragments of similar mass
//given prior likelihoods
computePredictedPeakHeights( pk_heights, log_spec_factors[energy], marginals[energy] );
//Use these to find the desired marginals for each fragment
desired_margs[energy].reset(marginals[energy].size());
Message::const_iterator it = marginals[energy].begin();
for( ; it != marginals[energy].end(); ++it ){
double log_sum = NULL_PROB;
for( unsigned int i = 0; i < pk_heights.size(); i++ ){
if( pk_heights.getIdx(i) > -10000.0 ){ //Ignore peaks that can't be realistically achieved
double tmp = peak_targets[energy].getIdx(i);
tmp -= pk_heights.getIdx(i);
log_sum = logAdd( log_sum, tmp + log_spec_factors[energy][i][it.index()]);
}
}
log_sum += *it;
desired_margs[energy].addToIdx(it.index(), log_sum);
}
}
}
/*
* CRF_StdSegStateNode::computeBeta
*
* Inputs: scale - scaling constant for result_beta array
*
* Returns:
*
* Read the beta vectors of next nodes directly from nextNode and store the result of the beta vector in betaArray.
*
* Compute the beta vector for the node before this one and store it in result_beta
*/
double CRF_StdSegStateNode::computeBeta(double scale)
{
// Logic desired:
// * Compute beta_i[size of alpha[]+1] to be all 1s
// * Multiply M_i[current] by beta_i[current+1] to get beta_i[current]
// QNUInt32 nLabs = this->crf_ptr->getNLabs();
//
// for (QNUInt32 clab=0; clab<nLabs; clab++) {
// this->tempBeta[clab]=this->betaArray[clab]+this->stateArray[clab];
// }
//
// for (QNUInt32 plab=0; plab<nLabs; plab++) {
// this->logAddAcc[0]=this->transMatrix[plab*nLabs+0]+this->tempBeta[0];
// double maxv=this->logAddAcc[0];
// for (QNUInt32 clab=1; clab<nLabs; clab++) {
// this->logAddAcc[clab]=this->transMatrix[plab*nLabs+clab]+this->tempBeta[clab];
// if (this->logAddAcc[clab]>maxv) {
// maxv=this->logAddAcc[clab];
// }
// }
// try {
// result_beta[plab]=logAdd(this->logAddAcc,maxv,nLabs);
// }
// catch (exception &e) {
// string errstr="CRF_StdSegStateNode::computeBeta() caught exception: "+string(e.what())+" while computing beta";
// throw runtime_error(errstr);
// return(-1);
// }
// }
checkNumNextNodes();
// if numNextNodes == 0, this is the last node of the sequence.
// Sets the beta value in this node to the special case for the end of the sequence.
if (this->numNextNodes == 0)
{
setTailBeta();
return this->alphaScale;
}
QNUInt32 nextlab = 0;
for (QNUInt32 dur = 1; dur <= this->numNextNodes; dur++)
{
CRF_StateNode* nextAdjacentSeg = this->nextNodes[dur - 1];
double* next_adj_seg_beta = nextAdjacentSeg->getBeta();
for (QNUInt32 lab = 0; lab < this->nActualLabs; lab++) {
this->tempBeta[nextlab] = next_adj_seg_beta[nextlab] + nextAdjacentSeg->getStateValue(nextlab);
nextlab++;
}
}
for (QNUInt32 clab = 0; clab < this->numAvailLabs; clab++)
{
CRF_StateNode* nextAdjacentSeg = this->nextNodes[0];
//this->transMatrix[plab*nLabs+0]+this->tempBeta[0]
this->logAddAcc[0] = nextAdjacentSeg->getTransValue(clab, 0) + this->tempBeta[0];
double maxv=this->logAddAcc[0];
QNUInt32 nextlab = 0;
for (QNUInt32 dur = 1; dur <= this->numNextNodes; dur++)
{
nextAdjacentSeg = this->nextNodes[dur - 1];
//TODO: It should be nActualLabs_of_nextNode instead of nActualLabs_of_thisNode as the number of iterations.
