int BIDSSensorLayer::updateState(double timef, double dt){
pvdata_t * output = getCLayer()->V;
pvdata_t * input = blayer->getCLayer()->activity->data;
int index;
//Iterate through post layer
for (int i = 0; i < nx * ny; i++){
assert(nf == 1);
//Iterate through features
// std::cout << "Node (" << coords[i].xCoord << ", " << coords[i].yCoord << ")\n";
for (int k = 0; k < nf; k++){
int x = i % nx;
int y = (int) floor(i/nx);
index = kIndex(x, y, k, nx, ny, nf);
data[i][buf_index] = input[index] - (neutral_val / 256);
// std::cout << "\tBuf_index: " << buf_index << ": " << data[i][buf_index] << "\n";
//Next buf index, or reset if at end
float out = matchFilter(i, (int)(timef * dt));
output[index] = out * weight;
}
}
if(buf_index < buf_size - 1){
buf_index++;
}
else{
buf_index = 0;
}
HyPerLayer::setActivity();
return PV_SUCCESS;
}
int CIFARGTLayer::updateState(double timef, double dt) {
//getline (inputfile,inputString);
inputString = std::string(imageLayer->getFilename());
unsigned found = inputString.find_last_of("/\\");
//CIFAR is 0 indexed
char cVal = inputString.at(found-1);
iVal = cVal - '0';
pvdata_t * A = getCLayer()->activity->data;
const PVLayerLoc * loc = getLayerLoc();
//std::cout << "time: " << parent->simulationTime() << " inputString:" << inputString << " iVal:" << iVal << "\n";
assert(iVal >= 0 && iVal < 10);
//NF must be 10, one for each class
assert(loc->nf == 10);
for(int ni = 0; ni < getNumNeurons(); ni++){
int nExt = kIndexExtended(ni, 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);
if(fi == iVal){
A[nExt] = 1;
}
else{
if(negativeGt){
A[nExt] = -1;
}
else{
A[nExt] = 0;
}
}
}
return PV_SUCCESS;
}
int MLPOutputLayer::updateState(double timef, double dt) {
int status = MLPSigmoidLayer::updateState(timef, dt);
//Collect stats if needed
if(statProgressPeriod > 0){
//TODO add more if statements for different cases
if(!localTarget){
if(getLayerLoc()->nf == 1){
binaryNonlocalStats();
}
else{
multiclassNonlocalStats();
}
}
else{
if(getLayerLoc()->nf == 1){
binaryLocalStats();
}
else{
std::cout << "Not implemented\n";
exit(EXIT_FAILURE);
//TODO
//multiclassNonlocalStats();
}
}
}
//For testing purposes
pvdata_t * A = getCLayer()->activity->data;
// for(int ni = 0; ni < getNumNeurons(); ni++){
//int nExt = kIndexExtended(ni, loc->nx, loc->ny, loc->nf, loc->nb);
// std::cout << timef <<": ni: " << ni << " A: " << A[ni] << "\n";
// }
return PV_SUCCESS;
}
int MaskError::updateState(double time, double dt)
{
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int num_neurons = nx*ny*nf;
//Reset pointer of gSynHead to point to the inhib channel
pvdata_t * GSynExt = getChannel(CHANNEL_EXC);
pvdata_t * GSynInh = getChannel(CHANNEL_INH);
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * V = getV();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int ni = 0; ni < num_neurons; ni++){
int next = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//expected - actual only if expected isn't 0
if(GSynExt[ni] == 0){
A[next] = 0;
}
else{
A[next] = GSynExt[ni] - GSynInh[ni];
}
}
return PV_SUCCESS;
}
int BinaryThresh::updateState(double time, double dt)
{
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int num_neurons = nx*ny*nf;
//Reset pointer of gSynHead to point to the inhib channel
pvdata_t * GSynExt = getChannel(CHANNEL_EXC);
pvdata_t * GSynInh = getChannel(CHANNEL_INH);
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * V = getV();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int ni = 0; ni < num_neurons; ni++){
int next = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//Activity is either 0 or 1 based on if it's active
