public void estimate(
InstanceList documents,
int numIterations,
int showTopicsInterval,
int outputModelInterval,
String outputModelFilename,
Randoms r) {
ilist = documents;
uniAlphabet = ilist.getDataAlphabet();
biAlphabet = ((FeatureSequenceWithBigrams) ilist.get(0).getData()).getBiAlphabet();
numTypes = uniAlphabet.size();
numBitypes = biAlphabet.size();
int numDocs = ilist.size();
topics = new int[numDocs][];
grams = new int[numDocs][];
docTopicCounts = new int[numDocs][numTopics];
typeNgramTopicCounts = new int[numTypes][2][numTopics];
unitypeTopicCounts = new int[numTypes][numTopics];
bitypeTopicCounts = new int[numBitypes][numTopics];
tokensPerTopic = new int[numTopics];
bitokensPerTopic = new int[numTypes][numTopics];
tAlpha = alpha * numTopics;
vBeta = beta * numTypes;
vGamma = gamma * numTypes;
long startTime = System.currentTimeMillis();
// Initialize with random assignments of tokens to topics
// and finish allocating this.topics and this.tokens
int topic, gram, seqLen, fi;
for (int di = 0; di < numDocs; di++) {
FeatureSequenceWithBigrams fs = (FeatureSequenceWithBigrams) ilist.get(di).getData();
seqLen = fs.getLength();
numTokens += seqLen;
topics[di] = new int[seqLen];
grams[di] = new int[seqLen];
// Randomly assign tokens to topics
int prevFi = -1, prevTopic = -1;
for (int si = 0; si < seqLen; si++) {
// randomly sample a topic for the word at position si
topic = r.nextInt(numTopics);
// if a bigram is allowed at position si, then sample a gram status for it.
gram = (fs.getBiIndexAtPosition(si) == -1 ? 0 : r.nextInt(2));
if (gram != 0) biTokens++;
topics[di][si] = topic;
grams[di][si] = gram;
docTopicCounts[di][topic]++;
fi = fs.getIndexAtPosition(si);
if (prevFi != -1) typeNgramTopicCounts[prevFi][gram][prevTopic]++;
if (gram == 0) {
unitypeTopicCounts[fi][topic]++;
tokensPerTopic[topic]++;
} else {
bitypeTopicCounts[fs.getBiIndexAtPosition(si)][topic]++;
bitokensPerTopic[prevFi][topic]++;
}
prevFi = fi;
prevTopic = topic;
}
}
for (int iterations = 0; iterations < numIterations; iterations++) {
sampleTopicsForAllDocs(r);
if (iterations % 10 == 0) System.out.print(iterations);
else System.out.print(".");
System.out.flush();
if (showTopicsInterval != 0 && iterations % showTopicsInterval == 0 && iterations > 0) {
System.out.println();
printTopWords(5, false);
}
if (outputModelInterval != 0 && iterations % outputModelInterval == 0 && iterations > 0) {
this.write(new File(outputModelFilename + '.' + iterations));
}
}
System.out.println(
"\nTotal time (sec): " + ((System.currentTimeMillis() - startTime) / 1000.0));
}
private void sampleTopicsForOneDoc(
FeatureSequenceWithBigrams oneDocTokens,
int[] oneDocTopics,
int[] oneDocGrams,
int[] oneDocTopicCounts, // indexed by topic index
double[] uniTopicWeights, // length==numTopics
double[] biTopicWeights, // length==numTopics*2: joint topic/gram sampling
Randoms r) {
int[] currentTypeTopicCounts;
int[] currentBitypeTopicCounts;
int[] previousBitokensPerTopic;
int type, bitype, oldGram, nextGram, newGram, oldTopic, newTopic;
double topicWeightsSum, tw;
// xxx int docLen = oneDocTokens.length;
int docLen = oneDocTokens.getLength();
// Iterate over the positions (words) in the document
for (int si = 0; si < docLen; si++) {
type = oneDocTokens.getIndexAtPosition(si);
bitype = oneDocTokens.getBiIndexAtPosition(si);
// if (bitype == -1) System.out.println ("biblock "+si+" at "+uniAlphabet.lookupObject(type));
oldTopic = oneDocTopics[si];
oldGram = oneDocGrams[si];
nextGram = (si == docLen - 1) ? -1 : oneDocGrams[si + 1];
// nextGram = (si == docLen-1) ? -1 : (oneDocTokens.getBiIndexAtPosition(si+1) == -1 ? 0 : 1);
boolean bigramPossible = (bitype != -1);
