public virtual ICounter <L> ProbabilityOf(IDatum <L, F> example)
{
// calculate the feature indices and feature values
int[] featureIndices = LogisticUtils.IndicesOf(example.AsFeatures(), featureIndex);
double[] featureValues;
if (example is RVFDatum <object, object> )
{
ICollection <double> featureValuesCollection = ((RVFDatum <object, object>)example).AsFeaturesCounter().Values();
featureValues = LogisticUtils.ConvertToArray(featureValuesCollection);
}
else
{
featureValues = new double[example.AsFeatures().Count];
Arrays.Fill(featureValues, 1.0);
}
// calculate probability of each class
ICounter <L> result = new ClassicCounter <L>();
int numClasses = labelIndex.Size();
double[] sigmoids = LogisticUtils.CalculateSigmoids(weights, featureIndices, featureValues);
for (int c = 0; c < numClasses; c++)
{
L label = labelIndex.Get(c);
result.IncrementCount(label, sigmoids[c]);
}
return(result);
}
protected internal override void Calculate(double[] thetasArray)
{
ClearResults();
double[][] thetas = new double[][] { };
LogisticUtils.Unflatten(thetasArray, thetas);
for (int i = 0; i < data.Length; i++)
{
int[] featureIndices = data[i];
double[] featureValues = dataValues[i];
double[] sums = LogisticUtils.CalculateSums(thetas, featureIndices, featureValues);
for (int c = 0; c < numClasses; c++)
{
double sum = sums[c];
value -= sum * labels[i][c];
if (c == 0)
{
continue;
}
int offset = (c - 1) * numFeatures;
double error = Math.Exp(sum) - labels[i][c];
for (int f = 0; f < featureIndices.Length; f++)
{
int index = featureIndices[f];
double x = featureValues[f];
derivative[offset + index] -= error * x;
}
}
}
// incorporate prior
if (prior.GetType().Equals(LogPrior.LogPriorType.Null))
{
return;
}
double sigma = prior.GetSigma();
for (int c_1 = 0; c_1 < numClasses; c_1++)
{
if (c_1 == 0)
{
continue;
}
int offset = (c_1 - 1) * numFeatures;
for (int j = 0; j < numL2Parameters; j++)
{
double theta = thetasArray[offset + j];
value += theta * theta / (sigma * 2.0);
derivative[offset + j] += theta / sigma;
}
}
}
public virtual MultinomialLogisticClassifier <L, F> TrainClassifier(GeneralDataset <L, F> dataset)
{
numClasses = dataset.NumClasses();
numFeatures = dataset.NumFeatures();
data = dataset.GetDataArray();
if (dataset is RVFDataset <object, object> )
{
dataValues = dataset.GetValuesArray();
}
else
{
dataValues = LogisticUtils.InitializeDataValues(data);
}
AugmentFeatureMatrix(data, dataValues);
labels = dataset.GetLabelsArray();
return(new MultinomialLogisticClassifier <L, F>(TrainWeights(), dataset.featureIndex, dataset.labelIndex));
}
private double[][] TrainWeights()
{
QNMinimizer minimizer = new QNMinimizer(15, true);
minimizer.UseOWLQN(true, lambda);
IDiffFunction objective = new ShiftParamsLogisticObjectiveFunction(data, dataValues, ConvertLabels(labels), numClasses, numFeatures + data.Length, numFeatures, prior);
double[] augmentedThetas = new double[(numClasses - 1) * (numFeatures + data.Length)];
augmentedThetas = minimizer.Minimize(objective, 1e-4, augmentedThetas);
// calculate number of non-zero parameters, for debugging
int count = 0;
for (int j = numFeatures; j < augmentedThetas.Length; j++)
{
if (augmentedThetas[j] != 0)
{
count++;
}
}
Redwood.Log("NUM NONZERO PARAMETERS: " + count);
double[][] thetas = new double[][] { };
LogisticUtils.Unflatten(augmentedThetas, thetas);
return(thetas);
}
