Reorganized TokenClassificationHead to use the Sequential module. Added optional nonlinearity. #59
Conversation
|
@kwalcock : I added you as reviewer here and I'd be happy if you are willing to continue keeping an eye on these projects. I see you as a key contributor to these projects, which will continue as open source. But, of course, this is completely optional! |
| if(nonLin.isDefined) { | ||
| for (matrix <- outputs) { | ||
| for (i <- 0 until Math.rows(matrix)) { | ||
| val row = Math.row(matrix, i) | ||
| for (j <- 0 until Math.cols(matrix)) { | ||
| val orig = Math.get(row, j) | ||
| Math.set(row, j, nonLin.get.compute(orig)) | ||
| } | ||
| } | ||
| } | ||
| } |
There was a problem hiding this comment.
It looks like we don't have a map method. It might be nice to hide these details and maybe make it more efficient with code like
def map(matrix: MathRowMatrix, f: MathValue => MathValue): Unit = {
val iterator = matrix.iterator(true, 0, 0, matrix.getNumRows, matrix.getNumCols)
while (iterator.hasNext) {
val oldValue = iterator.next()
val newValue = f(oldValue)
iterator.set(newValue)
}
}Then the Encoder would have
nonLinOpt.foreach { nonLin =>
outputs.foreach { matrix =>
Math.map(matrix, nonLin.compute)
}
}There will be an example PR with other details soon.
| @@ -0,0 +1,15 @@ | |||
| package org.clulab.scala_transformers.encoder | |||
|
|
|||
| import org.clulab.scala_transformers.encoder.math.EjmlMath.MathValue | |||
There was a problem hiding this comment.
There's a way to make this independent of EjmlMath.
| val row = Math.row(matrix, i) | ||
| for (j <- 0 until Math.cols(matrix)) { | ||
| val orig = Math.get(row, j) | ||
| Math.set(row, j, nonLin.get.compute(orig)) |
There was a problem hiding this comment.
If this does stay, the nonLin.get might be moved outside the triply nested loops.
|
PS I'm not sure what happened to sbt that is preventing testing. I'll try to check up on it. |
No description provided.