int MaskFromMemoryBuffer::updateState(double time, double dt)
{
if (imageLayer->getDataLeft() == dataLeft &&
imageLayer->getDataTop() == dataTop &&
imageLayer->getDataWidth() == dataRight-dataLeft &&
imageLayer->getDataHeight() && dataBottom-dataTop) {
return PV_SUCCESS; // mask only needs to change if the imageLayer changes its active region
}
dataLeft = imageLayer->getDataLeft();
dataRight = dataLeft+imageLayer->getDataWidth();
dataTop = imageLayer->getDataTop();
dataBottom = dataTop + imageLayer->getDataHeight();
PVLayerLoc const * loc = getLayerLoc();
for(int b = 0; b < loc->nbatch; b++) {
pvdata_t * ABatch = getActivity() + b * getNumExtended();
int const num_neurons = getNumNeurons();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int ni = 0; ni < num_neurons; ni++) {
PVHalo const * halo = &loc->halo;
int const nx = loc->nx;
int const ny = loc->ny;
int const nf = loc->nf;
int x = kxPos(ni, nx, ny, nf);
int y = kyPos(ni, nx, ny, nf);
pvadata_t a = (pvadata_t) (x>=dataLeft && x < dataRight && y >= dataTop && y < dataBottom);
int nExt = kIndexExtended(ni, nx, ny, nf, halo->lt, halo->rt, halo->dn, halo->up);
ABatch[nExt] = a;
}
}
return PV_SUCCESS;
}
int ImageTestLayer::updateStateWrapper(double time, double dt)
{
Image::updateStateWrapper(time, dt);
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nbatch = loc->nbatch;
for(int b = 0; b < nbatch; b++){
pvdata_t * dataBatch = data + b * getNumExtended();
for(int nkRes = 0; nkRes < getNumNeurons(); nkRes++){
//Calculate extended index
int nkExt = kIndexExtended(nkRes, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//checkVal is the value from batch index 0
pvdata_t checkVal = dataBatch[nkExt] * 255;
int kxGlobal = kxPos(nkRes, nx, ny, nf) + loc->kx0;
int kyGlobal = kyPos(nkRes, nx, ny, nf) + loc->ky0;
int kf = featureIndex(nkRes, nx, ny, nf);
pvdata_t expectedVal = kIndex(kxGlobal, kyGlobal, kf, loc->nxGlobal, loc->nyGlobal, nf);
if(fabs(checkVal - expectedVal) >= 1e-5){
std::cout << "ImageFileIO test Expected: " << expectedVal << " Actual: " << checkVal << "\n";
exit(-1);
}
}
}
return PV_SUCCESS;
}
int SegmentifyTest::updateState(double timef, double dt){
//Do update state first
Segmentify::updateState(timef, dt);
const PVLayerLoc * loc = getLayerLoc();
pvdata_t * A = getActivity();
assert(A);
for(int bi = 0; bi < loc->nbatch; bi++){
pvdata_t * batchA = A + bi * getNumExtended();
for(int yi = 0; yi < loc->ny; yi++){
for(int xi = 0; xi < loc->nx; xi++){
for(int fi = 0; fi < loc->nf; fi++){
int extIdx = (yi + loc->halo.up) * (loc->nx + loc->halo.lt + loc->halo.rt) * loc->nf + (xi + loc->halo.lt) * loc->nf + fi;
float actualVal = batchA[extIdx];
float targetVal = getTargetVal(yi+loc->ky0, xi+loc->kx0, fi);
checkOutputVals(yi+loc->ky0, xi+loc->kx0, fi, targetVal, actualVal);
//std::cout << "Idx: (" << bi << "," << yi << "," << xi << "," << fi << ") Val: " << actualVal << " Target: " << targetVal << "\n";
}
}
}
}
return PV_SUCCESS;
}
// set activity to global x/y/f position, using position in border/margin as required
int PlasticConnTestLayer::setActivitytoGlobalPos(){
for (int kLocalExt = 0; kLocalExt < getNumExtended(); kLocalExt++){
int kxLocalExt = kxPos(kLocalExt, clayer->loc.nx + clayer->loc.halo.lt + clayer->loc.halo.rt, clayer->loc.ny + clayer->loc.halo.dn + clayer->loc.halo.up, clayer->loc.nf) - clayer->loc.halo.lt;
int kxGlobalExt = kxLocalExt + clayer->loc.kx0;
float xScaleLog2 = clayer->xScale;
float x0 = xOriginGlobal(xScaleLog2);
float dx = deltaX(xScaleLog2);
float x_global_pos = (x0 + dx * kxGlobalExt);
clayer->activity->data[kLocalExt] = x_global_pos;
}
return PV_SUCCESS;
}
int AccumulateLayer::setActivity() {
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int num_neurons = nx*ny*nf;
int status = PV_SUCCESS;
memset(clayer->activity->data, 0, sizeof(pvdata_t)*getNumExtended());
