private void writeString(final ByteBuf buf, final AsciiString s) {
final int encodedLength = Huffman.encodedLength(s);
if (encodedLength < s.length()) {
Hpack.writeHuffmanString(buf, s, encodedLength);
} else {
Hpack.writeRawString(buf, s);
}
}
@Test
public void testHuffmanDecode() throws Exception {
final String expected = "https://www.example.com";
final byte[] encoded =
TestUtil.bytes(
0x9d, 0x29, 0xad, 0x17, 0x18, 0x63, 0xc7, 0x8f, 0x0b, 0x97, 0xc8, 0xe9, 0xae, 0x82,
0xae, 0x43, 0xd3);
final ByteBuf in = Unpooled.wrappedBuffer(encoded);
final ByteBuf out = Unpooled.buffer();
Huffman.decode(in, out);
final String decoded = out.toString(US_ASCII);
assertThat(decoded, is(expected));
}
// Training word2vec based on corpus
public void train(JavaRDD<String> corpusRDD) throws Exception {
log.info("Start training ...");
// SparkContext
final JavaSparkContext sc = new JavaSparkContext(corpusRDD.context());
// Pre-defined variables
Map<String, Object> tokenizerVarMap = getTokenizerVarMap();
Map<String, Object> word2vecVarMap = getWord2vecVarMap();
// Variables to fill in in train
// final JavaRDD<AtomicLong> sentenceWordsCountRDD;
final JavaRDD<List<VocabWord>> vocabWordListRDD;
// final JavaPairRDD<List<VocabWord>, Long> vocabWordListSentenceCumSumRDD;
final VocabCache vocabCache;
final JavaRDD<Long> sentenceCumSumCountRDD;
// Start Training //
//////////////////////////////////////
log.info("Tokenization and building VocabCache ...");
// Processing every sentence and make a VocabCache which gets fed into a LookupCache
Broadcast<Map<String, Object>> broadcastTokenizerVarMap = sc.broadcast(tokenizerVarMap);
TextPipeline pipeline =
new TextPipeline(corpusRDD.repartition(numPartitions), broadcastTokenizerVarMap);
pipeline.buildVocabCache();
pipeline.buildVocabWordListRDD();
// Get total word count and put into word2vec variable map
word2vecVarMap.put("totalWordCount", pipeline.getTotalWordCount() / numPartitions);
// 2 RDDs: (vocab words list) and (sentence Count).Already cached
// sentenceWordsCountRDD = pipeline.getSentenceCountRDD();
vocabWordListRDD = pipeline.getVocabWordListRDD();
// Get vocabCache and broad-casted vocabCache
Broadcast<VocabCache> vocabCacheBroadcast = pipeline.getBroadCastVocabCache();
vocabCache = vocabCacheBroadcast.getValue();
//////////////////////////////////////
log.info("Building Huffman Tree ...");
// Building Huffman Tree would update the code and point in each of the vocabWord in vocabCache
Huffman huffman = new Huffman(vocabCache.vocabWords());
huffman.build();
/////////////////////////////////////
log.info("Training word2vec sentences ...");
word2vecVarMap.put("vecNum", vocabCache.numWords());
// Map<Tuple2<Integer,Integer>, INDArray> s0 = new HashMap();
Map<Pair<Integer, Integer>, INDArray> s0 = new HashMap();
for (int k = 0; k < K; k++) {
for (int i = 0; i < vocabCache.numWords(); i++) {
s0.put(new Pair(i, k), getRandomSyn0Vec(vectorLength));
}
}
for (int i = vocabCache.numWords(); i < vocabCache.numWords() * 2 - 1; i++) {
s0.put(new Pair(i, 0), Nd4j.zeros(1, vectorLength));
}
for (int i = 0; i < iterations; i++) {
System.out.println("iteration: " + i);
word2vecVarMap.put("alpha", alpha - (alpha - minAlpha) / iterations * i);
word2vecVarMap.put("minAlpha", alpha - (alpha - minAlpha) / iterations * (i + 1));
FlatMapFunction firstIterationFunction =
new FirstIterationFunction(word2vecVarMap, expTable, sc.broadcast(s0));
class MapPairFunction
implements PairFunction<Map.Entry<Integer, INDArray>, Integer, INDArray> {
public Tuple2<Integer, INDArray> call(Map.Entry<Integer, INDArray> pair) {
return new Tuple2(pair.getKey(), pair.getValue());
}
}
class Sum implements Function2<INDArray, INDArray, INDArray> {
public INDArray call(INDArray a, INDArray b) {
return a.add(b);
}
}
// @SuppressWarnings("unchecked")
JavaPairRDD<Pair<Integer, Integer>, INDArray> indexSyn0UpdateEntryRDD =
vocabWordListRDD
.mapPartitions(firstIterationFunction)
.mapToPair(new MapPairFunction())
.cache();
Map<Pair<Integer, Integer>, Object> count = indexSyn0UpdateEntryRDD.countByKey();
indexSyn0UpdateEntryRDD = indexSyn0UpdateEntryRDD.reduceByKey(new Sum());
// Get all the syn0 updates into a list in driver
List<Tuple2<Pair<Integer, Integer>, INDArray>> syn0UpdateEntries =
indexSyn0UpdateEntryRDD.collect();
// Updating syn0
s0 = new HashMap();
for (Tuple2<Pair<Integer, Integer>, INDArray> syn0UpdateEntry : syn0UpdateEntries) {
int cc = Integer.parseInt(count.get(syn0UpdateEntry._1).toString());
// int cc = 1;
if (cc > 0) {
INDArray tmp = Nd4j.zeros(1, vectorLength).addi(syn0UpdateEntry._2).divi(cc);
s0.put(syn0UpdateEntry._1, tmp);
}
}
}
syn0 = Nd4j.zeros(vocabCache.numWords() * K, vectorLength);
for (Map.Entry<Pair<Integer, Integer>, INDArray> ss : s0.entrySet()) {
if (ss.getKey().getFirst() < vocabCache.numWords()) {
syn0.getRow(ss.getKey().getSecond() * vocabCache.numWords() + ss.getKey().getFirst())
.addi(ss.getValue());
}
}
vocab = vocabCache;
syn0.diviRowVector(syn0.norm2(1));
BufferedWriter write = new BufferedWriter(new FileWriter(new File(path), false));
for (int i = 0; i < syn0.rows(); i++) {
String word = vocab.wordAtIndex(i % vocab.numWords());
if (word == null) {
continue;
}
word = word + "(" + i / vocab.numWords() + ")";
StringBuilder sb = new StringBuilder();
sb.append(word.replaceAll(" ", "_"));
sb.append(" ");
INDArray wordVector = syn0.getRow(i);
for (int j = 0; j < wordVector.length(); j++) {
sb.append(wordVector.getDouble(j));
if (j < wordVector.length() - 1) {
sb.append(" ");
}
}
sb.append("\n");
write.write(sb.toString());
}
write.flush();
write.close();
}
