private static RealVector BatchGradient(double[] logloss, TrainingInstance[] batch)
{
int dimensions = batch[0].GetFeatureCount();
int batchSize = batch.Length;
RealVector toReplicate = new RealVector(logloss);
RealVector labels = new RealVector(batchSize);
for (int i = 0 ; i < batchSize; i++)
{
labels.GetFeatures()[i] = batch[i].GetLabel();
}
labels.Add(1).ScaleThis(0.5);
toReplicate.Subtract(labels);
RealVector[] repmat = new RealVector[dimensions];
for (int i = 0; i < dimensions; i++) {
repmat[i] = new RealVector(toReplicate);
}
for (int i = 0; i < dimensions; i++) {
for(int j = 0; j < batchSize; j++) {
repmat[i].w[j] *= batch[j].GetFeatures().w[i];
}
}
RealVector result = new RealVector(dimensions);
for (int i = 0; i < dimensions;i++) {
result.w[i] = repmat[i].Average();
}
return result;
}
public RealVector Train(int maxIter)
{
var w = new RealVector(defaultWeights);
for (int i = 0; i < maxIter; i++) {
TrainingInstance[] batch = CreateBatch(SelectSubset(subsampleSize), w);
RealVector grad = BatchGradient(LogisticLoss(batch, w), batch);
w.Subtract(grad);
if(grad.GetNorm() < epsilon) {
break;
}
}
return w;
}
