Meaning of negative index in System.Numerics.Tensors

[abgoswam]

I am curious whats a negative index in System.Numerics.Tensors represents ?

            var arrayA = new[, ,]
            {
                {
                    {0, 1, 2},
                    {3, 4, 5}
                },
                {
                    {6, 7 ,8 },
                    {9, 10 ,11 },
                },
                {
                    {12, 13 ,14 },
                    {15, 16 ,17 },
                },
                {
                    {18, 19 ,20 },
                    {21, 22 ,23 },
                }
            };

            var tensorA = arrayA.ToTensor<int>(); // tensorA shape (4, 2, 3)
            var y = tensorA[2, -1, -2]; // **7**
            var z = tensorA[0, -1, -2]; // **throws error**

I was trying to draw parallels with numpy. Numpy results made sense.

a = np.array([ [[0,1,2], [3,4,5]],  [[6,7,8], [9,10,11]],  [[12,13,14],[15,16,17]]])
y = a[2,-1,-2]  # 16,  looks ok
z = a[0,-1,-2]  # 4, looks ok

@eerhardt

Tagging subscribers to this area: @tannergooding
Notify danmosemsft if you want to be subscribed.

[eerhardt]

What you want here is a System.Index with a ^ which means "from the end".

https://docs.microsoft.com/en-us/dotnet/api/system.index?view=netcore-3.1

int[] someArray = new int[5] { 1, 2, 3, 4, 5 };
int lastElement = someArray[^1]; // lastElement = 5

However, the issue is that System.Numerics.Tensors targets netstandard1.1, but System.Index is only available on netcoreapp3.0+.

@terrajobst @tarekgh @GrabYourPitchforks - maybe since Range and Index are now stable - we could consider making a Microsoft.Bcl package for them?

[tarekgh]

maybe since Range and Index are now stable - we could consider making a Microsoft.Bcl package for them?

I believe it was decided before we are not going to create a package for Index/Range.

CC @jaredpar

[abgoswam]

thanks for the discussion.

@eerhardt . what really worries me is this:

var y = tensorA[2, -1, -2]; // **7**

the library silently returned me an answer (7)

I would have expected this to throw a 'System.IndexOutOfRangeException' similar to what multi-dimensional arrays in .NET does for negative indices

[eerhardt]

what really worries me is this:
the library silently returned me an answer (7)

Yeah, that looks like a bug. It doesn't do any argument checking:

runtime/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/Tensor.cs

Lines 668 to 673 in 103fa2c

	public virtual T this[ReadOnlySpan<int> indices]
	{
	get
	{
	return GetValue(ArrayUtilities.GetIndex(strides, indices));
	}

I believe it was decided before we are not going to create a package for Index/Range.

The discussion keeps coming up though. It recently came up for making the Utf8String package support netstandard2.0: #33357.

[jaredpar]

I believe it was decided before we are not going to create a package for Index/Range.

@terrajobst drives this process

[pgovind]

Yeah, that looks like a bug. It doesn't do any argument checking:

I randomly ran into this issue today and spent a little bit of time debugging it. In the example, tensorA is a DenseTensor. When I look at its strides property, it returns [6, 3, 1]. I'm not sure that is correct. What exactly are Strides?

Also, AFAICT, there don't seem to be any unit tests testing negative indices in the indexer!

Tagging subscribers to this area: @eiriktsarpalis, @pgovind
See info in area-owners.md if you want to be subscribed.

Issue Details

---

I am curious whats a negative index in System.Numerics.Tensors represents ?

            var arrayA = new[, ,]
            {
                {
                    {0, 1, 2},
                    {3, 4, 5}
                },
                {
                    {6, 7 ,8 },
                    {9, 10 ,11 },
                },
                {
                    {12, 13 ,14 },
                    {15, 16 ,17 },
                },
                {
                    {18, 19 ,20 },
                    {21, 22 ,23 },
                }
            };

            var tensorA = arrayA.ToTensor<int>(); // tensorA shape (4, 2, 3)
            var y = tensorA[2, -1, -2]; // **7**
            var z = tensorA[0, -1, -2]; // **throws error**

I was trying to draw parallels with numpy. Numpy results made sense.

a = np.array([ [[0,1,2], [3,4,5]],  [[6,7,8], [9,10,11]],  [[12,13,14],[15,16,17]]])
y = a[2,-1,-2]  # 16,  looks ok
z = a[0,-1,-2]  # 4, looks ok

@eerhardt

Author:	abgoswam
Assignees:	-
Labels:	area-System.Numerics.Tensors
Milestone:	Future