/** This methods reset counters for all words in vocabulary */
public void resetWordCounters() {
for (VocabularyWord word : getVocabulary()) {
word.setHuffmanNode(null);
word.setFrequencyShift(null);
word.setCount(0);
}
}
public long totalWordsBeyondLimit() {
if (totalWordOccurencies == 0) {
for (VocabularyWord word : vocabulary.values()) {
totalWordOccurencies += word.getCount();
}
return totalWordOccurencies;
} else return totalWordOccurencies;
}
public void consumeVocabulary(VocabularyHolder holder) {
for (VocabularyWord word : holder.getVocabulary()) {
if (!this.containsWord(word.getWord())) {
this.addWord(word);
} else {
holder.incrementWordCounter(word.getWord());
}
}
}
/**
* All words with frequency below threshold wii be removed
*
* @param threshold exclusive threshold for removal
*/
public void truncateVocabulary(int threshold) {
logger.debug("Truncating vocabulary to minWordFrequency: [" + threshold + "]");
Set<String> keyset = vocabulary.keySet();
for (String word : keyset) {
VocabularyWord vw = vocabulary.get(word);
// please note: we're not applying threshold to SPECIAL words
if (!vw.isSpecial() && vw.getCount() < threshold) {
vocabulary.remove(word);
if (vw.getHuffmanNode() != null) idxMap.remove(vw.getHuffmanNode().getIdx());
}
}
}
// TODO: investigate, if it's worth to make this internally synchronized and virtually thread-safe
public void addWord(String word) {
if (!vocabulary.containsKey(word)) {
VocabularyWord vw = new VocabularyWord(word);
/*
TODO: this should be done in different way, since this implementation causes minWordFrequency ultimate ignoral if markAsSpecial set to TRUE
Probably the best way to solve it, is remove markAsSpecial option here, and let this issue be regulated with minWordFrequency
*/
// vw.setSpecial(markAsSpecial);
// initialize frequencyShift only if hugeModelExpected. It's useless otherwise :)
if (hugeModelExpected) vw.setFrequencyShift(new byte[retentionDelay]);
vocabulary.put(word, vw);
if (hugeModelExpected
&& minWordFrequency > 1
&& hiddenWordsCounter.incrementAndGet() % scavengerThreshold == 0) activateScavenger();
return;
}
}
/**
* Builds VocabularyHolder from VocabCache.
*
* <p>Basically we just ignore tokens, and transfer VocabularyWords, supposing that it's already
* truncated by minWordFrequency.
*
* <p>Huffman tree data is ignored and recalculated, due to suspectable flaw in dl4j huffman impl,
* and it's exsessive memory usage.
*
* <p>This code is required for compatibility between dl4j w2v implementation, and standalone w2v
*
* @param cache
*/
protected VocabularyHolder(@NonNull VocabCache cache, boolean markAsSpecial) {
this.vocabCache = cache;
for (VocabWord word : cache.tokens()) {
VocabularyWord vw = new VocabularyWord(word.getWord());
vw.setCount((int) word.getWordFrequency());
// since we're importing this word from external VocabCache, we'll assume that this word is
// SPECIAL, and should NOT be affected by minWordFrequency
vw.setSpecial(markAsSpecial);
// please note: we don't transfer huffman data, since proper way is to recalculate it after
// new words being added
if (word.getPoints() != null && !word.getPoints().isEmpty()) {
vw.setHuffmanNode(
buildNode(word.getCodes(), word.getPoints(), word.getCodeLength(), word.getIndex()));
}
vocabulary.put(vw.getWord(), vw);
}
// there's no sense building huffman tree just for UNK word
if (numWords() > 1) updateHuffmanCodes();
logger.info("Init from VocabCache is complete. " + numWords() + " word(s) were transferred.");
}
/**
* build binary tree ordered by counter.
*
* <p>Based on original w2v by google
*/
public List<VocabularyWord> updateHuffmanCodes() {
int min1i;
int min2i;
int b;
int i;
// get vocabulary as sorted list
List<VocabularyWord> vocab = this.words();
int count[] = new int[vocab.size() * 2 + 1];
int parent_node[] = new int[vocab.size() * 2 + 1];
byte binary[] = new byte[vocab.size() * 2 + 1];
// at this point vocab is sorted, with descending order
for (int a = 0; a < vocab.size(); a++) count[a] = vocab.get(a).getCount();
for (int a = vocab.size(); a < vocab.size() * 2; a++) count[a] = Integer.MAX_VALUE;
int pos1 = vocab.size() - 1;
int pos2 = vocab.size();
for (int a = 0; a < vocab.size(); a++) {
// First, find two smallest nodes 'min1, min2'
if (pos1 >= 0) {
if (count[pos1] < count[pos2]) {
min1i = pos1;
pos1--;
} else {
min1i = pos2;
pos2++;
}
} else {
min1i = pos2;
pos2++;
}
if (pos1 >= 0) {
if (count[pos1] < count[pos2]) {
min2i = pos1;
pos1--;
} else {
min2i = pos2;
pos2++;
}
} else {
min2i = pos2;
pos2++;
}
count[vocab.size() + a] = count[min1i] + count[min2i];
parent_node[min1i] = vocab.size() + a;
parent_node[min2i] = vocab.size() + a;
binary[min2i] = 1;
}
// Now assign binary code to each vocabulary word
byte[] code = new byte[MAX_CODE_LENGTH];
int[] point = new int[MAX_CODE_LENGTH];
for (int a = 0; a < vocab.size(); a++) {
b = a;
i = 0;
byte[] lcode = new byte[MAX_CODE_LENGTH];
int[] lpoint = new int[MAX_CODE_LENGTH];
while (true) {
code[i] = binary[b];
point[i] = b;
i++;
b = parent_node[b];
if (b == vocab.size() * 2 - 2) break;
}
lpoint[0] = vocab.size() - 2;
for (b = 0; b < i; b++) {
lcode[i - b - 1] = code[b];
lpoint[i - b] = point[b] - vocab.size();
}
vocab.get(a).setHuffmanNode(new HuffmanNode(lcode, lpoint, a, (byte) i));
}
idxMap.clear();
for (VocabularyWord word : vocab) {
idxMap.put(word.getHuffmanNode().getIdx(), word);
}
return vocab;
}
/**
* This method removes low-frequency words based on their frequency change between activations.
