/** Update centroid images for 1 channel
* @param d
* @param images
* @param weighParameter
*/
static void _Cig_UpdateCentroidImages(Destin * d,
float *** images, // preallocated images to update
float weightParameter, // higher value means higher contrast
int channel
){
int child_section, i, l, c;
_Cig_UpdateLayerZeroImages(d, images, channel);
int child_image_width = sqrt(GetNodeFromDestin(d, 0, 0, 0)->ni);
for(l = 1 ; l < d->nLayers ; l++){
Node * n = GetNodeFromDestin(d, l, 0, 0);
float ** combined_images;
MALLOC(combined_images, float *, 4);
for(i = 0 ; i < 4 ; i++){
MALLOC(combined_images[i], float, child_image_width * child_image_width);
}
for(c = 0 ; c < d->nb[l]; c++){
// combine the child layer's shared centroids according
// to the weights given in
for(child_section = 0 ; child_section < 4 ; child_section++){
// For the current sub section of the current centroid,
// generate its representative image and store it in
// the appropriate section of combined_images;
_Cig_BlendImages(images[l - 1],
&n->mu[c][child_section * d->nb[l - 1]],
d->nb[l - 1],
child_image_width * child_image_width,
weightParameter,
combined_images[child_section]);
}
Cig_ConcatImages(combined_images, child_image_width, child_image_width, images[l][c] );
}
for(i = 0 ; i < 4 ; i++){
FREE(combined_images[i]);
combined_images[i] = NULL;
}
FREE(combined_images);
combined_images = NULL;
child_image_width = child_image_width * 2;
}
return;
}
void DestinNetworkAlt::Uniform_addRescaledCentroids(Destin *destin,int layer,int iter,int rescaleiter,int toplayer)
{
int row,col,width;
int i,j,k,l,m;
int index=0;
float newmu[50000];
Node *currnode,*destnode;
std::vector<float> v;
v.clear();
currnode=GetNodeFromDestin(destin,toplayer,0,0);
i=layer;
{
width=sqrt(destin->layerSize[i]);
row=0;col=0;
for(row=0;row<width;row++)
{
for(col=0;col<width;col++)
{
destnode=GetNodeFromDestin(destin,i,row,col);
for(j=0;j<currnode->nb;j++)
{
v=rescaleCentroid(toplayer,j,i);
if(iter>rescaleiter)
{
DeleteUniformCentroid(destin,i,j);
}
memset(newmu,0,sizeof(newmu));
for(k=0;k<v.size();k++)
{
newmu[k]=v[k];
}
AddRescaledCentroids(destin,i,newmu,0,index);
}
}
}
}
destin->rescale=1;
}
// CZT: uniform;
void Uniform_ApplyDeltas(Destin * d, uint layer, float ** layerSharedSigma){
uint c, s, ns;
float diff, learnRate, dt;
//iterate over shared centroids
Node * n = GetNodeFromDestin(d, layer, 0,0);
for(c = 0 ; c < d->nb[layer]; c++){
//use learning strategy function pointer to get the learning rate
learnRate = d->centLearnStratFunc(d, NULL, layer, c);
//get the first node of the current layers
ns = n->ns;
for(s = 0 ; s < ns ; s++){
//move the centroid with the averaged delta
dt = d->uf_avgDelta[layer][c][s];
diff = dt * learnRate;
n->mu[c][s] += diff; //all nodes in a layer share this n->mu pointer
#ifdef CHECK_BIG_MU
if ( n->mu[c][s] > 1.0){
oops("Big mu value:%e at line %i\n",n->mu[c][s],__LINE__ );
}
#endif
n->muSqDiff += diff * diff; //only 0th node of each layer gets a muSqDiff
//TODO: write unit test for layerSharedSigma
//layerSharedSigma[c * ns + s] += n->beta * (dt * dt - layerSharedSigma[c * ns + s]);