if( status == PV_SUCCESS ) status = applyVThresh_ANNLayer(num_neurons, getV(), AMin, VThresh, AShift, VWidth, getCLayer()->activity->data, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
if( status == PV_SUCCESS ) status = applyVMax_ANNLayer(num_neurons, getV(), AMax, getCLayer()->activity->data, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
return status;
}
int RunningAverageLayer::updateState(double timef, double dt) {
numUpdateTimes++;
int status = PV_SUCCESS;
double deltaT = parent->getDeltaTime();
//Check if an update is needed
//Done in cloneVLayer
//if(checkIfUpdateNeeded()){
int numNeurons = originalLayer->getNumNeurons();
pvdata_t * A = clayer->activity->data;
const pvdata_t * originalA = originalLayer->getCLayer()->activity->data;
const PVLayerLoc * loc = getLayerLoc();
const PVLayerLoc * locOriginal = originalLayer->getLayerLoc();
int nbatch = loc->nbatch;
//Make sure all sizes match
//assert(locOriginal->nb == loc->nb);
assert(locOriginal->nx == loc->nx);
assert(locOriginal->ny == loc->ny);
assert(locOriginal->nf == loc->nf);
for(int b = 0; b < nbatch; b++){
const pvdata_t * originalABatch = originalA + b * originalLayer->getNumExtended();
pvdata_t * ABatch = A + b * getNumExtended();
if (numUpdateTimes < numImagesToAverage*deltaT){
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif // PV_USE_OPENMP_THREADS
for(int k=0; k<numNeurons; k++) {
int kExt = kIndexExtended(k, loc->nx, loc->ny, loc->nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int kExtOriginal = kIndexExtended(k, locOriginal->nx, locOriginal->ny, locOriginal->nf,
locOriginal->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
ABatch[kExt] = ((numUpdateTimes/deltaT-1) * ABatch[kExt] + originalABatch[kExtOriginal]) * deltaT / numUpdateTimes;
}
}
else{
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif // PV_USE_OPENMP_THREADS
for(int k=0; k<numNeurons; k++) {
int kExt = kIndexExtended(k, loc->nx, loc->ny, loc->nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int kExtOriginal = kIndexExtended(k, locOriginal->nx, locOriginal->ny, locOriginal->nf,
locOriginal->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
ABatch[kExt] = ((numImagesToAverage-1) * ABatch[kExt] + originalABatch[kExtOriginal]) / numImagesToAverage;
}
}
}
//Update lastUpdateTime
lastUpdateTime = parent->simulationTime();
//}
return status;
}
//Makes a layer such that the restricted space is the index, but with spinning order be [x, y, f] as opposed to [f, x, y]
int InputLayer::updateState(double timef, double dt){
//Grab layer size
const PVLayerLoc* loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nxGlobal = loc->nxGlobal;
int nyGlobal = loc->nyGlobal;
int kx0 = loc->kx0;
int ky0 = loc->ky0;
for(int b = 0; b < parent->getNBatch(); b++){
pvdata_t * A = getActivity() + b * getNumExtended();
//looping over ext
for(int iY = 0; iY < ny+loc->halo.up+loc->halo.dn; iY++){
for(int iX = 0; iX < nx+loc->halo.lt+loc->halo.rt; iX++){
//Calculate x and y global extended
int xGlobalExt = iX + loc->kx0;
int yGlobalExt = iY + loc->ky0;
//Calculate x and y in restricted space
int xGlobalRes = xGlobalExt - loc->halo.lt;
int yGlobalRes = yGlobalExt - loc->halo.up;
//Calculate base value
//xGlobal and yGlobalRes can be negative
int baseActivityVal = yGlobalRes * nxGlobal + xGlobalRes;
for(int iFeature = 0; iFeature < nf; iFeature++){
int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
//Feature gives an offset, since it spins slowest
int activityVal = baseActivityVal + iFeature * nxGlobal * nyGlobal;
A[ext_idx] = activityVal;
}
}
}
}
////Printing for double checking
//printf("\nOutMat\n");
////looping over ext
//for(int iFeature = 0; iFeature < nf; iFeature++){
// for(int iY = 0; iY < ny+loc->halo.up+loc->halo.dn; iY++){
// for(int iX = 0; iX < nx+loc->halo.lt+loc->halo.rt; iX++){
