Shape manipulation

Overview

Shape manipulation changes how the same data is interpreted across dimensions. Some operations return views over the same storage, while others create new arrays. Use arr.shape and arr.dim to verify the resulting layout.

Reshape

reshape reinterprets the data with a new shape without changing element order. The new shape must contain the same number of elements as the original array.

val a = mk.ndarray(mk[1, 2, 3, 4, 5, 6])
val b = a.reshape(2, 3)
/*
[[1, 2, 3],
 [4, 5, 6]]
 */

val c = b.reshape(3, 2)
/*
[[1, 2],
 [3, 4],
 [5, 6]]
 */

There is no -1 dimension inference, so all dimensions must be specified explicitly. When possible, reshape returns a view; otherwise it may copy data to keep storage consistent.

Flatten

flatten always returns a one-dimensional copy of the array.

val m = mk.ndarray(mk[mk[1, 2], mk[3, 4]])
val flat = m.flatten()
// [1, 2, 3, 4]

Transpose

transpose reorders axes. With no arguments, it reverses the axes order.

val m = mk.ndarray(mk[mk[1, 2], mk[3, 4], mk[5, 6]])
val t = m.transpose()
/*
[[1, 3, 5],
 [2, 4, 6]]
 */

For higher dimensions, provide the new axes order:

val x = mk.d3array(2, 3, 4) { it }
val y = x.transpose(1, 0, 2) // swap axes 0 and 1

transpose returns a view with new shape and strides.

Squeeze and unsqueeze

squeeze removes axes of size 1. Use it to drop redundant dimensions.

val a = mk.zeros<Double>(1, 3, 1)
val b = a.squeeze()
// shape is (3)

unsqueeze inserts axes of size 1 at the specified positions:

val v = mk.ndarray(mk[1, 2, 3])
val col = v.unsqueeze(1) // shape (3, 1)
val row = v.unsqueeze(0) // shape (1, 3)

You can also use expandDims/expandNDims as helpers for adding size-1 axes.

Concatenate and stack

Use cat to concatenate arrays along an axis. The shapes must match on all other axes.

val left = mk.ndarray(mk[mk[1, 2], mk[3, 4]])
val right = mk.ndarray(mk[mk[5, 6], mk[7, 8]])

val byRows = left.cat(right, axis = 0)
/*
[[1, 2],
 [3, 4],
 [5, 6],
 [7, 8]]
 */

val byCols = left.cat(right, axis = 1)
/*
[[1, 2, 5, 6],
 [3, 4, 7, 8]]
 */

Use mk.stack to add a new axis and combine arrays of the same shape:

val a = mk.ndarray(mk[1, 2, 3])
val b = mk.ndarray(mk[4, 5, 6])
val stacked = mk.stack(a, b) // shape (2, 3)

cat and stack create new arrays and copy data.

19 February 2026