layerSharedSigma[c][s] += n->beta * (d->uf_avgSquaredDelta[layer][c][s] - layerSharedSigma[c][s]);
}
}
for(s = 0 ; s < ns ; s++){
d->uf_absvar[layer][s] += n->beta * (d->uf_avgAbsDelta[layer][s] - d->uf_absvar[layer][s]);
}
return;
}
void Uniform_AddNewCentroids(Destin *d)
{
uint l, i;
float absvar;
for (l = 0; l < d->nLayers; l++)
{
// layer not trained or belief dimensionality reached
if (d->layerMask[l] == 0 || d->nb[l] >= d->layerMaxNb[l])
{
continue;
}
Node * n = GetNodeFromDestin(d, l, 0, 0);
absvar = 0;
for (i = 0; i < n->ns; i++)
{
absvar += d->uf_absvar[l][i];
}
absvar /= n->ns;
float rnd = (float) rand() / (float) RAND_MAX;
if (rnd < d->addCoeff * absvar)
{
AddUniformCentroid(d, l);
}
}
}
void Uniform_AddRescale(Destin *d,float *rcentroid,int destlayer,int srclayer)
{
uint l, i,j;
float absvar;
l=destlayer;
//for(l=srclayer-1;l>=0;l--)
{
if (d->layerMask[l] == 0 || d->nb[l] >= d->layerMaxNb[l])
{
//continue;
return;
}
Node *n=GetNodeFromDestin(d, l, 0, 0);
absvar=0;
for (i = 0; i < n->ns; i++)
{
absvar += d->uf_absvar[l][i];
}
absvar /= n->ns;
float rnd = (float) rand() / (float) RAND_MAX;
//float rnd=0.654;
if (rnd < d->addCoeff * absvar)
{
//for(j=0;j<min(n->nb,d->nb[srclayer]);j++)
{
AddRescaledCentroids(d,l,rcentroid,0,n->ns+1);
}
}
}
}
void _generateNewUniformCentroidMu(Destin *d, uint l, uint vicinity, float * newMu)
{
uint i, j, k, ni, nb, np, ns;
float sigma;
Node * n = GetNodeFromDestin(d, l, 0, 0);
ni = n->ni;
nb = n->nb;
np = n->np;
ns = n->ns;
float weights[nb];
uint picked[vicinity];
float pickedWeights[vicinity];
// calculate weights
float sumWeight = 0;
for (i = 0; i < nb; i++)
{
sigma = 0;
for (j = 0; j < ns; j++)
{
sigma += d->uf_sigma[l][i][j];
}
weights[i] = sigma; // d->uf_winFreqs[l][i] * sigma;
sumWeight += weights[i];
}
// pick centroids with probability proportional to weights
float sumPickedWeights = 0;
for (i = 0; i < vicinity; i++)
{
float pickWeight = 0;
float rnd = sumWeight * rand() / (float) RAND_MAX;
for (j = 0; (j < nb) && (rnd >= pickWeight); j++)
{
pickWeight += weights[j];
}
picked[i] = j - 1;
pickedWeights[i] = rand() / (float) RAND_MAX;
sumPickedWeights += pickedWeights[i];
}
float centerMu[ns];
float centerSigma[ns];
for (i = 0; i < ns; i++)
{
centerMu[i] = 0;
centerSigma[i] = INIT_SIGMA;
for (j = 0; j < vicinity; j++)
{
centerMu[i] += d->uf_mu[l][picked[j]][i] * pickedWeights[j] / sumPickedWeights;
}
}
_sampleNormalDistributedCentroid(newMu, centerMu, centerSigma, ns);
// TODO: in recurrent mode take care about position nb+1
_normalizeMu(newMu, ni, nb+1, np);
}
void DestinNetworkAlt::printWinningCentroidGrid(int layer){
uint width = (uint)sqrt(destin->layerSize[layer]);
uint nidx = 0;
Node * nodelist = GetNodeFromDestin(destin, layer, 0, 0);
printf("Winning centroids grid - layer %i:\n", layer);
for( int r = 0 ; r < width ; r++){
for(int c = 0 ; c < width ; c++ , nidx++){
printf("%i ", nodelist[nidx].winner);
}
printf("\n");
}
}
static void _Cig_UpdateLayerZeroImages(Destin * d,
float *** images, // preallocated images to update
int channel
){
int pixel, c;
Node * n = GetNodeFromDestin(d, 0, 0, 0);
for(c = 0 ; c < d->nb[0]; c++){
