bool Softmax::trainSoftmaxModel(UINT classLabel,SoftmaxModel &model,ClassificationData &data){
Float error = 0;
Float errorSum = 0;
Float lastErrorSum = 0;
Float delta = 0;
const UINT N = data.getNumDimensions();
const UINT M = data.getNumSamples();
UINT iter = 0;
bool keepTraining = true;
Random random;
VectorFloat y(M);
VectorFloat batchMean(N);
Vector< UINT > randomTrainingOrder(M);
Vector< VectorFloat > batchData(batchSize,VectorFloat(N));
//Init the model
model.init( classLabel, N );
//Setup the target vector, the input data is relabelled as positive samples (with label 1.0) and negative samples (with label 0.0)
for(UINT i=0; i<M; i++){
y[i] = data[i].getClassLabel()==classLabel ? 1.0 : 0;
}
//In most cases, the training data is grouped into classes (100 samples for class 1, followed by 100 samples for class 2, etc.)
//This can cause a problem for stochastic gradient descent algorithm. To avoid this issue, we randomly shuffle the order of the
//training samples. This random order is then used at each epoch.
for(UINT i=0; i<M; i++){
randomTrainingOrder[i] = i;
}
std::random_shuffle(randomTrainingOrder.begin(), randomTrainingOrder.end());
//Clear any previous training results
trainingResults.clear();
trainingResults.reserve( maxNumEpochs );
TrainingResult epochResult;
//Run the main stochastic gradient descent training algorithm
while( keepTraining ){
//Run one epoch of training using stochastic gradient descent
errorSum = 0;
UINT m=0;
while( m < M ){
//Get the batch data for this update
UINT roundSize = m+batchSize < M ? batchSize : M-m;
batchMean.fill(0.0);
for(UINT i=0; i<roundSize; i++){
for(UINT j=0; j<N; j++){
batchData[i][j] = data[ randomTrainingOrder[m+i] ][j];
batchMean[j] += batchData[i][j];
}
}
for(UINT j=0; j<N; j++) batchMean[j] /= roundSize;
//Compute the error on this batch, given the current weights
error = 0.0;
for(UINT i=0; i<roundSize; i++){
error += y[ randomTrainingOrder[m+i] ] - model.compute( batchData[i] );
}
error /= roundSize;
errorSum += error;
//Update the weights
for(UINT j=0; j<N; j++){
model.w[j] += learningRate * error * batchMean[j];
}
model.w0 += learningRate * error;
m += roundSize;
}
//Compute the error
delta = fabs( errorSum-lastErrorSum );
lastErrorSum = errorSum;
//Check to see if we should stop
if( delta <= minChange ){
keepTraining = false;
}
if( ++iter >= maxNumEpochs ){
keepTraining = false;
}
trainingLog << "Class: " << classLabel << " Epoch: " << iter << " TotalError: " << errorSum << " Delta: " << delta << std::endl;
epochResult.setClassificationResult( iter, errorSum, this );
trainingResults.push_back( epochResult );
}
return true;
}
bool Softmax::trainSoftmaxModel(UINT classLabel,SoftmaxModel &model,ClassificationData &data){
Float error = 0;
Float errorSum = 0;
Float lastErrorSum = 0;
Float delta = 0;
UINT N = data.getNumDimensions();
UINT M = data.getNumSamples();
UINT iter = 0;
bool keepTraining = true;
Random random;
VectorFloat y(M);
Vector< UINT > randomTrainingOrder(M);
//Init the model
model.init( classLabel, N );
//Setup the target vector, the input data is relabelled as positive samples (with label 1.0) and negative samples (with label 0.0)
for(UINT i=0; i<M; i++){
y[i] = data[i].getClassLabel()==classLabel ? 1.0 : 0;
}
//In most cases, the training data is grouped into classes (100 samples for class 1, followed by 100 samples for class 2, etc.)
//This can cause a problem for stochastic gradient descent algorithm. To avoid this issue, we randomly shuffle the order of the
//training samples. This random order is then used at each epoch.
for(UINT i=0; i<M; i++){
randomTrainingOrder[i] = i;
}
std::random_shuffle(randomTrainingOrder.begin(), randomTrainingOrder.end());
//Run the main stochastic gradient descent training algorithm
while( keepTraining ){
//Run one epoch of training using stochastic gradient descent
errorSum = 0;
for(UINT m=0; m<M; m++){
//Select the random sample
UINT i = randomTrainingOrder[m];
//Compute the error, given the current weights
error = y[i] - model.compute( data[i].getSample() );
errorSum += error;
//Update the weights
for(UINT j=0; j<N; j++){
model.w[j] += learningRate * error * data[i][j];
}
model.w0 += learningRate * error;
}
//Compute the error
delta = fabs( errorSum-lastErrorSum );
lastErrorSum = errorSum;
//Check to see if we should stop
if( delta <= minChange ){
keepTraining = false;
}
if( ++iter >= maxNumEpochs ){
keepTraining = false;
}
trainingLog << "Epoch: " << iter << " TotalError: " << errorSum << " Delta: " << delta << std::endl;
}
return true;
}
