Package de.jungblut.math

Class ViterbiUtils

java.lang.Object

de.jungblut.math.ViterbiUtils

public final class ViterbiUtils
extends java.lang.Object

Viterbi Utilities for forward backward passes and his famous decoding algorithm for hidden markov models.

Author:

thomas.jungblut

Method Summary

Modifier and Type	Method	Description
static de.jungblut.math.DoubleMatrix	decode(de.jungblut.math.DoubleMatrix weights, de.jungblut.math.DoubleMatrix features, de.jungblut.math.DoubleMatrix featuresPerState, int classes)	Do a decoding pass on the given HMM weights, the features to decode and how many classes to predict.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Detail

decode

public static de.jungblut.math.DoubleMatrix decode(de.jungblut.math.DoubleMatrix weights,
                                                   de.jungblut.math.DoubleMatrix features,
                                                   de.jungblut.math.DoubleMatrix featuresPerState,
                                                   int classes)

Do a decoding pass on the given HMM weights, the features to decode and how many classes to predict. The output will contain a vector that contains a 1 at the index of the predicted label.

Parameters:

weights - the HMM weights.

features - the features to predict on.

featuresPerState - the matrix containing the feature vectors, precomputed for each possible state in classes. The layout is that the same feature was computed n-times, so class 0 first, class 1 next and so on and this is layed out in rows (Feature 1 | class 0, Feature 1 | class 1 ...). Feature 0 is only contained once, because it only had class zero as previous class.

classes - how many classes? 2 if binary.

Returns:

a n x m matrix where n is the number of featurevectors and m is the number of classes (in binary prediction this is just 1, 0 and 1 are the predicted labels at index 0 then).