for(pixel = 0 ; pixel < n->ni ; pixel++){
//images[layer][centroid][pixel]
images[0][c][pixel] = n->mu[c][pixel]; // layer 0 centroids directly represent their images
}
if(channel > 0){
for(pixel=0; pixel < n->ni; ++pixel){
images[0][c][pixel] = n->mu[c][n->ni*channel + n->nb + n->np + pixel];
}
}
}
}
void DestinNetworkAlt::load(const char * fileName){
if(centroidImages != NULL){
Cig_DestroyCentroidImages(destin, centroidImages);
centroidImages = NULL;
}
destin = LoadDestin(destin, fileName);
if(destin == NULL){
throw std::runtime_error("load: could not open destin file.\n");
}
Node * n = GetNodeFromDestin(destin, 0, 0, 0);
this->gamma = n->gamma;
this->isUniform = destin->isUniform;
this->lambdaCoeff = n->lambdaCoeff;
if(temperatures != NULL){
delete [] temperatures;
}
temperatures = new float[getLayerCount()];
memcpy(temperatures, destin->temp, sizeof(float) * getLayerCount());
imageMode = extRatioToImageMode(destin->extRatio);
}
int Cig_GetCentroidImageWidth(Destin * d, int layer){
return sqrt(GetNodeFromDestin(d, 0, 0, 0)->ni) * pow(2, layer);
}
void Cig_UpdateCentroidImages(Destin * d,
float *** images, // preallocated images to update
float weighParameter // higher value means higher contrast
){
int p, l, c, prev_image_width,image_width;
Node * n;
image_width = sqrt(GetNodeFromDestin(d, 0, 0, 0)->ni);
prev_image_width = image_width;
for(l = 0 ; l < 1 ; l++){
n = GetNodeFromDestin(d, l, 0, 0);
for(c = 0 ; c < d->nb[l]; c++){
for(p = 0 ; p < n->ni ; p++){
images[l][c][p] = n->mu[c * n->ns + p];
}
int nRatio;
for(nRatio=1; nRatio < d->extRatio; ++nRatio)
{
for(p=0; p<n->ni; ++p)
{
images[l][c][n->ni*nRatio + p] = n->mu[c*n->ns + n->ni*nRatio + n->nb + n->np + p];
}
}
}
prev_image_width = image_width;
image_width *= 2;
}
int child_section, i;
for(l = 1 ; l < d->nLayers ; l++){
n = GetNodeFromDestin(d, l, 0, 0);
float ** combined_images;
MALLOC(combined_images, float *, 4);
for(i = 0 ; i < 4 ; i++){
MALLOC(combined_images[i], float, prev_image_width * prev_image_width);
}
for(c = 0 ; c < d->nb[l]; c++){
// combine the child layer's shared centroids according
// to the weights given in
for(child_section = 0 ; child_section < 4 ; child_section++){
// For the current sub section of the current centroid,
// generate its representative image and store it in
// the appropriate section of combined_images;
Cig_BlendImages(images[l - 1],
&n->mu[c * n->ns + child_section * d->nb[l - 1]],
d->nb[l - 1],
prev_image_width * prev_image_width,
weighParameter,
combined_images[child_section],
l,
d->extRatio);
}
Cig_ConcatImages(combined_images, prev_image_width, prev_image_width, images[l][c] );
}
for(i = 0 ; i < 4 ; i++){
FREE(combined_images[i]);
combined_images[i] = NULL;
}
FREE(combined_images);
combined_images = NULL;
prev_image_width = image_width;
image_width *= 2;
}
return;
}
// The method used to distribute values of beliefs associated with centroid that
// is going to be deleted. It helps when deleted centroid is very close to good
// learned centroids. This centroids from neighbourhood get additional weighted
// belief of deleted centroid where more weights have centroids that are closer.