assert (!(!bigramPossible && oldGram == 1));
if (!bigramPossible) {
// Remove this token from all counts
oneDocTopicCounts[oldTopic]--;
tokensPerTopic[oldTopic]--;
unitypeTopicCounts[type][oldTopic]--;
if (si != docLen - 1) {
typeNgramTopicCounts[type][nextGram][oldTopic]--;
assert (typeNgramTopicCounts[type][nextGram][oldTopic] >= 0);
}
assert (oneDocTopicCounts[oldTopic] >= 0);
assert (tokensPerTopic[oldTopic] >= 0);
assert (unitypeTopicCounts[type][oldTopic] >= 0);
// Build a distribution over topics for this token
Arrays.fill(uniTopicWeights, 0.0);
topicWeightsSum = 0;
currentTypeTopicCounts = unitypeTopicCounts[type];
for (int ti = 0; ti < numTopics; ti++) {
tw =
((currentTypeTopicCounts[ti] + beta) / (tokensPerTopic[ti] + vBeta))
* ((oneDocTopicCounts[ti]
+ alpha)); // additional term is constance across all topics
topicWeightsSum += tw;
uniTopicWeights[ti] = tw;
}
// Sample a topic assignment from this distribution
newTopic = r.nextDiscrete(uniTopicWeights, topicWeightsSum);
// Put that new topic into the counts
oneDocTopics[si] = newTopic;
oneDocTopicCounts[newTopic]++;
unitypeTopicCounts[type][newTopic]++;
tokensPerTopic[newTopic]++;
if (si != docLen - 1) typeNgramTopicCounts[type][nextGram][newTopic]++;
} else {
// Bigram is possible
int prevType = oneDocTokens.getIndexAtPosition(si - 1);
int prevTopic = oneDocTopics[si - 1];
// Remove this token from all counts
oneDocTopicCounts[oldTopic]--;
typeNgramTopicCounts[prevType][oldGram][prevTopic]--;
if (si != docLen - 1) typeNgramTopicCounts[type][nextGram][oldTopic]--;
if (oldGram == 0) {
unitypeTopicCounts[type][oldTopic]--;
tokensPerTopic[oldTopic]--;
} else {
bitypeTopicCounts[bitype][oldTopic]--;
bitokensPerTopic[prevType][oldTopic]--;
biTokens--;
}
assert (oneDocTopicCounts[oldTopic] >= 0);
assert (typeNgramTopicCounts[prevType][oldGram][prevTopic] >= 0);
assert (si == docLen - 1 || typeNgramTopicCounts[type][nextGram][oldTopic] >= 0);
assert (unitypeTopicCounts[type][oldTopic] >= 0);
assert (tokensPerTopic[oldTopic] >= 0);
assert (bitypeTopicCounts[bitype][oldTopic] >= 0);
assert (bitokensPerTopic[prevType][oldTopic] >= 0);
assert (biTokens >= 0);
// Build a joint distribution over topics and ngram-status for this token
Arrays.fill(biTopicWeights, 0.0);
topicWeightsSum = 0;
currentTypeTopicCounts = unitypeTopicCounts[type];
currentBitypeTopicCounts = bitypeTopicCounts[bitype];
previousBitokensPerTopic = bitokensPerTopic[prevType];
for (int ti = 0; ti < numTopics; ti++) {
newTopic = ti << 1; // just using this variable as an index into [ti*2+gram]
// The unigram outcome
tw =
(currentTypeTopicCounts[ti] + beta)
/ (tokensPerTopic[ti] + vBeta)
* (oneDocTopicCounts[ti] + alpha)
* (typeNgramTopicCounts[prevType][0][prevTopic] + delta1);
topicWeightsSum += tw;
biTopicWeights[newTopic] = tw;
// The bigram outcome
newTopic++;
tw =
(currentBitypeTopicCounts[ti] + gamma)
/ (previousBitokensPerTopic[ti] + vGamma)
* (oneDocTopicCounts[ti] + alpha)
* (typeNgramTopicCounts[prevType][1][prevTopic] + delta2);
topicWeightsSum += tw;
biTopicWeights[newTopic] = tw;
}
// Sample a topic assignment from this distribution
newTopic = r.nextDiscrete(biTopicWeights, topicWeightsSum);
// Put that new topic into the counts
newGram = newTopic % 2;
newTopic /= 2;
// Put that new topic into the counts
oneDocTopics[si] = newTopic;
oneDocGrams[si] = newGram;
oneDocTopicCounts[newTopic]++;
typeNgramTopicCounts[prevType][newGram][prevTopic]++;
if (si != docLen - 1) typeNgramTopicCounts[type][nextGram][newTopic]++;
if (newGram == 0) {
unitypeTopicCounts[type][newTopic]++;
tokensPerTopic[newTopic]++;
} else {
bitypeTopicCounts[bitype][newTopic]++;
bitokensPerTopic[prevType][newTopic]++;
biTokens++;
}
}
}
}
