The module lacks a 'where' procedure.

[filipeclduarte]

I was looking for a procedure similar to numpy.where() from python numpy but I cannot find it. So, I think it is beneficial to create a procedure that do this task, i.e., return the indices where the condition is true.

https://github.com/numpy/numpy/blob/1b8b46b3f2f68f5be8e52e798eb91c2ac5952745/numpy/ma/core.py#L7249

[Clonkk]

Isn't this a direct application of apply_inline / map_inline ?

var xT = randomTensor[float64]([2, 3], 100.0)
xT.apply_inline:
  if x > 10:
    2*x+5
  else: 
    x*10

[Vindaar]

@Clonkk numpy.where has multiple use cases depending on the arguments (yeah, that's freaking annoying indeed). From the docs https://numpy.org/doc/stable/reference/generated/numpy.where.html:

Note
When only condition is provided, this function is a shorthand for np.asarray(condition).nonzero().

If one only hands the condition it returns the indices at which some condition is true.
Thus, it's neither served by masked/index_select, because these return / use boolean masks nor by apply/map_inline, because one does not want to do anything with the indices.

One can always approximate where in this use case like this:

let t = ... # some tensor t which we want to use `where` for
let cond = ... # our boolean condition
let mask = t.masked_select(cond) # or t[cond]
let wT = toSeq(0 ..< t.size).toTensor()[mask] 

to get the global indices. For 1D tensors that's the same as numpy.where but for ND of course we would want to get indices along each dimension.

The equivalent would be easy to do if we had nonzero. Well, we kind of do, but because I can't finish what I start apparently not. I added it here #447:

Arraymancer/src/tensor/aggregate.nim

Lines 200 to 248 in bce53ef

	proc nonzero*[T](t: Tensor[T]): Tensor[int] =
	## Returns the indices, which are non zero as a `Tensor[int]`.
	##
	## The resulting tensor is 2 dimensional and has one element for each
	## dimension in ``t``. Each of those elements contains the indicies along
	## the corresponding axis (element 0 == axis 0), which are non zero.
	##
	## Input:
	## - A tensor
	##
	## Returns:
	## - A 2D tensor with N elements, where N is the rank of ``t``
	##
	## Example:
	## .. code:: nim
	## let a = [[3, 0, 0],
	## [0, 4, 0],
	## [5, 6, 0]].toTensor()
	## assert a.nonzero == [[0, 1, 2, 2], [0, 1, 0, 1]].toTensor
	## # ^-- indices.. ^ ..for axis 0
	## # |-- indices for axis 1
	## # axis 0: [0, 1, 2, 2] refers to:
	## # - 0 -> 3 in row 0
	## # - 1 -> 4 in row 1
	## # - 2 -> 5 in row 2
	## # - 2 -> 6 in row 2
	## # axis 1: [0, 1, 0, 1] refers to:
	## # - 0 -> 3 in col 0
	## # - 1 -> 4 in col 1
	## # - 0 -> 5 in col 0
	## # - 1 -> 6 in col 1
	var count = 0 # number of non zero elements
	let mask = map_inline(t):
	block:
	let cond = x != 0.T
	if cond:
	inc count
	cond
	result = newTensor[int]([t.shape.len, count])
	var ax = 0 # current axis
	var k = 0 # counter for indices in one axis
	for idx, x in mask:
	if x:
	ax = 0
	for j in idx:
	result[ax, k] = j
	inc ax
	inc k

but I still haven't gotten around to finishing it (sorry :( ).
I'll finish up what I'm doing now and then wrap up that PR soon.

[Clonkk]

Interesting, I've never used np.where that way.

[Vindaar]

Interesting, I've never used np.where that way.

I've only ever used np.where that way, haha.

[filipeclduarte]

@Clonkk numpy.where has multiple use cases depending on the arguments (yeah, that's freaking annoying indeed). From the docs https://numpy.org/doc/stable/reference/generated/numpy.where.html:

Note
When only condition is provided, this function is a shorthand for np.asarray(condition).nonzero().

If one only hands the condition it returns the indices at which some condition is true.
Thus, it's neither served by masked/index_select, because these return / use boolean masks nor by apply/map_inline, because one does not want to do anything with the indices.

One can always approximate where in this use case like this:

let t = ... # some tensor t which we want to use `where` for
let cond = ... # our boolean condition
let mask = t.masked_select(cond) # or t[cond]
let wT = toSeq(0 ..< t.size).toTensor()[mask] 

to get the global indices. For 1D tensors that's the same as numpy.where but for ND of course we would want to get indices along each dimension.

The equivalent would be easy to do if we had nonzero. Well, we kind of do, but because I can't finish what I start apparently not. I added it here #447:

Arraymancer/src/tensor/aggregate.nim

Lines 200 to 248 in bce53ef

	proc nonzero*[T](t: Tensor[T]): Tensor[int] =
	## Returns the indices, which are non zero as a `Tensor[int]`.
	##
	## The resulting tensor is 2 dimensional and has one element for each
	## dimension in ``t``. Each of those elements contains the indicies along
	## the corresponding axis (element 0 == axis 0), which are non zero.
	##
	## Input:
	## - A tensor
	##
	## Returns:
	## - A 2D tensor with N elements, where N is the rank of ``t``
	##
	## Example:
	## .. code:: nim
	## let a = [[3, 0, 0],
	## [0, 4, 0],
	## [5, 6, 0]].toTensor()
	## assert a.nonzero == [[0, 1, 2, 2], [0, 1, 0, 1]].toTensor
	## # ^-- indices.. ^ ..for axis 0
	## # |-- indices for axis 1
	## # axis 0: [0, 1, 2, 2] refers to:
	## # - 0 -> 3 in row 0
	## # - 1 -> 4 in row 1
	## # - 2 -> 5 in row 2
	## # - 2 -> 6 in row 2
	## # axis 1: [0, 1, 0, 1] refers to:
	## # - 0 -> 3 in col 0
	## # - 1 -> 4 in col 1
	## # - 0 -> 5 in col 0
	## # - 1 -> 6 in col 1
	var count = 0 # number of non zero elements
	let mask = map_inline(t):
	block:
	let cond = x != 0.T
	if cond:
	inc count
	cond
	result = newTensor[int]([t.shape.len, count])
	var ax = 0 # current axis
	var k = 0 # counter for indices in one axis
	for idx, x in mask:
	if x:
	ax = 0
	for j in idx:
	result[ax, k] = j
	inc ax
	inc k

but I still haven't gotten around to finishing it (sorry :( ).
I'll finish up what I'm doing now and then wrap up that PR soon.

@Vindaar this help with rank-1 Tensors. numpy.where is a very useful function. I didn't know all its functionalities.