问题
This is what I tried. Depends on what does user put into the function I want to add String or Double to new Chunk.
package org.apache.spark.h2o.utils
import water.fvec.{NewChunk, Frame, Chunk}
import water._
import water.parser.ValueString
class ReplaceNa[T >: Any](a: T) extends MRTask{
override def map(c: Chunk, nc: NewChunk): Unit = {
for (row <- 0 until c.len()) {
a match{
case s: ValueString if(c.isNA(row)) => nc.addStr(s)
case d: Double if(c.isNA(row)) => nc.addNum(d)
}
}
}
}
But I got error
error: value outputFrame is not a member of Nothing
pred.add(new ReplaceNa(3).doAll(1, pred.vec(4)).outputFrame(Array("s"), null))
Thanks for your help!
回答1:
I have several comments:
- check for NA outside the switch branch
- you are missing non-NA case hence you are generating vector which is shorter than input vector (i expect you would like to generate the same length vector)
Regarding generics, you need to provide type specialization. For example, something like the following snippet:
class ReplaceNA[T](val value: T)(implicit add: TAdd[T]) extends MRTask[ReplaceNA[T]] {
override def map(c: Chunk, nc: NewChunk): Unit = {
for (row <- 0 until c.len()) {
// Replace NAs by given value
if (c.isNA(row)) {
add.addValue(nc, value)
} else {
// Do something with default value
nc.addNA()
}
}
}
}
trait TAdd[T] extends Serializable {
def addValue(nc: NewChunk, value: T)
}
object TAdd extends Serializable {
implicit val addDouble = new TAdd[Double] { def addValue(nc: NewChunk, value: Double) = nc.addNum(value) }
implicit val addFloat = new TAdd[Float] { def addValue(nc: NewChunk, value: Float) = nc.addNum(value) }
implicit val addValueString = new TAdd[ValueString] { def addValue(nc: NewChunk, value: ValueString) = nc.addStr(value) }
}
来源:https://stackoverflow.com/questions/33442536/how-to-parametrize-class-and-implement-method-depends-on-type-in-scala