Filter a DataSet in terms of another DataSet in Scala flink

Question

I am trying to replicate this python code:

cond_entropy_x = np.array([entropy(x[y == v]) for v in uy])

Where x and y are vectors, and uy are the unique values on y , for example 0,1 .

In flink, I have:

val uy = y.distinct.collect
val condHx = for (i ← uy)
    yield entropy(x.filterWithBcVariable(y)((_, yy) ⇒ yy == i))

However, it seems filterWithBcVariable does not take every value on y , it only takes the first one.

I've also tried:

for (i ← values) yield y.join(x).where(a ⇒ a).equalTo(_ ⇒ i)

But I ran out of memory.

How could I filter x in terms of the values on y ?

Something like x.zip(y) would do it, but it is not supported.

Any ideas?

Answer 1

I came out with one solution, may be is not the best, but at least its working.

Now, instead of passing x and y as separated DataSets , I am passing a DataSet[LabeledVector] with only a column:

val xy = input.map(lv ⇒ LabeledVector(lv.label, DenseVector(lv.vector(0))))

Then I pass xy to my function:

def conditionalEntropy(xy: DataSet[LabeledVector]): Double = {
    // Get the label
    val y = xy map (_.label)
    // Get probs for the label
    val p = probs(y).toArray.asBreeze
    // Get unique values in label
    val values = y.distinct.collect
    // Compute Conditional Entropy
    val condH = for (i ← values)
      yield entropy(xy.filter(_.label == i))
    p.dot(seq2Breeze(condH))
  }