//for (QNUInt32 lab = 0; lab < this->nActualLabs; lab++) {
for (QNUInt32 lab = 0; lab < this->nextNodeNActualLabs; lab++) {
//this->logAddAcc[clab]=this->transMatrix[plab*nLabs+clab]+this->tempBeta[clab];
this->logAddAcc[nextlab] = nextAdjacentSeg->getTransValue(clab, nextlab) + this->tempBeta[nextlab];
if (this->logAddAcc[nextlab]>maxv) {
maxv=logAddAcc[nextlab];
}
nextlab++;
}
}
try {
//TODO:It should be nActualLabs_of_nextNode instead of nActualLabs_of_thisNode.
//this->betaArray[clab]=logAdd(this->logAddAcc,maxv,this->nActualLabs * this->numNextNodes);
this->betaArray[clab]=logAdd(this->logAddAcc,maxv,this->nextNodeNActualLabs * this->numNextNodes);
}
catch (exception &e) {
string errstr="CRF_StdSegStateNode::computeBeta() caught exception: "+string(e.what())+" while computing beta";
throw runtime_error(errstr);
return(-1);
}
}
return this->alphaScale;
}
void Gear_VideoTexture::onUpdateSettings()
{
// Timer timer;
// timer.reset();
char tempstr[1024];
// XXX todo : parameters
double power = 1.0;
int nEpochs = 3;
double alpha = 0.999;
pthread_mutex_lock(_mutex);
// Initialize (open) the movie.
strcpy(tempstr,_settings.get(SETTING_FILENAME)->valueStr().c_str());
std::cout << "opening movie : " << tempstr << std::endl;
if (_file!=NULL)
mpeg3_close(_file);
_file = mpeg3_open(tempstr);
if (_file==NULL)
{
std::cout << "error opening movie : " << tempstr << std::endl;
pthread_mutex_unlock(_mutex);
return;
}
_sizeX = mpeg3_video_width(_file, 0);
_sizeY = mpeg3_video_height(_file, 0);
//_nFrames = (int)CLAMP((long int)_settings.get(SETTING_NFRAMES)->valueInt(), 1l, mpeg3_video_frames(_file, 0));
_nFrames = (int)_settings.get(SETTING_NFRAMES)->valueInt();
ASSERT_ERROR(_nFrames >= 1);
std::cout << "movie size X : " << _sizeX << std::endl;
std::cout << "movie size Y : " << _sizeY << std::endl;
std::cout << "numframes : " << _nFrames << " / " << mpeg3_video_frames(_file, 0) << std::endl;
std::cout << "movie samplerate : " << mpeg3_sample_rate(_file,0) << std::endl;
for (int i=0;i<_sizeY-1;i++)
_frame[i] = (RGBA*) realloc(_frame[i], _sizeX * sizeof(RGBA));
//from the doc :
//You must allocate 4 extra bytes in the last output_row. This is scratch area for the MMX routines.
_frame[_sizeY-1] = (RGBA*) realloc(_frame[_sizeY-1], (_sizeX * sizeof(RGBA)) + 4);
// std::cout << "Time to initialize things: " << timer.getTime() << std::endl;
// timer.reset();
// Fill sequences and distance matrix.
_distances.resize(_nFrames, _nFrames);
_sequences.resize(_nFrames);
_size = _sizeX * _sizeY;
NOTICE("Filling distance and sequence matrix.");
for (int i=0; i<_nFrames; ++i)
{
Array2D<RGBA>& currImage = _sequences[i];
currImage.resize(_sizeX, _sizeY);
// Read current image.
mpeg3_read_frame(_file, (unsigned char**)_frame, 0, 0, _sizeX, _sizeY, _sizeX, _sizeY, MPEG3_RGBA8888, 0);
// Add image to sequences.
for(int y=0;y<_sizeY;y++)
memcpy(currImage.row(y), _frame[y], sizeof(RGBA) * _sizeX);
// Update distance matrix (this is the bottleneck of the whole algorithm).
