void AdaGradCpp::max(MPI_Comm comm, const double tol, const int MaxNumIterations) {
int i, IterationNum, BatchNum, BatchBegin, BatchEnd, BatchSize, GlobalBatchSize, CurrentBatchSize, WBatchBegin, WBatchEnd;
double GlobalBatchSizeDouble;
double CurrentLearningRate, LocalSumGradients;
this->SumdZdWSquared = (double*)calloc(this->lengthW, sizeof(double));
for (IterationNum = 0; IterationNum < MaxNumIterations; ++IterationNum) {//IterationNum layer
//Calculate CurrentLearningRate
CurrentLearningRate = this->LearningRate/pow((double)(IterationNum + 1), this->LearningRatePower);
//this->NumBatches is inherited from the Optimiser class
for (BatchNum = 0; BatchNum < this->NumBatches; ++BatchNum) {//BatchNum layer
//We must find out our current values for BatchBegin and BatchEnd. We do so by calling this->CalcBatchBeginEnd, which is inherited from the optimiser class.
this->CalcBatchBeginEnd(BatchBegin, BatchEnd, BatchSize, BatchNum);
//VERY IMPORTANT CONVENTION: When passing this->localdZdw to g(), always set to zero first.
for (i=0; i<this->lengthW; ++i) this->localdZdW[i] = 0.0;
//If weight is greater than wMax or smaller than wMin, then clip
//If there is no wMin or wMax, this will have no effect
wClip(this->W);
//Barrier: Wait until all processes have reached this point
MPI_Barrier(comm);
//Call g()
//Note that it is the responsibility of whoever writes the MLalgorithm to make sure that this->dZdW and this->SumdZdW are passed to ALL processes
//It is, however, your responsibility to place a barrier after that, if required
this->MLalgorithm->g(comm, this->dZdW, this->localdZdW, this->W, BatchBegin, BatchEnd, BatchSize, BatchNum, IterationNum);
//Add all BatchSize and store the result in GlobalBatchSize
MPI_Allreduce(&BatchSize, &GlobalBatchSize, 1, MPI_INT, MPI_SUM, comm);
GlobalBatchSizeDouble = (double)GlobalBatchSize;
//If weight equals wMin (wMax) and dZdW is smaller than zero (greater than zero), set dZdW to zero
//If there is no wMin or wMax, this will have no effect
dZdWClipMax();
//Barrier: Wait until all processes have reached this point
MPI_Barrier(comm);
for (i=0; i<this->lengthW; ++i) this->SumdZdW[i] += this->dZdW[i];
//Calculate "dampening parameter"
for (i=0; i<this->lengthW; ++i) this->SumdZdWSquared[i] += this->dZdW[i]*this->dZdW[i];
//Update W
for (i=0; i<this->lengthW; ++i) if (this->SumdZdWSquared[i] > 0.0) this->W[i] += (this->dZdW[i]/sqrt(this->SumdZdWSquared[i]))*CurrentLearningRate;
}//BatchNum layer
//The following lines should be left unchanged for all gradient-based-optimisers
//Calculate LocalSumGradients
this->MLalgorithm->SumGradients[IterationNum] = 0.0;
for (i=0; i<this->lengthW; ++i) this->MLalgorithm->SumGradients[IterationNum] += this->SumdZdW[i]*this->SumdZdW[i];
//Set SumdZdW to 0 for next iteration
for (i=0; i<this->lengthW; ++i) this->SumdZdW[i] = 0.0;
//Check whether convergence condition is met. If yes, break
if (this->MLalgorithm->SumGradients[IterationNum]/((double)(this->lengthW)) < tol) break;
}//IterationNum layer
//If weight is greater than wMax or smaller than wMin, then clip
//If there is no wMin or wMax, this will have no effect
wClip(this->W);
//Store number of IterationNums needed
this->MLalgorithm->IterationsNeeded = IterationNum;
free(this->SumdZdWSquared);
}
void GradientDescentCpp::min(MPI_Comm comm, const double tol, const int MaxNumIterations) {
int i, IterationNum, BatchNum, BatchBegin, BatchEnd, BatchSize, GlobalBatchSize;
double GlobalBatchSizeDouble;
double CurrentLearningRate;
for (IterationNum = 0; IterationNum < MaxNumIterations; ++IterationNum) {//IterationNum layer
//Calculate CurrentLearningRate
CurrentLearningRate = this->LearningRate/pow((double)(IterationNum + 1), this->LearningRatePower);
//this->NumBatches is inherited from the Optimiser class
for (BatchNum = 0; BatchNum < this->NumBatches; ++BatchNum) {//BatchNum layer
//We must find out our current values for BatchBegin and BatchEnd. We do so by calling this->CalcBatchBeginEnd, which is inherited from the optimiser class.