* I.e. if word has appeared only once, and it's retained the same frequency over consequence
* activations, we can assume it can be removed freely
*/
protected synchronized void activateScavenger() {
int initialSize = vocabulary.size();
List<VocabularyWord> words = new ArrayList<>(vocabulary.values());
for (VocabularyWord word : words) {
// scavenging could be applied only to non-special tokens that are below minWordFrequency
if (word.isSpecial()
|| word.getCount() >= minWordFrequency
|| word.getFrequencyShift() == null) {
word.setFrequencyShift(null);
continue;
}
// save current word counter to byte array at specified position
word.getFrequencyShift()[word.getRetentionStep()] = (byte) word.getCount();
/*
we suppose that we're hunting only low-freq words that already passed few activations
so, we assume word personal threshold as 20% of minWordFrequency, but not less then 1.
so, if after few scavenging cycles wordCount is still <= activation - just remove word.
otherwise nullify word.frequencyShift to avoid further checks
*/
int activation = Math.max(minWordFrequency / 5, 2);
logger.debug(
"Current state> Activation: ["
+ activation
+ "], retention info: "
+ Arrays.toString(word.getFrequencyShift()));
if (word.getCount() <= activation && word.getFrequencyShift()[this.retentionDelay - 1] > 0) {
// if final word count at latest retention point is the same as at the beginning - just
// remove word
if (word.getFrequencyShift()[this.retentionDelay - 1] <= activation
&& word.getFrequencyShift()[this.retentionDelay - 1] == word.getFrequencyShift()[0]) {
vocabulary.remove(word.getWord());
}
}
// shift retention history to the left
if (word.getRetentionStep() < retentionDelay - 1) {
word.incrementRetentionStep();
} else {
for (int x = 1; x < retentionDelay; x++) {
word.getFrequencyShift()[x - 1] = word.getFrequencyShift()[x];
}
}
}
logger.info(
"Scavenger was activated. Vocab size before: ["
+ initialSize
+ "], after: ["
+ vocabulary.size()
+ "]");
}
public void addWord(VocabularyWord word) {
vocabulary.put(word.getWord(), word);
}
/**
* This method is required for compatibility purposes. It just transfers vocabulary from
* VocabHolder into VocabCache
*
* @param cache
*/
public void transferBackToVocabCache(VocabCache cache, boolean emptyHolder) {
if (!(cache instanceof InMemoryLookupCache))
throw new IllegalStateException("Sorry, only InMemoryLookupCache use implemented.");
// make sure that huffman codes are updated before transfer
List<VocabularyWord> words = words(); // updateHuffmanCodes();
for (VocabularyWord word : words) {
if (word.getWord().isEmpty()) continue;
VocabWord vocabWord = new VocabWord(1, word.getWord());
// if we're transferring full model, it CAN contain HistoricalGradient for AdaptiveGradient
// feature
if (word.getHistoricalGradient() != null) {
INDArray gradient = Nd4j.create(word.getHistoricalGradient());
vocabWord.setHistoricalGradient(gradient);
}
// put VocabWord into both Tokens and Vocabs maps
((InMemoryLookupCache) cache).getVocabs().put(word.getWord(), vocabWord);
((InMemoryLookupCache) cache).getTokens().put(word.getWord(), vocabWord);
// update Huffman tree information
if (word.getHuffmanNode() != null) {
vocabWord.setIndex(word.getHuffmanNode().getIdx());
vocabWord.setCodeLength(word.getHuffmanNode().getLength());
vocabWord.setPoints(
arrayToList(word.getHuffmanNode().getPoint(), word.getHuffmanNode().getLength()));
vocabWord.setCodes(
arrayToList(word.getHuffmanNode().getCode(), word.getHuffmanNode().getLength()));
// put word into index
cache.addWordToIndex(word.getHuffmanNode().getIdx(), word.getWord());
}
// update vocabWord counter. substract 1, since its the base value for any token
// >1 hack is required since VocabCache impl imples 1 as base word count, not 0
if (word.getCount() > 1) cache.incrementWordCount(word.getWord(), word.getCount() - 1);
}
// at this moment its pretty safe to nullify all vocabs.
if (emptyHolder) {
idxMap.clear();
vocabulary.clear();
}
}