void _distributeEvidenceOfDeletedCentroidToNeighbours(Destin *d, uint l, uint idx, uint nearest)
{
uint i, j, k, nb, ns;
Node * n = GetNodeFromDestin(d, l, 0, 0);
nb = n->nb;
ns = n->ns;
if (nb < 2)
{
fprintf (stderr, "DistributeWeightsOfDeletedCentroid(): called when the layer %d has only single centroid!", l);
return;
}
_Neighbour neighbours[nb];
for (i = 0; i < nb; i++)
{
neighbours[i].index = i;
neighbours[i].distance = _calcNeighboursDistanceEuc(d->uf_mu[l][i], d->uf_mu[l][idx], ns) + EPSILON;
neighbours[i].weight = 0;
}
qsort((void *)neighbours, nb, sizeof(_Neighbour), (void *)&_compareNeighboursByDistance);
nearest = _MIN(nearest, nb-1);
float sumDistance = 0;
for (i = 1; i <= nearest; i++)
{
sumDistance += neighbours[i].distance;
}
float sumWeights = 0;
for (i = 1; i <= nearest; i++)
{
neighbours[i].weight = sumDistance / neighbours[i].distance;
sumWeights += neighbours[i].weight;
}
for (i = 1; i <= nearest; i++)
{
neighbours[i].weight /= sumWeights;
}
// Layer l+1
if (l+1 < d->nLayers)
{
n = GetNodeFromDestin(d, l+1, 0, 0);
for (i = 0; i < n->nb; i++)
{
for (j = 0; j < n->nChildren; j++)
{
for (k = 1; k <= nearest; k++)
{
d->uf_mu[l+1][i][nb*j + neighbours[k].index] += d->uf_mu[l+1][i][nb*j + idx] * neighbours[k].weight;
d->uf_sigma[l+1][i][nb*j + neighbours[k].index] += d->uf_sigma[l+1][i][nb*j + idx] * neighbours[k].weight;
}
// centroid evidence has been distributed
d->uf_mu[l+1][i][nb*j + idx] = 0;
d->uf_sigma[l+1][i][nb*j + idx] = 0;
}
}
}
// TODO: RECURRENT MODE: move evidence for l layer and l-1 layer
}
void AddUniformCentroid(Destin *d, uint l)
{
uint i, j, ni, nb, np, ns, idx;
Node * n = GetNodeFromDestin(d, l, 0, 0);
nb = n->nb;
ni = n->ni;
np = n->np;
ns = n->ns;
idx = ni + nb;
// Layer l
float * newMu, * newSigma, * newAvgDelta, * newAvgSquaredDelta;
MALLOCV(newMu, float, (ns+1));
MALLOCV(newSigma, float, (ns+1));
MALLOCV(newAvgDelta, float, (ns+1));
MALLOCV(newAvgSquaredDelta, float, (ns+1));
_generateNewUniformCentroidMu(d, l, VICINITY_OF_ADDED_CENTROID, newMu);
for (j=0; j < ns+1; j++)
{
newSigma[j] = INIT_SIGMA;
newAvgDelta[j] = 0;
newAvgSquaredDelta[j] = 0;
}
for (i = 0; i < nb; i++)
{
ArrayInsertFloat(&d->uf_mu[l][i], ns, idx, 0);
_normalizeMu(d->uf_mu[l][i], ni, nb+1, np);
ArrayInsertFloat(&d->uf_sigma[l][i], ns, idx, INIT_SIGMA);
ArrayInsertFloat(&d->uf_avgDelta[l][i], ns, idx, 0);
ArrayInsertFloat(&d->uf_avgSquaredDelta[l][i], ns, idx, 0);
}
ArrayInsertFloat(&d->uf_absvar[l], ns, idx, 0);
ArrayInsertFloat(&d->uf_avgAbsDelta[l], ns, idx, 0);
ArrayAppendPtr((void *)&d->uf_mu[l], nb, newMu);
ArrayAppendPtr((void *)&d->uf_sigma[l], nb, newSigma);
ArrayAppendUInt(&d->uf_winCounts[l], nb, 0);
ArrayAppendLong(&d->uf_persistWinCounts[l], nb, 0);
ArrayAppendLong(&d->uf_persistWinCounts_detailed[l], nb, 0);
ArrayAppendPtr((void *)&d->uf_avgDelta[l], nb, newAvgDelta);
ArrayAppendPtr((void *)&d->uf_avgSquaredDelta[l], nb, newAvgSquaredDelta);
ArrayAppendFloat(&d->uf_starv[l], nb, 1);
ArrayAppendFloat(&d->uf_winFreqs[l], nb, 1/(float) nb);