this->CalcBatchBeginEnd(BatchBegin, BatchEnd, BatchSize, BatchNum);
//VERY IMPORTANT CONVENTION: When passing this->localdZdw to g(), always set to zero first.
for (i=0; i<this->lengthW; ++i) this->localdZdW[i] = 0.0;
//If weight is greater than wMax or smaller than wMin, then clip
//If there is no wMin or wMax, this will have no effect
wClip(this->W);
//Barrier: Wait until all processes have reached this point
MPI_Barrier(comm);
//Call g()
//Note that it is the responsibility of whoever writes the MLalgorithm to make sure that this->dZdW and this->SumdZdW are passed to ALL processes
//It is, however, your responsibility to place a barrier after that, if required
this->MLalgorithm->g(comm, this->dZdW, this->localdZdW, this->W, BatchBegin, BatchEnd, BatchSize, BatchNum, IterationNum);
//Add all BatchSize and store the result in GlobalBatchSize
MPI_Allreduce(&BatchSize, &GlobalBatchSize, 1, MPI_INT, MPI_SUM, comm);
GlobalBatchSizeDouble = (double)GlobalBatchSize;
//If weight equals wMin (wMax) and dZdW is smaller than zero (greater than zero), set dZdW to zero
//If there is no wMin or wMax, this will have no effect
dZdWClipMin();
//Barrier: Wait until all processes have reached this point
MPI_Barrier(comm);
for (i=0; i<this->lengthW; ++i) {
//Record this->SumdZdW for SumGradients
this->SumdZdW[i] += this->dZdW[i];
//Record this->SumdZdW for SumSubgradients
this->MLalgorithm->SumSubgradients[this->NumBatches*IterationNum + BatchNum] += this->dZdW[i]*this->dZdW[i];
//Update W
//Learning rates are always divided by the sample size
this->W[i] -= this->dZdW[i]*CurrentLearningRate/GlobalBatchSizeDouble;
}
//Run post-update manipulation
//Some algorithms need the ability to change W once it has been updated by the optimiser, for instance to ensure that certain conditions be met.
//If there is no post-update manipulation function defined, this will have no impact whatsoever.
this->MLalgorithm->PostUpdateManipulation(comm, BatchBegin, BatchEnd, BatchSize, BatchNum, IterationNum);
}//BatchNum layer
//The following lines should be left unchanged for all gradient-based-optimisers
//Calculate SumGradients
this->MLalgorithm->SumGradients[IterationNum] = 0.0;
for (i=0; i<this->lengthW; ++i) {
this->MLalgorithm->SumGradients[IterationNum] += this->SumdZdW[i]*this->SumdZdW[i];
this->SumdZdW[i] = 0.0;//Set SumdZdW to 0 for next iteration
}
//Check whether convergence condition is met. If yes, break
if (this->MLalgorithm->SumGradients[IterationNum]/((double)(this->lengthW)) < tol) break;
}//IterationNum layer
//If weight is greater than wMax or smaller than wMin, then clip
//If there is no wMin or wMax, this will have no effect
wClip(this->W);
//Store number of IterationNums needed
this->MLalgorithm->SumGradientsLength = IterationNum;
this->MLalgorithm->SumSubgradientsLength = this->NumBatches*IterationNum;
}