A[next] = GSynExt[ni] == 0 ? 0 : 1;
}
return PV_SUCCESS;
}
int InputLayer::updateState(double timef, double dt) {
if(!constantValue || firstRun){
char cVal = inputString.at(int(parent->simulationTime()-1)%numExamples);
iVal = cVal - '0';
}
pvdata_t * A = getCLayer()->activity->data;
const PVLayerLoc * loc = getLayerLoc();
assert(loc->nf == 2);
//Set binary values of xor values
std::cout << timef << ": input val:" << iVal << "\n";
int negVal;
negVal = -1;
for(int ni = 0; ni < getNumNeurons(); ni++){
int nExt = kIndexExtended(ni, 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.lt+loc->halo.rt, loc->ny+loc->halo.dn+loc->halo.up, loc->nf);
switch(iVal){
case 0:
if(fi == 0){
A[nExt] = negVal;
}
if(fi == 1){
A[nExt] = negVal;
}
break;
case 1:
if(fi == 0){
A[nExt] = negVal;
}
if(fi == 1){
A[nExt] = 1;
}
break;
case 2:
if(fi == 0){
A[nExt] = 1;
}
if(fi == 1){
A[nExt] = negVal;
}
break;
case 3:
if(fi == 0){
A[nExt] = 1;
}
if(fi == 1){
A[nExt] = 1;
}
break;
}
}
firstRun = false;
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 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 KmeansLayer::setActivity()
{
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
PVHalo const * halo = &loc->halo;
int num_neurons = nx*ny*nf;
int nbatch = loc->nbatch;
int status;
status = setActivity_HyPerLayer(nbatch, num_neurons, getCLayer()->activity->data, getV(), nx, ny, nf, halo->lt, halo->rt, halo->dn, halo->up);
return status;
}
int TrainingLayer::updateState(double timed, double dt) {
int status = updateState(timed, dt, getLayerLoc(), getCLayer()->activity->data, getV(), numTrainingLabels, trainingLabels, curTrainingLabelIndex, strength);
return status;
}
int SigmoidLayer::updateState(double timef, double dt) {
int status;
status = updateState(timef, dt, getLayerLoc(), getCLayer()->activity->data, getV(), 0, NULL, Vth, V0, SigmoidAlpha, SigmoidFlag, InverseFlag);
return status;
}
void MLPOutputLayer::multiclassNonlocalStats(){
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int numNeurons = getNumNeurons();
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * gtA = gtLayer->getCLayer()->activity->data;
float sumsq = 0;
//Winner take all in the output layer
int currNumRight = 0;
int currNumWrong = 0;
assert(classBuffer);
//Clear classBuffer
for(int i = 0; i < nf; i++){
classBuffer[i] = 0;
}
//Only go through restricted
//Calculate the sum squared error
for(int ni = 0; ni < numNeurons; ni++){
int nExt = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int fi = featureIndex(nExt, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
//Sum over x and y direction
classBuffer[fi] += A[nExt];
sumsq += pow(A[nExt] - gtA[nExt], 2);
}
//Normalize classBuffer to find mean
for(int i = 0; i < nf; i++){
classBuffer[i] /= nx*ny;
}
//Reduce all classBuffers through a mean
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &sumsq, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, classBuffer, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
//Normalize classBuffer across processors
for(int i = 0; i < nf; i++){
classBuffer[i] /= parent->icCommunicator()->commSize();
}
#endif // PV_USE_MPI
//Find max
float estMaxF = -1000;
int estMaxFi = -1;
float actualMaxF = -1000;
int actualMaxFi = -1;
for(int i = 0; i < nf; i++){
if(classBuffer[i] >= estMaxF){
estMaxF = classBuffer[i];
estMaxFi = i;
}
int nExt = kIndex(loc->halo.lt, loc->halo.up, i, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
if(gtA[nExt] >= actualMaxF){
actualMaxF = gtA[nExt];
actualMaxFi = i;
}
}
//Calculate stats
//Found winning feature, compare to ground truth
if(estMaxFi == actualMaxFi){