_normalizeFloatArray(d->uf_winFreqs[l], nb+1);
d->nb[l]++; // increase global number of centroids for layer l
for (j = 0; j < d->layerSize[l]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l] + j];
ArrayAppendFloat(&n->belief, nb, 0);
ArrayAppendFloat(&n->beliefEuc, nb, 0);
ArrayAppendFloat(&n->beliefMal, nb, 0);
ArrayAppendFloat(&n->outputBelief, nb, 0);
ArrayInsertFloat(&n->delta, ns, idx, 0);
ArrayInsertFloat(&n->observation, ns, idx, 0);
// increase centroid dimensionality for each node from layer l
UpdateNodeSizes(n, n->ni, nb+1, n->np, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l];
n->starv = d->uf_starv[l];
}
// Layer l+1
if (l+1 < d->nLayers)
{
n = GetNodeFromDestin(d, l+1, 0, 0);
ns = n->ns;
uint childIndexes[n->nChildren]; // indexes of added centroids for all childs
float childValues[n->nChildren]; // initial values for all childs
float childSigmas[n->nChildren];
for (i = 0; i < n->nChildren; i++)
{
childIndexes[i] = (i+1)*nb;
childValues[i] = 0;
childSigmas[i] = INIT_SIGMA;
}
for (i = 0; i < n->nb; i++)
{
ArrayInsertFloats(&d->uf_mu[l+1][i], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&d->uf_sigma[l+1][i], ns, childIndexes, childSigmas, n->nChildren);
ArrayInsertFloats(&d->uf_avgDelta[l+1][i], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&d->uf_avgSquaredDelta[l+1][i], ns, childIndexes, childValues, n->nChildren);
}
ArrayInsertFloats(&d->uf_absvar[l+1], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&d->uf_avgAbsDelta[l+1], ns, childIndexes, childValues, n->nChildren);
for (j = 0; j < d->layerSize[l+1]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l+1] + j];
ArrayInsertFloats(&n->delta, ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&n->observation, ns, childIndexes, childValues, n->nChildren);
// increase centroid dimensionality for each node from layer l+1
UpdateNodeSizes(n, n->ni + n->nChildren, n->nb, n->np, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l+1];
n->starv = d->uf_starv[l+1];
}
}
// Layer l-1
if (l > 0)
{
n = GetNodeFromDestin(d, l-1, 0, 0);
ns = n->ns;
uint pIdx = n->ni + n->nb + n->np; // index of added parent centroid
for (i = 0; i < n->nb; i++)
{
ArrayInsertFloat(&d->uf_mu[l-1][i], ns, pIdx, 0);
_normalizeMu(d->uf_mu[l-1][i], n->ni, n->nb, n->np+1);
ArrayInsertFloat(&d->uf_sigma[l-1][i], ns, pIdx, INIT_SIGMA);
ArrayInsertFloat(&d->uf_avgDelta[l-1][i], ns, pIdx, 0);
ArrayInsertFloat(&d->uf_avgSquaredDelta[l-1][i], ns, pIdx, 0);
}
ArrayInsertFloat(&d->uf_absvar[l-1], ns, pIdx, 0);
ArrayInsertFloat(&d->uf_avgAbsDelta[l-1], ns, pIdx, 0);
for (j = 0; j < d->layerSize[l-1]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l-1] + j];
ArrayInsertFloat(&n->delta, ns, pIdx, 0);
ArrayInsertFloat(&n->observation, ns, pIdx, 0);
// increase centroid dimensionality for each node from layer l-1
UpdateNodeSizes(n, n->ni, n->nb, n->np+1, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l-1];
n->starv = d->uf_starv[l-1];
}
}
}
//function for adding rescaled centroids to the existing system.
