public void consumeVocabulary(VocabularyHolder holder) {
for (VocabularyWord word : holder.getVocabulary()) {
if (!this.containsWord(word.getWord())) {
this.addWord(word);
} else {
holder.incrementWordCounter(word.getWord());
}
}
}
/**
* 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();
}
}
/**
* 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()
+ "]");
}
/**
* 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.");
}
public void addWord(VocabularyWord word) {
vocabulary.put(word.getWord(), word);
}