currNumRight++;
}
else{
currNumWrong++;
}
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &currNumRight, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currNumWrong, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
#endif // PV_USE_MPI
numRight += currNumRight;
numWrong += currNumWrong;
progressNumRight += currNumRight;
progressNumWrong += currNumWrong;
//Print if need
float timef = parent->simulationTime();
if(timef >= nextStatProgress){
//Update nextStatProgress
nextStatProgress += statProgressPeriod;
if (parent->columnId()==0) {
float totalScore = 100*float(numRight)/float(numRight+numWrong);
float progressScore = 100*float(progressNumRight)/float(progressNumRight+progressNumWrong);
fprintf(stdout, "time:%f layer:\"%s\" total:%f%% progressStep:%f%% energy:%f\n", timef, name, totalScore, progressScore, sumsq/2);
}
//Reset progressStats
progressNumRight = 0;
progressNumWrong = 0;
}
}
void MLPOutputLayer::binaryNonlocalStats(){
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
assert(nf == 1);
int numNeurons = getNumNeurons();
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * gtA = gtLayer->getCLayer()->activity->data;
float sumsq = 0;
float sum = 0;
float gtSum = 0;
int currNumRight = 0;
int currNumWrong = 0;
int totNum = 0;
//Only go through restricted
//Calculate the sum squared error
for(int ni = 0; ni < numNeurons; ni++){
int nExt = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int fi = featureIndex(nExt, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
//Sum over x and y direction
sumsq += pow(A[nExt] - gtA[nExt], 2);
//Sum over activity to find mean
sum += A[nExt];
gtSum += gtA[nExt];
}
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &sumsq, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &sum, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, >Sum, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
#endif // PV_USE_MPI
//Normalize sum to find mean
sum /= loc->nxGlobal * loc->nyGlobal;
gtSum /= loc->nxGlobal * loc->nyGlobal;
//gtSum should be the same as the values
assert(gtSum == gtA[0]);
//Calculate stats
if(sum < 0 && gtSum < 0){
currNumRight++;
}
else if(sum > 0 && gtSum > 0){
currNumRight++;
}
else{
currNumWrong++;
}
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &currNumRight, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currNumWrong, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
#endif // PV_USE_MPI
numRight += currNumRight;
numWrong += currNumWrong;
progressNumRight += currNumRight;
progressNumWrong += currNumWrong;
//Print if need
float timef = parent->simulationTime();
if(timef >= nextStatProgress){
//Update nextStatProgress
nextStatProgress += statProgressPeriod;
if (parent->columnId()==0) {
float totalScore = 100*float(numRight)/float(numRight+numWrong);
float progressScore = 100*float(progressNumRight)/float(progressNumRight+progressNumWrong);
fprintf(stdout, "time:%f layer:\"%s\" total:%f%% progressStep:%f%% energy:%f\n", timef, name, totalScore, progressScore, sumsq/2);
}
//Reset progressStats
progressNumRight = 0;
progressNumWrong = 0;
}
}
int MLPErrorLayer::updateState(double time, double dt)
{
//update_timer->start();
//assert(getNumChannels()>= 3);
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int num_neurons = nx*ny*nf;
//Reset pointer of gSynHead to point to the inhib channel
pvdata_t * GSynExt = getChannel(CHANNEL_EXC);
pvdata_t * GSynInh = getChannel(CHANNEL_INH);
pvdata_t Vth, sig_scale;
if(!symSigmoid){
//Calculate constants for derivitive of sigmoid layer
Vth = (VthRest+Vrest)/2.0;
sig_scale = -logf(1.0f/sigmoid_alpha - 1.0f)/(Vth - Vrest);
