C# Tensor Extensions

[Onslauth (onslauth)]

Hi,

Are there any libraries that provide extensions to the tensor class that implement standard functions like permute or flatten, etc. for C# ?

[Advait Kawale (Advait251206)]

Hi,

If you mean the Tensor<T> type in Microsoft.ML.OnnxRuntime.Tensors, ONNX Runtime provides some tensor utilities, including Reshape, but its C# API does not appear to provide a general PyTorch-style permute() method. The API reference lists the available tensor classes and utilities here: https://onnxruntime.ai/docs/api/csharp/api/Microsoft.ML.OnnxRuntime.Tensors.html

A few options:

• For basic shape changes: use the built-in Reshape() where the tensor’s element order is already suitable. Note that reshaping is not the same as permuting axes.
• For broader tensor manipulation: consider a library such as [TorchSharp](https://github.com/dotnet/TorchSharp), which provides tensor operations such as permute and flatten. You would need to account for conversion or data-copying when moving between TorchSharp tensors and ONNX Runtime tensors.
• If you only need a few operations: writing small extension methods for DenseTensor<T> may be simpler than introducing another dependency. Be sure to handle dimensions, strides, and element ordering correctly.

Are you looking specifically for extensions that operate directly on ONNX Runtime’s Tensor<T>/DenseTensor<T>, or would a separate C# tensor library also work?

If this answer helped or pointed you in the right direction, I'd appreciate it if you could mark it as the accepted answer so it's easier for others with the same issue to find.

Also, if you found my contribution useful, I'd appreciate it if you could check out my GitHub profile, follow me, and star any repositories you find interesting.

GitHub: https://github.com/Advait251206

[jinwoo (dev-jinwoohong)]

For a .NET 10 project, another option is Microsoft's System.Numerics.Tensors package. It has PermuteDimensions, Reshape, and FlattenTo:

using NumericsTensor = System.Numerics.Tensors.Tensor;

float[] data = [0, 1, 2, 3, 4, 5];
var t = NumericsTensor.Create(data, [2, 3]);
var p = NumericsTensor.PermuteDimensions(t, [1, 0]);
float[] flat = new float[6];
p.FlattenTo(flat);
// p has shape [3, 2]; flat contains [0, 3, 1, 4, 2, 5].

These operate on System.Numerics.Tensors.Tensor<T>, a separate type from ORT's DenseTensor<T>. You would pass the resulting buffer and shape back to ORT; they are not extensions on the ORT class itself. Use NumericsTensor.Reshape(t, [6]) if you want a 1D tensor rather than a flat array.

The higher-level tensor APIs became stable with .NET 10. The released permutation implementation supports arbitrary axis ordering.