_distances(i,i) = 0.0;
for (int j=0; j<i; ++j)
_distances(i,j) = _distances(j,i) = L2((unsigned char*)currImage.data(), (unsigned char*)_sequences[j].data(), (size_t)_size*SIZE_RGBA);
}
// std::cout << "Time to fill sequence and distance: " << timer.getTime() << std::endl;
// timer.reset();
#if DEBUG_NOTICE
std::cout << "Distances: " << std::endl;
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
std::cout << _distances(j,i) << " ";
std::cout << std::endl;
}
std::cout << std::endl;
#endif
// Add temporal coherence to distance matrix by smoothing the distance with linear interpolation.
NOTICE("Computing smoothed distances.");
int timeWindowLength = CLAMP(_settings.get(SETTING_TIMEWINDOWLENGTH)->valueInt(), 0, 2);
_smoothedDistances.resize(_nFrames, _nFrames);
switch (timeWindowLength)
{
case 0:
memcpy(_smoothedDistances.data(), _distances.data(), _distances.size());
break;
case 1:
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
{
int iPrev = MAX(i-1,0);
int jPrev = MAX(j-1,0);
_smoothedDistances(j,i) =
0.5 * _distances(jPrev,iPrev) +
0.5 * _distances(j,i);
}
}
break;
case 2:
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
{
int iPrev = MAX(i-1,0);
int jPrev = MAX(j-1,0);
int iNext = MIN(i+1,_nFrames-1);
int jNext = MIN(j+1,_nFrames-1);
int iPrev2 = MAX(i-2,0);
int jPrev2 = MAX(j-2,0);
_smoothedDistances(j,i) =
0.125 * _distances(jPrev2,iPrev2) +
0.375 * _distances(jPrev,iPrev) +
0.375 * _distances(j,i) +
0.125 * _distances(jNext,iNext);
}
}
break;
default:;
error("Wrong time window length specified, please check");
}
// std::cout << "Time to compute smoothed distances: " << timer.getTime() << std::endl;
// timer.reset();
#if DEBUG_NOTICE
std::cout << "Smoothed distances: " << std::endl;
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
std::cout << _smoothedDistances(i,j) << " ";
std::cout << std::endl;
}
std::cout << std::endl;
#endif
// Using Q-learning, recompute the matrix of distances.
NOTICE("Computing final set of distances using Q-learning.");
_minDistances.resize(_nFrames);
// Initialize distances.
for (int i=0; i<_nFrames; ++i)
for (int j=0; j<=i; ++j)
_distances(i,j) = _distances(j,i) = _smoothedDistances(i,j) = _smoothedDistances(j,i) = pow(_smoothedDistances(i,j), power);
// Q-learning.
for (int t=0; t<nEpochs; ++t)
{
#if DEBUG_NOTICE
std::cout << "Q-learn distances step " << t << " : " << std::endl;
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
std::cout << _distances(j,i) << " ";
std::cout << std::endl;
}
std::cout << std::endl;
#endif
// Init min distances.
for (int j=0; j<_nFrames; ++j)
_minDistances[j] = min(_distances.row(j), (size_t)_nFrames);
for (int i=_nFrames-1; i>=0; --i)
for (int j=0; j<_nFrames; ++j)
{
// Update distances.
_distances(j,i) = _smoothedDistances(j,i) + alpha * _minDistances[j];
// Update min distances.
_minDistances[j] = min(_distances.row(j), (size_t)_nFrames);
}
}
// std::cout << "Time to compute Q-learned distances: " << timer.getTime() << std::endl;
// timer.reset();
// Calculate the mean (smoothed) distance.
double meanDistance = sum(_distances.data(), _distances.size()) / (double)_distances.size();
// Compute cumulative probabilities.
_logCumProbs.resize(_nFrames, _nFrames);
for (int i=0; i<_nFrames; ++i)
{
// Compute logCumProbs.