// int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
// printf("%03d ", (int)A[ext_idx]);
// }
// printf("\n");
// }
// printf("\n\n");
//}
return PV_SUCCESS;
}
int FilenameParsingGroundTruthLayer::updateState(double time, double dt)
{
update_timer->start();
pvdata_t * A = getCLayer()->activity->data;
const PVLayerLoc * loc = getLayerLoc();
int num_neurons = getNumNeurons();
if (num_neurons != numClasses)
{
pvError() << "The number of neurons in " << getName() << " is not equal to the number of classes specified in " << parent->getOutputPath() << "/classes.txt\n";
}
for(int b = 0; b < loc->nbatch; b++){
char * currentFilename = NULL;
int filenameLen = 0;
//TODO depending on speed of this layer, more efficient way would be to preallocate currentFilename buffer
if(parent->icCommunicator()->commRank()==0){
currentFilename = strdup(movieLayer->getFilename(b));
//Get length of currentFilename and broadcast
int filenameLen = (int) strlen(currentFilename) + 1; //+1 for the null terminator
//Using local communicator, as each batch MPI will handle it's own run
MPI_Bcast(&filenameLen, 1, MPI_INT, 0, parent->icCommunicator()->communicator());
//Braodcast filename to all other local processes
MPI_Bcast(currentFilename, filenameLen, MPI_CHAR, 0, parent->icCommunicator()->communicator());
}
else{
//Receive broadcast about length of filename
MPI_Bcast(&filenameLen, 1, MPI_INT, 0, parent->icCommunicator()->communicator());
currentFilename = (char*)calloc(sizeof(char), filenameLen);
//Receive filename
MPI_Bcast(currentFilename, filenameLen, MPI_CHAR, 0, parent->icCommunicator()->communicator());
}
std::string fil = currentFilename;
pvdata_t * ABatch = A + b * getNumExtended();
for(int i = 0; i < num_neurons; i++){
int nExt = kIndexExtended(i, loc->nx, loc->ny, loc->nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int fi = featureIndex(nExt, loc->nx+loc->halo.rt+loc->halo.lt, loc->ny+loc->halo.dn+loc->halo.up, loc->nf);
int match = fil.find(classes[i]);
if(0 <= match){
ABatch[nExt] = gtClassTrueValue;
}
else{
ABatch[nExt] = gtClassFalseValue;
}
}
//Free buffer, TODO, preallocate buffer to avoid this
free(currentFilename);
}
update_timer->stop();
return PV_SUCCESS;
}
int GatePoolTestLayer::updateState(double timef, double dt) {
//Do update state of ANN Layer first
ANNLayer::updateState(timef, dt);
//Grab layer size
const PVLayerLoc* loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nxGlobal = loc->nxGlobal;
int nyGlobal = loc->nyGlobal;
int nf = loc->nf;
int kx0 = loc->kx0;
int ky0 = loc->ky0;
bool isCorrect = true;
//Grab the activity layer of current layer
for(int b = 0; b < loc->nbatch; b++) {
const pvdata_t * A = getActivity() + b * getNumExtended();
//We only care about restricted space, but iY and iX are extended
for(int iY = loc->halo.up; iY < ny + loc->halo.up; iY++) {
for(int iX = loc->halo.lt; iX < nx + loc->halo.lt; iX++) {
for(int iFeature = 0; iFeature < nf; iFeature++) {
int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
float actualvalue = A[ext_idx];
int xval = (iX + kx0 - loc->halo.lt)/2;
int yval = (iY + ky0 - loc->halo.up)/2;
assert(xval >= 0 && xval < loc->nxGlobal);
assert(yval >= 0 && yval < loc->nxGlobal);
float expectedvalue;
expectedvalue = iFeature * 64 + yval * 16 + xval * 2 + 4.5;
expectedvalue*=4;
if(fabs(actualvalue - expectedvalue) >= 1e-4) {
pvErrorNoExit() << "Connection " << name << " Mismatch at (" << iX << "," << iY << ") : actual value: " << actualvalue << " Expected value: " << expectedvalue << ". Discrepancy is a whopping " << actualvalue - expectedvalue << "! Horrors!" << "\n";
isCorrect = false;
}
}
}
}
}
if(!isCorrect) {
InterColComm * icComm = parent->icCommunicator();
MPI_Barrier(icComm->communicator()); // If there is an error, make sure that MPI doesn't kill the run before process 0 reports the error.