//Here each node has an extra belief dimensionality(an extra centroid)
void AddRescaledCentroids(Destin *destin,uint l,float * rcentroid,uint centid)
{
uint i,j,ns,nb,np,idx,ni;
Node * n = GetNodeFromDestin(destin, l, 0, 0);
nb = n->nb;ni = n->ni;np = n->np;ns = n->ns;
idx = ni + nb;
float * rMu, * rSigma, * rAvgDelta, * rAvgSquaredDelta;
//if rescaled function was already called before we delete duplicate centroids
if(destin->rescale==1)
{
for(i=rindex;i<destin->nb[l];i++)
{
DeleteUniformCentroid(destin,l,i);
}
}
//allocate space for uniform centroid location
destin->rindex=ni+nb;
MALLOCV(rMu, float, (ns+1));
MALLOCV(rSigma, float, (ns+1));
MALLOCV(rAvgDelta, float, (ns+1));
MALLOCV(rAvgSquaredDelta, float, (ns+1));
//initializing rMu with rescaled centroid size
for(i=0;i<ns;i++)
{
rMu[i]=rcentroid[i];
}
//initializing rsigma,rAvgDelta and rAvgSquaredDelta
for (j=0; j < ns+1; j++)
{
rSigma[j] = INIT_SIGMA;
rAvgDelta[j] = 0;
rAvgSquaredDelta[j] = 0;
}
//similiar to add uniform centroids in centroid.c
for (i = 0; i < nb; i++)
{
ArrayInsertFloat(&destin->uf_mu[l][i], ns, idx, 0);
_normalizeMu(destin->uf_mu[l][i], ni, nb+1, np);
ArrayInsertFloat(&destin->uf_sigma[l][i], ns, idx, INIT_SIGMA);
ArrayInsertFloat(&destin->uf_avgDelta[l][i], ns, idx, 0);
ArrayInsertFloat(&destin->uf_avgSquaredDelta[l][i], ns, idx, 0);
}
ArrayInsertFloat(&destin->uf_absvar[l], ns, idx, 0);
ArrayInsertFloat(&destin->uf_avgAbsDelta[l], ns, idx, 0);
ArrayAppendPtr((void *)&destin->uf_mu[l], nb, rMu);
ArrayAppendPtr((void *)&destin->uf_sigma[l], nb, rSigma);
ArrayAppendUInt(&destin->uf_winCounts[l], nb, 0);
ArrayAppendLong(&destin->uf_persistWinCounts[l], nb, 0);
ArrayAppendLong(&destin->uf_persistWinCounts_detailed[l], nb, 0);
ArrayAppendPtr((void *)&destin->uf_avgDelta[l], nb, rAvgDelta);
ArrayAppendPtr((void *)&destin->uf_avgSquaredDelta[l], nb, rAvgSquaredDelta);
ArrayAppendFloat(&destin->uf_starv[l], nb, 1);
ArrayAppendFloat(&destin->uf_winFreqs[l], nb, 1/(float) nb);
_normalizeFloatArray(destin->uf_winFreqs[l], nb+1);
destin->nb[l]++; // increase global number of centroids for layer l
for (j = 0; j < destin->layerSize[l]; j++)
{
n =& destin->nodes[destin->layerNodeOffsets[l] + j];
ArrayAppendFloat(&n->belief, nb, 0);
ArrayAppendFloat(&n->beliefEuc, nb, 0);
ArrayAppendFloat(&n->beliefMal, nb, 0);
ArrayAppendFloat(&n->outputBelief, nb, 0);
ArrayInsertFloat(&n->delta, ns, idx, 0);
ArrayInsertFloat(&n->observation, ns, idx, 0);
// increase centroid dimensionality for each node from layer l
UpdateNodeSizes(n, n->ni, nb+1, n->np, n->nc);
// pointers may change due to reallocation
n->mu = destin->uf_mu[l];
n->starv = destin->uf_starv[l];
}
if (l+1 < destin->nLayers)
{
n = GetNodeFromDestin(destin, l+1, 0, 0);
ns = n->ns;
uint childIndexes[n->nChildren]; // indexes of added centroids for all childs
float childValues[n->nChildren]; // initial values for all childs
float childSigmas[n->nChildren];
for (i = 0; i < n->nChildren; i++)
{
childIndexes[i] = (i+1)*nb;
childValues[i] = 0;
