Closed
Description
Description
StackAlongDimension always use the first parameter to calculate. please see the [Other information] for detailed analyze.
Reproduction Steps
use the code as below
using System.Numerics.Tensors;
#pragma warning disable SYSLIB5001 // Type is for evaluation purposes only and is subject to change or removal in future updates. Suppress this diagnostic to proceed.
int[] f1 = [1, 2, 3, 4];
int[] f2 = [10, 20, 30, 40];
var v1=Tensor.Create(f1, [2,2]);
var v2 = Tensor.Create(f2, [2,2]);
var v3 = StackAlongDimension1(1,[v1,v2]);
Console.WriteLine(v3.ToString(9, 9, 9));Expected behavior
output should be
[2x2x2], type = System.Int32, isPinned = False
{
{1,2}
{10,20}
{3,4}
{30,40}
}
Actual behavior
output is
[2x2x2], type = System.Int32, isPinned = False
{
{1,2}
{1,2}
{3,4}
{3,4}
}
Regression?
No response
Known Workarounds
No response
Configuration
VS 2022, .NET 8
Windows 11 [Version 10.0.26100.1742]
x64
Other information
by dive into the source code, I found the code is always using the first tensor for stacking.
public static Tensor<T> StackAlongDimension<T>(int dimension, params ReadOnlySpan<Tensor<T>> tensors)
{
if (tensors.Length < 2)
ThrowHelper.ThrowArgument_StackTooFewTensors();
for (int i = 1; i < tensors.Length; i++)
{
if (!TensorHelpers.AreLengthsTheSame<T>(tensors[0], tensors[i]))
ThrowHelper.ThrowArgument_StackShapesNotSame();
}
if (dimension < 0)
dimension = tensors[0].Rank - dimension;
Tensor<T>[] outputs = new Tensor<T>[tensors.Length];
for (int i = 0; i < tensors.Length; i++)
{
outputs[i] = Tensor.Unsqueeze(tensors[0], dimension); **<--it always use the first tensor**
}
return Tensor.ConcatenateOnDimension<T>(dimension, outputs);
}
I created a PoC version, this version works as exptected.
Tensor<T> StackAlongDimension1<T>(int dimension,Tensor<T>[] tensors)
{
if (dimension < 0)
{
dimension = tensors[0].Rank - dimension;
}
Tensor<T>[] array = new Tensor<T>[tensors.Length];
for (int i = 0; i < tensors.Length; i++)
{
array[i] = tensors[i].Unsqueeze(dimension); **<--use the correct tensor**
}
return Tensor.ConcatenateOnDimension<T>(dimension, array);
}