exit(-1);
}
return PV_SUCCESS;
}
int BIDSCloneLayer::allocateDataStructures() {
int status = CloneVLayer::allocateDataStructures();
assert(GSyn==NULL);
BIDSMovieCloneMap *blayer = dynamic_cast<BIDSMovieCloneMap*> (originalLayer->getParent()->getLayerFromName(jitterSourceName));
if (blayer==NULL) {
fprintf(stderr, "BIDSCloneLayer \"%s\": jitterSource \"%s\" must be a BIDSMovieCloneMap.\n", name, jitterSourceName);
abort();
}
coords = blayer->getCoords();
numNodes = blayer->getNumNodes();
for(int i = 0; i < getNumExtended(); i++){
this->clayer->activity->data[i] = 0;
}
return status;
}
int AccumulateLayer::doUpdateState(double time, double dt, const PVLayerLoc * loc, pvdata_t * A,
pvdata_t * V, int num_channels, pvdata_t * gSynHead)
{
bool needsUpdate = false;
if (syncedInputLayer != NULL) {
if (getPhase() > syncedInputLayer->getPhase()) {
needsUpdate = syncedInputLayer->getLastUpdateTime() >= lastUpdateTime;
}
else {
needsUpdate = syncedInputLayer->getLastUpdateTime() > lastUpdateTime;
}
}
if (needsUpdate) {
memset(clayer->activity->data, 0, sizeof(pvdata_t)*getNumExtended());
}
//update_timer->start();
//#ifdef PV_USE_OPENCL
// if(gpuAccelerateFlag) {
// updateStateOpenCL(time, dt);
// //HyPerLayer::updateState(time, dt);
// }
// else {
//#endif
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int num_neurons = nx*ny*nf;
updateV_AccumulateLayer(num_neurons, V, num_channels, gSynHead, A,
AMax, AMin, VThresh, AShift, VWidth, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//Moved to publish
//if (this->writeSparseActivity){
// updateActiveIndices(); // added by GTK to allow for sparse output, can this be made an inline function???
//}
//#ifdef PV_USE_OPENCL
// }
//#endif
//update_timer->stop();
return PV_SUCCESS;
}
int PursuitLayer::updateState(double time, double dt) {
if (!updateReady) return PV_SUCCESS;
int nx = getLayerLoc()->nx;
int ny = getLayerLoc()->ny;
int nf = getLayerLoc()->nf;
PVHalo const * halo = &getLayerLoc()->halo;
pvdata_t * activity = getActivity();
memset(activity, 0, getNumExtended()*sizeof(*activity));
int nxy = nx*ny;
for (int kxy=0; kxy<nxy; kxy++) {
int kf = foundFeatures[kxy];
if (kf>=0) {
int kx = kxPos(kxy,nx,ny,1);
int ky = kyPos(kxy,nx,ny,1);
int kex = kIndex(kx+halo->lt, ky+halo->up, kf, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, nf); /* Is this correct? Before splitting x- and y- margin widths, the ny argument was ny*nb, which seems weird. */
activity[kex] = gSynSparse[kxy];
}
}
//resetGSynBuffers_HyPerLayer(getNumNeurons(), getNumChannels(), GSyn[0]);
updateReady = false;
return PV_SUCCESS;
}
int MoviePvpTestLayer::updateStateWrapper(double time, double dt)
{
MoviePvp::updateStateWrapper(time, dt);
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nbatch = loc->nbatch;
for(int b = 0; b < nbatch; b++){
pvdata_t * dataBatch = data + b * getNumExtended();
int frameIdx;
if(strcmp(getBatchMethod(), "byImage") == 0){
frameIdx = (time-1) * nbatch + b;
}
else if(strcmp(getBatchMethod(), "byMovie") == 0){
frameIdx = b * 2 + (time-1);
}
for(int nkRes = 0; nkRes < getNumNeurons(); nkRes++){
//Calculate extended index
int nkExt = kIndexExtended(nkRes, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//checkVal is the value from batch index 0
pvdata_t checkVal = dataBatch[nkExt];
int kxGlobal = kxPos(nkRes, nx, ny, nf) + loc->kx0;
int kyGlobal = kyPos(nkRes, nx, ny, nf) + loc->ky0;
int kf = featureIndex(nkRes, nx, ny, nf);
pvdata_t expectedVal = kIndex(kxGlobal, kyGlobal, kf, loc->nxGlobal, loc->nyGlobal, nf) + frameIdx*192;
if(fabs(checkVal - expectedVal) >= 1e-5){
std::cout << "ImageFileIO " << name << " test Expected: " << expectedVal << " Actual: " << checkVal << "\n";