childSigmas[i] = INIT_SIGMA;
}
for (i = 0; i < n->nb; i++)
{
ArrayInsertFloats(&destin->uf_mu[l+1][i], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&destin->uf_sigma[l+1][i], ns, childIndexes, childSigmas, n->nChildren);
ArrayInsertFloats(&destin->uf_avgDelta[l+1][i], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&destin->uf_avgSquaredDelta[l+1][i], ns, childIndexes, childValues, n->nChildren);
}
ArrayInsertFloats(&destin->uf_absvar[l+1], ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&destin->uf_avgAbsDelta[l+1], ns, childIndexes, childValues, n->nChildren);
for (j = 0; j < destin->layerSize[l+1]; j++)
{
n =& destin->nodes[destin->layerNodeOffsets[l+1] + j];
ArrayInsertFloats(&n->delta, ns, childIndexes, childValues, n->nChildren);
ArrayInsertFloats(&n->observation, ns, childIndexes, childValues, n->nChildren);
// increase centroid dimensionality for each node from layer l+1
UpdateNodeSizes(n, n->ni + n->nChildren, n->nb, n->np, n->nc);
// pointers may change due to reallocation
n->mu = destin->uf_mu[l+1];
n->starv = destin->uf_starv[l+1];
}
}
// Layer l-1
if (l > 0)
{
n = GetNodeFromDestin(destin, l-1, 0, 0);
ns = n->ns;
uint pIdx = n->ni + n->nb + n->np; // index of added parent centroid
for (i = 0; i < n->nb; i++)
{
ArrayInsertFloat(&destin->uf_mu[l-1][i], ns, pIdx, 0);
_normalizeMu(destin->uf_mu[l-1][i], n->ni, n->nb, n->np+1);
ArrayInsertFloat(&destin->uf_sigma[l-1][i], ns, pIdx, INIT_SIGMA);
ArrayInsertFloat(&destin->uf_avgDelta[l-1][i], ns, pIdx, 0);
ArrayInsertFloat(&destin->uf_avgSquaredDelta[l-1][i], ns, pIdx, 0);
}
ArrayInsertFloat(&destin->uf_absvar[l-1], ns, pIdx, 0);
ArrayInsertFloat(&destin->uf_avgAbsDelta[l-1], ns, pIdx, 0);
for (j = 0; j < destin->layerSize[l-1]; j++)
{
n =& destin->nodes[destin->layerNodeOffsets[l-1] + j];
ArrayInsertFloat(&n->delta, ns, pIdx, 0);
ArrayInsertFloat(&n->observation, ns, pIdx, 0);
// increase centroid dimensionality for each node from layer l-1
UpdateNodeSizes(n, n->ni, n->nb, n->np+1, n->nc);
// pointers may change due to reallocation
n->mu = destin->uf_mu[l-1];
n->starv = destin->uf_starv[l-1];
}
}
return;
}
void DeleteUniformCentroid(Destin *d, uint l, uint idx)
{
uint i, j, ni, nb, np, ns;
Node * n = GetNodeFromDestin(d, l, 0, 0);
nb = n->nb;
ni = n->ni;
np = n->np;
ns = n->ns;
_distributeEvidenceOfDeletedCentroidToNeighbours(d, l, idx, NEAREST_OF_DELETED_CENTROID);
// Layer l
for (i = 0; i < nb; i++)
{
ArrayDeleteFloat(&d->uf_mu[l][i], ns, ni+idx);
_normalizeMu(d->uf_mu[l][i], ni, nb-1, np);
ArrayDeleteFloat(&d->uf_sigma[l][i], ns, ni+idx);
ArrayDeleteFloat(&d->uf_avgDelta[l][i], ns, ni+idx);
ArrayDeleteFloat(&d->uf_avgSquaredDelta[l][i], ns, ni+idx);
}
ArrayDeleteFloat(&d->uf_absvar[l], ns, ni+idx);
ArrayDeleteFloat(&d->uf_avgAbsDelta[l], ns, ni+idx);
ArrayDeleteArray((void *)&d->uf_mu[l], nb, idx);
ArrayDeleteArray((void *)&d->uf_sigma[l], nb, idx);
ArrayDeleteUInt(&d->uf_winCounts[l], nb, idx);
ArrayDeleteLong(&d->uf_persistWinCounts[l], nb, idx);
