private float[] ComputePosterios(float[] componentScores, int numStreams)
{
float[] posteriors = componentScores;
int step = componentScores.Length / numStreams;
int startIdx = 0;
for (int i = 0; i < numStreams; i++)
{
float max = posteriors[startIdx];
for (int j = startIdx + 1; j < startIdx + step; j++)
{
if (posteriors[j] > max)
{
max = posteriors[j];
}
}
for (int j = startIdx; j < startIdx + step; j++)
{
posteriors[j] = (float)_logMath.LogToLinear(posteriors[j] - max);
}
startIdx += step;
}
return(posteriors);
}
/// <summary>
/// Gets the (linearly scaled) mixture weights of the component densities
/// </summary>
/// <returns></returns>
public float[] GetComponentWeights()
{
var mixWeights = new float[MixtureComponents.Length];
for (var i = 0; i < mixWeights.Length; i++)
mixWeights[i] = (float)LogMath.LogToLinear(MixtureWeights.Get(_Id, 0, i));
return mixWeights;
}
private void BuildSmearInfo()
{
double s0 = 0;
double r0 = 0;
_bigramSmearMap = new Dictionary <long, Float>();
var ugNumerator = new double[_unigrams.Length];
var ugDenominator = new double[_unigrams.Length];
var ugAvgLogProb = new double[_unigrams.Length];
_unigramSmearTerm = new float[_unigrams.Length];
foreach (var unigram in _unigrams)
{
var logp = unigram.LogProbability;
var p = LogMath.LogToLinear(logp);
s0 += p * logp;
r0 += p * logp * logp;
}
this.LogInfo("R0 S0 " + r0 + ' ' + s0);
for (var i = 0; i < _loadedBigramBuffers.Length; i++)
{
var bigram = GetBigramBuffer(i);
if (bigram == null)
{
_unigramSmearTerm[i] = LogMath.LogOne;
continue;
}
ugNumerator[i] = 0.0;
ugDenominator[i] = 0.0;
ugAvgLogProb[i] = 0.0;
var logugbackoff = _unigrams[i].LogBackoff;
var ugbackoff = LogMath.LogToLinear(logugbackoff);
for (var j = 0; j < bigram.NumberNGrams; j++)
{
var wordID = bigram.GetWordID(j);
var bgProb = bigram.GetNGramProbability(j);
var logugprob = _unigrams[wordID].LogProbability;
var logbgprob = _ngramProbTable[1][bgProb.ProbabilityID];
var ugprob = LogMath.LogToLinear(logugprob);
var bgprob = LogMath.LogToLinear(logbgprob);
var backoffbgprob = ugbackoff * ugprob;
double logbackoffbgprob = LogMath.LinearToLog(backoffbgprob);
ugNumerator[i] +=
(bgprob * logbgprob
- backoffbgprob * logbackoffbgprob) * logugprob;
ugDenominator[i] += (bgprob - backoffbgprob) * logugprob;
// dumpProbs(ugNumerator, ugDenominator, i, j, logugprob,
// logbgprob, ugprob, bgprob, backoffbgprob,
// logbackoffbgprob);
}
ugNumerator[i] += ugbackoff * (logugbackoff * s0 + r0);
ugAvgLogProb[i] = ugDenominator[i] + ugbackoff * s0;
ugDenominator[i] += ugbackoff * r0;
// System.out.println("n/d " + ugNumerator[i] + " " +
// ugDenominator[i]);
_unigramSmearTerm[i] = (float)(ugNumerator[i] / ugDenominator[i]);
// / unigramSmearTerm[i] =
// logMath.linearToLog(ugNumerator[i] / ugDenominator[i]);
// System.out.println("ugs " + unigramSmearTerm[i]);
}
for (var i = 0; i < _loadedBigramBuffers.Length; i++)
{
this.LogInfo("Processed " + i
+ " of " + _loadedBigramBuffers.Length);
var bigram = GetBigramBuffer(i);
if (bigram == null)
{
continue;
}
for (var j = 0; j < bigram.NumberNGrams; j++)
{
float smearTerm;
var bgProb = bigram.GetNGramProbability(j);
var logbgbackoff =
_ngramBackoffTable[2][bgProb.BackoffID];
var bgbackoff = LogMath.LogToLinear(logbgbackoff);
var k = bigram.GetWordID(j);
var trigram = LoadTrigramBuffer(i, k);
if (trigram == null)
{
smearTerm = _unigramSmearTerm[k];
}
else
{
double bgNumerator = 0;
double bgDenominator = 0;
for (var l = 0; l < trigram.NumberNGrams; l++)
{
var m = trigram.GetWordID(l);
var logtgprob =
_ngramProbTable[2][trigram.GetProbabilityID(l)];
var tgprob = LogMath.LogToLinear(logtgprob);
var logbgprob = GetBigramProb(k, m);
var bgprob = LogMath.LogToLinear(logbgprob);
var logugprob = _unigrams[m].LogProbability;
var backofftgprob = bgbackoff * bgprob;
double logbackofftgprob =
LogMath.LinearToLog(backofftgprob);
bgNumerator += (tgprob * logtgprob - backofftgprob
* logbackofftgprob) * logugprob;
bgDenominator += (tgprob - backofftgprob) * logugprob
* logugprob;
}
bgNumerator +=
bgbackoff * (logbgbackoff * ugAvgLogProb[k] - ugNumerator[k]);
bgDenominator += bgbackoff * ugDenominator[k];
// bigram.ugsmear = bg_numerator / bg_denominator;
smearTerm = (float)(bgNumerator / bgDenominator);
_smearTermCount++;
}
PutSmearTerm(i, k, smearTerm);
}
}
this.LogInfo("Smear count is " + _smearTermCount);
}