//exit(-1);
}
}
}
return PV_SUCCESS;
}
int SegmentLayer::updateState(double timef, double dt) {
pvdata_t* srcA = originalLayer->getActivity();
pvdata_t* thisA = getActivity();
assert(srcA);
assert(thisA);
const PVLayerLoc* loc = getLayerLoc();
//Segment input layer based on segmentMethod
if(strcmp(segmentMethod, "none") == 0){
int numBatchExtended = getNumExtendedAllBatches();
//Copy activity over
//Since both buffers should be identical size, we can do a memcpy here
memcpy(thisA, srcA, numBatchExtended * sizeof(pvdata_t));
}
else{
//This case should never happen
assert(0);
}
assert(loc->nf == 1);
//Clear centerIdxs
for(int bi = 0; bi < loc->nbatch; bi++){
centerIdx[bi].clear();
}
for(int bi = 0; bi < loc->nbatch; bi++){
pvdata_t* batchA = thisA + bi * getNumExtended();
//Reset max/min buffers
maxX.clear();
maxY.clear();
minX.clear();
minY.clear();
//Loop through this buffer to fill labelVec and idxVec
//Looping through restricted, but indices are extended
for(int yi = loc->halo.up; yi < loc->ny+loc->halo.up; yi++){
for(int xi = loc->halo.lt; xi < loc->nx+loc->halo.lt; xi++){
//Convert to local extended linear index
int niLocalExt = yi * (loc->nx+loc->halo.lt+loc->halo.rt) + xi;
//Convert yi and xi to global res index
int globalResYi = yi - loc->halo.up + loc->ky0;
int globalResXi = xi - loc->halo.lt + loc->kx0;
//Get label value
//Note that we're assuming that the activity here are integers,
//even though the buffer is floats
int labelVal = round(batchA[niLocalExt]);
//Calculate max/min x and y for a single batch
//If labelVal exists in map
if(maxX.count(labelVal)){
//Here, we're assuming the 4 maps are in sync, so we use the
//.at method, as it will throw an exception as opposed to the
//[] operator, which will simply add the key into the map
if(globalResXi > maxX.at(labelVal)){
maxX[labelVal] = globalResXi;
}
if(globalResXi < minX.at(labelVal)){
minX[labelVal] = globalResXi;
}
if(globalResYi > maxY.at(labelVal)){
maxY[labelVal] = globalResYi;
}
if(globalResYi < minY.at(labelVal)){
minY[labelVal] = globalResYi;
}
}
//If doesn't exist, add into map with current vals
else{
maxX[labelVal] = globalResXi;
minX[labelVal] = globalResXi;
maxY[labelVal] = globalResYi;
minY[labelVal] = globalResYi;
}
}
}
//We need to mpi across processors in case a segment crosses an mpi boundary
InterColComm * icComm = parent->icCommunicator();
int numMpi = icComm->commSize();
int rank = icComm->commRank();
//Local comm rank
//Non root processes simply send buffer size and then buffers
int numLabels = maxX.size();
if(rank != 0){
//Load buffers
loadLabelBuf();
//Send number of labels first
MPI_Send(&numLabels, 1, MPI_INT, 0, rank, icComm->communicator());
//Send labels, then max/min buffers
MPI_Send(labelBuf, numLabels, MPI_INT, 0, rank, icComm->communicator());
MPI_Send(maxXBuf, numLabels, MPI_INT, 0, rank, icComm->communicator());
MPI_Send(maxYBuf, numLabels, MPI_INT, 0, rank, icComm->communicator());
MPI_Send(minXBuf, numLabels, MPI_INT, 0, rank, icComm->communicator());
MPI_Send(minYBuf, numLabels, MPI_INT, 0, rank, icComm->communicator());
//Receive the full centerIdxBuf from root process
int numCenterIdx = 0;
MPI_Bcast(&numCenterIdx, 1, MPI_INT, 0, icComm->communicator());
checkIdxBufSize(numCenterIdx);
MPI_Bcast(allLabelsBuf, numCenterIdx, MPI_INT, 0, icComm->communicator());
MPI_Bcast(centerIdxBuf, numCenterIdx, MPI_INT, 0, icComm->communicator());
//Load buffer into centerIdx map
loadCenterIdxMap(bi, numCenterIdx);
}
//Root process stores everything
else{
//One recv per buffer
for(int recvRank = 1; recvRank < numMpi; recvRank++){
int numRecvLabels = 0;
MPI_Recv(&numRecvLabels, 1, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