ArrayDeleteLong(&d->uf_persistWinCounts_detailed[l], nb, idx);
ArrayDeleteArray((void *)&d->uf_avgDelta[l], nb, idx);
ArrayDeleteArray((void *)&d->uf_avgSquaredDelta[l], nb, idx);
ArrayDeleteFloat(&d->uf_starv[l], nb, idx);
ArrayDeleteFloat(&d->uf_winFreqs[l], nb, idx);
_normalizeFloatArray(d->uf_winFreqs[l], nb - 1);
d->nb[l]--; // decrease global number of centroids for layer l
for (j = 0; j < d->layerSize[l]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l] + j];
ArrayDeleteFloat(&n->belief, nb, idx);
ArrayDeleteFloat(&n->beliefEuc, nb, idx);
ArrayDeleteFloat(&n->beliefMal, nb, idx);
ArrayDeleteFloat(&n->outputBelief, nb, idx);
ArrayDeleteFloat(&n->delta, ns, ni+idx);
ArrayDeleteFloat(&n->observation, ns, ni+idx);
// decrease centroid dimensionality for each node from layer l
UpdateNodeSizes(n, n->ni, nb-1, n->np, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l];
n->starv = d->uf_starv[l];
}
// Layer l+1
if (l+1 < d->nLayers)
{
n = GetNodeFromDestin(d, l+1, 0, 0);
ns = n->ns;
uint childIndexes[n->nChildren]; // indexes of deleted centroids for all childs
for (i = 0; i < n->nChildren; i++)
{
childIndexes[i] = i*nb + idx;
}
for (i = 0; i < n->nb; i++)
{
ArrayDeleteFloats(&d->uf_mu[l+1][i], ns, childIndexes, n->nChildren);
ArrayDeleteFloats(&d->uf_sigma[l+1][i], ns, childIndexes, n->nChildren);
ArrayDeleteFloats(&d->uf_avgDelta[l+1][i], ns, childIndexes, n->nChildren);
ArrayDeleteFloats(&d->uf_avgSquaredDelta[l+1][i], ns, childIndexes, n->nChildren);
}
ArrayDeleteFloats(&d->uf_absvar[l+1], ns, childIndexes, n->nChildren);
ArrayDeleteFloats(&d->uf_avgAbsDelta[l+1], ns, childIndexes, n->nChildren);
for (j = 0; j < d->layerSize[l+1]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l+1] + j];
ArrayDeleteFloats(&n->delta, ns, childIndexes, n->nChildren);
ArrayDeleteFloats(&n->observation, ns, childIndexes, n->nChildren);
// decrease centroid dimensionality for each node from layer l+1
UpdateNodeSizes(n, n->ni - n->nChildren, n->nb, n->np, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l+1];
n->starv = d->uf_starv[l+1];
}
}
// Layer l-1
if (l > 0)
{
n = GetNodeFromDestin(d, l-1, 0, 0);
ns = n->ns;
uint pIdx = n->ni + n->nb + idx; // index of deleted parent centroid
for (i = 0; i < n->nb; i++)
{
ArrayDeleteFloat(&d->uf_mu[l-1][i], ns, pIdx);
_normalizeMu(d->uf_mu[l-1][i], n->ni, n->nb, n->np-1);
ArrayDeleteFloat(&d->uf_sigma[l-1][i], ns, pIdx);
ArrayDeleteFloat(&d->uf_avgDelta[l-1][i], ns, pIdx);
ArrayDeleteFloat(&d->uf_avgSquaredDelta[l-1][i], ns, pIdx);
}
ArrayDeleteFloat(&d->uf_absvar[l-1], ns, pIdx);
ArrayDeleteFloat(&d->uf_avgAbsDelta[l-1], ns, pIdx);
for (j = 0; j < d->layerSize[l-1]; j++)
{
n =& d->nodes[d->layerNodeOffsets[l-1] + j];
ArrayDeleteFloat(&n->delta, ns, pIdx);
ArrayDeleteFloat(&n->observation, ns, pIdx);
// decrease centroid dimensionality for each node from layer l-1
UpdateNodeSizes(n, n->ni, n->nb, n->np-1, n->nc);
// pointers may change due to reallocation
n->mu = d->uf_mu[l-1];
n->starv = d->uf_starv[l-1];
}
}
}