}
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * V = getV();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int ni = 0; ni < num_neurons; ni++){
int next = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//Update activity
//f'(V)*(error)
//error = gt - finalLayer iff error is last error
float errProp, gradient;
//exct is expected, inh is actual
if(lastError){
//0 is DCR
if(GSynExt[ni] == 0){
errProp = 0;
}
}
else{
if(strcmp(lossFunction, "squared") == 0){
//expected - actual
errProp = GSynExt[ni] - GSynInh[ni];
}
else if(strcmp(lossFunction, "entropy") == 0){
//expected/actual
errProp = GSynExt[ni]/GSynInh[ni];
}
else if(strcmp(lossFunction, "hidden") == 0){
errProp = GSynExt[ni];
}
if(symSigmoid){
gradient = 1.14393 * (1/(pow(cosh(((float)2/3) * V[ni]), 2))) + linAlpha;
}
else{
gradient = -.5 * sig_scale * (1/(pow(cosh(sig_scale*(Vth - V[ni])), 2)));
}
}
A[next] = dropout[ni] ? 0 : errProp * gradient;
}
//update_timer->stop();
return PV_SUCCESS;
}
int IncrementLayer::updateState(double timef, double dt) {
int status;
assert(numChannels>=2); // updateState uses gSynExc and gSynInh
status = doUpdateState(timef, dt, &VInited, &nextUpdateTime, firstUpdateTime, displayPeriod, getLayerLoc(), getCLayer()->activity->data, getV(), getVprev(), getNumChannels(), GSyn[0]);
return status;
}
int PtwiseProductLayer::updateState(double timef, double dt) {
int status;
status = doUpdateState(timef, dt, getLayerLoc(), getCLayer()->activity->data, getV(), getNumChannels(), GSyn[0]);
return status;
}
void MLPOutputLayer::binaryLocalStats(){
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int numNeurons = getNumNeurons();
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * gtA = gtLayer->getCLayer()->activity->data;
float sumsq = 0;
assert(nf == 1);
int currNumTotPos = 0;
int currNumTotNeg = 0;
int currTruePos = 0;
int currTrueNeg = 0;
for(int ni = 0; ni < numNeurons; ni++){
int nExt = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//DCR
if(gtA[nExt] == 0){
continue;
//Note that sumsq doesn't get updated in this case, so a dcr doesn't contribute to the score at all
}
//Negative
else if(gtA[nExt] == -1){
currNumTotNeg++;
if(A[nExt] < 0){
currTrueNeg++;
}
}
//Positive
else if(gtA[nExt] == 1){
currNumTotPos++;
if(A[nExt] > 0){
currTruePos++;
}
}
sumsq += pow(A[nExt] - gtA[nExt], 2);
}
//Do MPI
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &currNumTotPos, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currNumTotNeg, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currTruePos, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currTrueNeg, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &sumsq, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
#endif
numTotPos += currNumTotPos;
numTotNeg += currNumTotNeg;
truePos += currTruePos;
trueNeg += currTrueNeg;
progressNumTotPos += currNumTotPos;
progressNumTotNeg += currNumTotNeg;
progressTruePos += currTruePos;
progressTrueNeg += currTrueNeg;
//Print if need
float timef = parent->simulationTime();
if(timef >= nextStatProgress){
//Update nextStatProgress
nextStatProgress += statProgressPeriod;
if (parent->columnId()==0) {
float totalScore = 50*(float(truePos)/float(numTotPos) + float(trueNeg)/float(numTotNeg));
float progressScore = 50*(float(progressTruePos)/float(progressNumTotPos) + float(progressTrueNeg)/float(progressNumTotNeg));
fprintf(stdout, "time:%f layer:\"%s\" total:%f%% progressStep:%f%% energy:%f\n", timef, name, totalScore, progressScore, sumsq/2);
}
//Reset progressStats
progressNumTotPos = 0;
progressNumTotNeg = 0;
progressTruePos = 0;
progressTrueNeg = 0;
}
}
