I guess this is going a bit overboard with overloa...

# kotlindl

[Peter]

I guess this is going a bit overboard with overloading operator functions 🙂 , but wanted to see if it is possible to imitate Numpy array access with NDArray, and actually not very hard at all.

@file:Suppress("ConstantConditionIf")
import org.nd4j.linalg.api.ndarray.INDArray
import org.nd4j.linalg.indexing.NDArrayIndex

/**
 * Add stride syntax to int ranges: 0..100..2
 *
 * @param step
 * @return
 */
operator fun IntRange.rangeTo(step: Int) : IntProgression {
    return IntProgression.fromClosedRange(first, last, step)
}

/**
 * Simplify NDArray access: a[0..4, 16..32..2, END-10..END-5]
 *
 * @param indexes
 * @return
 */
operator fun INDArray.get(vararg indexes: IntProgression) : INDArray {
    var dimension = 0
    val intervals = indexes.map {
        val first = if (it.first < 0) size(dimension).toInt() + it.first + 1 else it.first
        val last = if (it.last < 0) size(dimension).toInt() + it.last + 1 else it.last
        dimension++
        NDArrayIndex.interval(first, it.step, last, false)!!
    }.toTypedArray()
    return get(*intervals)
}

// Small trick so the compiler doesn't infer empty ranges
val END = if (true) -1 else 1

PS I don't really use this, so might contain some bugs.

[altavir]

I think it is a wrong chat for that. Look into #mathematics. MultiK library does exactly numpy-like overloading and KMath has basic bindings for ND4J, we recently discussed just that.

[Ролан]

your get function returns a copy or not? If that's numpy behaviour.

[Ролан]

we implemented non copying slicing in kmath

[Ролан]

https://github.com/mipt-npm/kmath/blob/82d8394a9fd19461e90a331157e5ac4eee705aab/km[…]st/kotlin/space/kscience/kmath/tensors/core/TestDoubleTensor.kt

[Ролан]

more work is needed before we release that

[Ролан]

but should happen in the coming couple of months

[Ролан]

varargs caused us trouble on multiplatform actually so we decided to get a behaviour like array[i][j][k] etc. to subset each dimension

[Ролан]

but that's also familiar to numpy users

[Peter]

Will keep an eye out for kmath, sounds interesting.

BTW, in my sample snippet, the get shouldn't create a copy (at least according to some ND4J doc I read).