checkLabelBufSize(numRecvLabels);
MPI_Recv(labelBuf, numRecvLabels, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
MPI_Recv(maxXBuf, numRecvLabels, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
MPI_Recv(maxYBuf, numRecvLabels, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
MPI_Recv(minXBuf, numRecvLabels, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
MPI_Recv(minYBuf, numRecvLabels, MPI_INT, recvRank, recvRank, icComm->communicator(), NULL);
for(int i = 0; i < numRecvLabels; i++){
int label = labelBuf[i];
//Add on to maps
//If the label already exists, fill with proper max/min
if(maxX.count(label)){
if(maxXBuf[i] > maxX.at(label)){
maxX[label] = maxXBuf[i];
}
if(maxYBuf[i] > maxY.at(label)){
maxY[label] = maxYBuf[i];
}
if(minXBuf[i] < minX.at(label)){
minX[label] = minXBuf[i];
}
if(minYBuf[i] < minY.at(label)){
minY[label] = minYBuf[i];
}
}
else{
maxX[label] = maxXBuf[i];
maxY[label] = maxYBuf[i];
minX[label] = minXBuf[i];
minY[label] = minYBuf[i];
}
}
}
//Maps are now filled with all segments from the image
//Fill centerIdx based on max/min
for(std::map<int, int>::iterator it = maxX.begin();
it != maxX.end(); ++it){
int label = it->first;
int centerX = minX.at(label) + (maxX.at(label) - minX.at(label))/2;
int centerY = minY.at(label) + (maxY.at(label) - minY.at(label))/2;
//Convert centerpoints (in global res idx) to linear idx (in global res space)
int centerIdxVal = centerY * (loc->nxGlobal) + centerX;
//Add to centerIdxMap
centerIdx[bi][label] = centerIdxVal;
}
//Fill centerpoint buffer
int numCenterIdx = centerIdx[bi].size();
checkIdxBufSize(numCenterIdx);
int idx = 0;
for(std::map<int, int>::iterator it = centerIdx[bi].begin();
it != centerIdx[bi].end(); ++it){
allLabelsBuf[idx] = it->first;
centerIdxBuf[idx] = it->second;
idx++;
}
//Broadcast buffers
MPI_Bcast(&numCenterIdx, 1, MPI_INT, 0, icComm->communicator());
MPI_Bcast(allLabelsBuf, numCenterIdx, MPI_INT, 0, icComm->communicator());
MPI_Bcast(centerIdxBuf, numCenterIdx, MPI_INT, 0, icComm->communicator());
}
} //End batch loop
//centerIdx now stores each center coordinate of each segment
return PV_SUCCESS;
}
int BackwardsBatchNorm::updateState(double timef, double dt) {
int status = PV_SUCCESS;
//We are filling this activity buffer
pvdata_t * thisA = clayer->activity->data;
assert(thisA);
//We need the normalized input vals, orig input vals, and the input gradients
const pvdata_t * inputGradA = originalLayer->getCLayer()->activity->data;
const pvdata_t * forwardA = forwardLayer->getCLayer()->activity->data;
const pvdata_t * origInputA = forwardLayer->getOriginalLayer()->getCLayer()->activity->data;
assert(inputGradA && forwardA && origInputA);
//Get locs for all buffers
const PVLayerLoc * thisLoc = getLayerLoc();
const PVLayerLoc * inputGradLoc = originalLayer->getLayerLoc();
const PVLayerLoc * forwardLoc = forwardLayer->getLayerLoc();
const PVLayerLoc * origInputLoc = forwardLayer->getOriginalLayer()->getLayerLoc();
int nbatch = thisLoc->nbatch;
//All nx, ny, and nf should be the same
int nx = thisLoc->nx;
int ny = thisLoc->ny;
int nf = thisLoc->nf;
//Get buffer margins here
int xThisMargin = thisLoc->halo.lt + thisLoc->halo.rt;
int yThisMargin = thisLoc->halo.up + thisLoc->halo.dn;
int xInputGradMargin = inputGradLoc->halo.lt + inputGradLoc->halo.rt;
int yInputGradMargin = inputGradLoc->halo.up + inputGradLoc->halo.dn;
int xForwardMargin = forwardLoc->halo.lt + forwardLoc->halo.rt;
int yForwardMargin = forwardLoc->halo.up + forwardLoc->halo.dn;
int xOrigInputMargin = origInputLoc->halo.lt + origInputLoc->halo.rt;
int yOrigInputMargin = origInputLoc->halo.up + origInputLoc->halo.dn;