_logCumProbs(0,i) = LOG_ZERO;
for (int j=1; j<_nFrames; ++j)
_logCumProbs(j,i) = logAdd(_logCumProbs(j-1,i), -_distances(j,i) / meanDistance);
// Normalize to make a true probability.
double norm = _logCumProbs(_nFrames-1,i);
for (int j=0; j<_nFrames; ++j)
_logCumProbs(j,i) -= norm;
}
// std::cout << "Time to compute probabilities: "<< timer.getTime() << std::endl;
#if DEBUG_NOTICE
std::cout << "Distances: " << std::endl;
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
std::cout << _distances(j,i) << " ";
std::cout << std::endl;
}
std::cout << std::endl;
std::cout << "Probabilities: " << std::endl;
for (int i=0; i<_nFrames; ++i)
{
for (int j=0; j<_nFrames; ++j)
std::cout << exp(_logCumProbs(j,i)) << " ";
std::cout << std::endl;
}
std::cout << std::endl;
#endif
_currentFrame = 0;
pthread_mutex_unlock(_mutex);
}
double IPFP::computeBeliefs( ){
double diff = 0.0;
beliefs_t tmp_beliefs;
std::vector<double> norms;
norms.resize(cfg->model_depth);
const FragmentGraph *fg = moldata->getFragmentGraph();
//Compute Persistence Beliefs (and track norms)
tmp_beliefs.ps.resize(fg->getNumFragments());
for( unsigned int i = 0; i < fg->getNumFragments(); i++ ){
tmp_beliefs.ps[i].resize(cfg->model_depth);
for( unsigned int d = 0; d < cfg->model_depth; d++ ){
double tmp;
if( (d == 0 && i == 0) || (d > 0 && down_msgs[d-1].getIdx(i) > -DBL_MAXIMUM) ){
tmp = fprobs.ps[i][d];
if( d < cfg->model_depth-1 )
tmp += up_msgs[d].getIdx(i);
if( d > 0 ){
double msg_val = down_msgs[d-1].getIdx(i);
if( msg_val > -DBL_MAXIMUM )
tmp += msg_val;
else tmp = NULL_PROB;
}
if ( i == 0 ) norms[d] = tmp;
else norms[d] = logAdd(norms[d], tmp);
}else tmp = NULL_PROB;
tmp_beliefs.ps[i][d] = tmp;
}
}
//Compute Transition Beliefs (and track norms)
tmp_beliefs.tn.resize(fg->getNumTransitions());
for( unsigned int i = 0; i < fg->getNumTransitions(); i++ ){
const Transition *t = fg->getTransitionAtIdx(i);
tmp_beliefs.tn[i].resize(cfg->model_depth);
for( unsigned int d = 0; d < cfg->model_depth; d++ ){
double tmp;
if( (d == 0 && t->getFromId() == 0) || (d > 0 && down_msgs[d-1].getIdx(t->getFromId()) > -DBL_MAXIMUM) ){
tmp= fprobs.tn[i][d];
if( d < cfg->model_depth-1 )
tmp += up_msgs[d].getIdx(t->getToId());
if( d > 0 ){
double msg_val = down_msgs[d-1].getIdx(t->getFromId());
if( msg_val > -DBL_MAXIMUM )
tmp += msg_val;
else tmp = NULL_PROB;
}
norms[d] = logAdd(norms[d], tmp);
}else tmp = NULL_PROB;
tmp_beliefs.tn[i][d] = tmp;
}
}
//Normalise
for( unsigned int i = 0; i < tmp_beliefs.tn.size(); i++ ){
for( unsigned int d = 0; d < cfg->model_depth; d++ ){
tmp_beliefs.tn[i][d] -= norms[d];
}
}
for( unsigned int i = 0; i < tmp_beliefs.ps.size(); i++ ){
for( unsigned int d = 0; d < cfg->model_depth; d++ ){
tmp_beliefs.ps[i][d] -= norms[d];
}
}
//Replace the old beliefs with the new ones, tracking the total difference
diff = copyBeliefsAndComputeDiff( &tmp_beliefs );
return diff;
}
