之前询问过的，也给出了回复，可是自己能力不够，有些lambda不会转译。给出修改的代码和需要翻译的代码，麻烦大神帮忙一下，非常感谢！
```class MultiSenseSkipGramEmbeddingModel
negative = opts.negative.value
window = opts.window.value
rng = random
sample = opts.sample.value.toDouble
def process(doc: String):
Int = {
sen = doc.stripLineEnd.split(' ').map(word => vocab.getId(word.toLowerCase())).filter(id => id != -1)
wordCount = sen.size
```var rightSense = 0
if (kmeans == 1)
rightSense = cbow_predict_kmeans(currWord, contexts)
else if (dpmeans == 1)
rightSense = cbow_predict_dpmeans(currWord, contexts)
else
rightSense = cbow_predict(currWord, contexts)
contexts.foreach(context => {
trainer.processExample(new MSCBOWSkipGramNegSamplingExample(this, currWord, rightSense, context, 1))
(0 until negative).foreach(neg => trainer.processExample(new MSCBOWSkipGramNegSamplingExample(this, currWord, rightSense, vocab.getRandWordId, -1)))
})
}
return wordCount
}
def cbow_predict(word : Int, contexts: Seq[Int]): Int = {
val contextsEmbedding = new DenseTensor1(D, 0)
contexts.foreach(context => contextsEmbedding.+=(global_weights(context).value))
var sense = 0
if (learnMultiVec(word)) {
var maxdot = contextsEmbedding.dot(sense_weights(word)(0).value)
for (s <- 1 until S) {
val dot = contextsEmbedding.dot(sense_weights(word)(s).value)
if (dot > maxdot) {
maxdot = dot
sense = s
}
}
}
sense
}
def cbow_predict_kmeans(word: Int, contexts: Seq[Int]): Int = {
val contextsEmbedding = new DenseTensor1(D, 0)
contexts.foreach(context => contextsEmbedding.+=(global_weights(context).value))
var sense = 0
if (learnMultiVec(word)) {
var minDist = Double.MaxValue
for (s <- 0 until ncluster(word)) {
val mu = clusterCenter(word)(s)/(clusterCount(word)(s))
val dist = 1 - TensorUtils.cosineDistance(contextsEmbedding, mu)
if (dist < minDist) {
minDist = dist
sense = s
}
}
}
clusterCenter(word)(sense).+=(contextsEmbedding)
clusterCount(word)(sense) += 1
sense
}
def cbow_predict_dpmeans(word: Int, contexts: Seq[Int]): Int = {
val contextsEmbedding = new DenseTensor1(D, 0)
contexts.foreach(context => contextsEmbedding.+=(global_weights(context).value))
var sense = 0
if (learnMultiVec(word)) {
var minDist = Double.MaxValue
var ncluster_word = ncluster(word)
val nC = if (ncluster_word == S) S else ncluster_word + 1
var prob = new Array[Double](nC)
for (s <- 0 until ncluster_word) {
val mu = clusterCenter(word)(s) / (clusterCount(word)(s))
val dist = 1 - TensorUtils.cosineDistance(contextsEmbedding, mu)
prob(s) = dist
if (dist < minDist) {
minDist = dist
sense = s
}
}
if (ncluster_word < S) {
if (createClusterlambda < minDist) {
prob(ncluster_word) = createClusterlambda
sense = ncluster_word
ncluster(word) = ncluster_word + 1
}
}
}
clusterCenter(word)(sense).+=(contextsEmbedding)
clusterCount(word)(sense) += 1
sense
}
def subSample(word: Int): Int = {
val ran = vocab.getSubSampleProb(word)
val real_ran = rng.nextInt(0xFFFF) / 0xFFFF.toDouble
return if (ran < real_ran) -1 else word
}
}
class MSCBOWSkipGramNegSamplingExample(model: MultiSenseWordEmbeddingModel, word: Int, sense : Int, context : Int, label: Int) extends Example {
def accumulateValueAndGradient(value: DoubleAccumulator, gradient: WeightsMapAccumulator): Unit = {
val wordEmbedding = model.sense_weights(word)(sense).value
val contextEmbedding = model.global_weights(context).value
val score: Double = wordEmbedding.dot(contextEmbedding)
val exp: Double = math.exp(-score) // TODO : pre-compute expTable similar to word2vec
var objective: Double = 0.0
var factor: Double = 0.0
// for POS Label
if (label == 1) {
objective = -math.log1p(exp) // log1p -> log(1+x)
factor = exp / (1 + exp)
}
// for NEG Label
if (label == -1) {
objective = -score - math.log1p(exp)
factor = -1 / (1 + exp)
}
if (value ne null) value.accumulate(objective)
if (gradient ne null) {
gradient.accumulate(model.sense_weights(word)(sense), contextEmbedding, factor)
// don;t update if global_weights is fixed.
if (model.updateGlobal == 1) gradient.accumulate(model.global_weights(context), wordEmbedding, factor)
}
}
}
比较多，麻烦了！