//We also need various mean and var buffers from the forward layer
const float* batchMean = forwardLayer->getBatchMean();
const float* batchVar = forwardLayer->getBatchVar();
float* batchMeanShift = forwardLayer->getBatchMeanShift();
float* batchVarShift = forwardLayer->getBatchVarShift();
float epsilon = forwardLayer->getEpsilon();
//Total number of neurons to divide by for each feature
float normVal = parent->getNBatchGlobal() * thisLoc->nyGlobal * thisLoc->nxGlobal;
//We're accumulating into delta buffers, so clear
clearDelta();
//Ioffe et. al. Batch Normalization
//Calculate deltaVar
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float secondTerm = -.5*(powf(batchVar[iF] + epsilon, -1.5));
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
deltaVar[iF] += deltaNorm * (batchOrigInputA[kExtOrigInput] - batchMean[iF]);
}
}
}
//Multiply deltaVar by secondTerm
deltaVar[iF] = deltaVar[iF] * secondTerm;
}
//Reduce deltaVar
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaVar, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//Calculate deltaMean
//Calculate first term first
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float multiplier = -1.0/(sqrtf(batchVar[iF]+epsilon));
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
deltaMean[iF] += deltaNorm * multiplier;
}
}
}
}
//Reduce deltaMean across mpi
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaMean, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//Calculate second term
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float tmpMean = 0;
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
tmpMean += -2 * (batchOrigInputA[kExtOrigInput] - batchMean[iF]);
}
}
}
//Reduce tmpMean
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &tmpMean, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
tmpMean = tmpMean / normVal;
//Add second term to first term
deltaMean[iF] += deltaVar[iF] * tmpMean;
}
//No more sums, go with efficient loop
//TODO Is the efficient loop better for optimization or do we put
//features on the outer most loop for precalculation of constants over features?
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
pvdata_t* batchThisA = thisA + b * getNumExtended();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for collapse(3)
#endif
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
for(int iF = 0; iF < nf; iF++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
int kExtThis = kIndex(iX, iY, iF, nx+xThisMargin, ny+yThisMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
float firstTerm = deltaNorm/sqrtf(batchVar[iF] + epsilon);
float secondTerm = deltaVar[iF] * (2*(batchOrigInputA[kExtOrigInput] - batchMean[iF])/normVal);
float thirdTerm = deltaMean[iF]/normVal;
batchThisA[kExtThis] = firstTerm + secondTerm + thirdTerm;
}
}
}
}
//We calculate delta varShift and deltaMeanShift here
//TODO parallize over threads
//Since we're summing into delta*shift buffers, we have to sequentialize over features
for(int iF = 0; iF < nf; iF++) {
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchForwardA = forwardA + b * forwardLayer->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
int kExtForwardA = kIndex(iX, iY, iF, nx+xForwardMargin, ny + yForwardMargin, nf);
deltaVarShift[iF] += batchInputGradA[kExtInputGrad] * batchForwardA[kExtForwardA];
deltaMeanShift[iF] += batchInputGradA[kExtInputGrad];
}
}
}
}
//Reduce delta*Shift across all mpi
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaVarShift, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
MPI_Allreduce(MPI_IN_PLACE, deltaMeanShift, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//TODO implement learning rule for meanShift and varShift
return status;
}