Torch and numpy interoperability?

[GeorgeS2019]

Is there a plan to support Torch and numpy interoperability in .NET?

Numpy is a popular open source library used for mathematical and scientific computing in Python. It enables efficient operations on large multi-dimensional arrays, and has a large ecosystem of supporting libraries:

Instead of reinventing the wheel, PyTorch interoperates really well with Numpy to leverage its existing ecosystem of tools and libraries.

Currently, users of Tensorflow.NET use either

• Numpy.NET or
• NumSharp or
• NumSharp.Lite
• {Non SciSharp} numpy.net or NumpyDotNet@nuget

The .NET community need Microsoft to support the community to (perhaps) coordinate/prioritize the different efforts of bringing Numpy to .NET.

[aiqinxuancai]

When translating python scripts, I often encounter the need for libraries such as sklearn, and currently I can only go to the reference py implementation for rewriting, which is really tiring.

[GeorgeS2019]

This issue has been discussed many months before.

The key is to gather community effort and reusability. If someone has ported some python code e.g. from sklearn to .NET, we need to think of reusability on top of existing framework e.g. ML.NET. (this is just my opinion, we need a further discussion on that :-))

We need a concerted community effort to plan how to move forwards so we have .NET substitutions for e.g. sklearn Spacy etc.

The more developers like yourself feedback this, the more the dotnet team will take this into consideration.

Keep providing feedback.

Thanks for keep contributing!

[filipeclduarte]

Which numerical library above mentioned should be used to port the Sklearn library to .NET and maintain interoperability with TorchSharp?

Also, as another route, I think it is possible to reborn Accord.NET, which is now archived. It is simple to Sklearn with the peculiarity of C#.

[GeorgeS2019]

TorchSharp's active community has grown significantly since the same time last year. What seems inconceivable which is daily with python is now within reach using dotnet thanks to Torchsharp.

YES, we need more discussions.

For Accord.NET its license needs to be made available as MIT license. Someone needs to persuade this transition since it is now a public archive.

We need to discuss different strategies for how to do that.

One of the key reasons why TorchSharp is catching growing traction, in my view, was the decision and active commitment to making the portability from PyTorch to dotnet simpler AND keeping Torchsharp up-to-date with abundant and high coverage tests to demonstrate how to do the portability.

We need to learn from this success, to make machine learning for dotnet a more complete experience.

[filipeclduarte]

I understand what you have said and think those advances made really awesome. It is essential to continue discussing and elaborate a plan to build a fully interoperable companion of libraries for data manipulation, numerical computing in CPU (for small matrices and interoperable with TorchSharp), probability, statistics, general machine learning, data visualization, etc.

Right now, is there any .NET numerical computing library that is interoperable with TorchSharp?

[GeorgeS2019]

https://en.m.wikipedia.org/wiki/List_of_numerical_libraries

Can u be more specific?

The discussion need to be e.g. for pytotch, what are the typical options, are these available in .NET? If so, are they interoperable with TorchSharp?

[filipeclduarte]

I mean ... Do any of those libraries mentioned in the first post have integration with TorchSharp? Considering we are talking about the library on the .NET platform. Maybe some method like tensor.toNumpy() conversion?

From what I've seen, NumpyDotNet is the most compatible with the NumPy API.

[GeorgeS2019]

I do not speak for the main TorchSharp developers from Microsoft.

I hope, if there are sufficient users here who have similar wishes for torch.from_numpy and/or PyTorch Tensor to Numpy, Perhaps the wishes will be considered.

[GeorgeS2019]

@filipeclduarte => FYI

[GeorgeS2019]

@kaiidams

In your most recent TorchAudio example using TorchSharp: SPEECH COMMAND CLASSIFICATION WITH TORCHAUDIO

There are a few Pytorch Numpy interoperability statements

waveform.t().numpy()
waveform_first.numpy()

Could you please share your views on how would TorchSharp community approach these interoperability statements? Primarily based on the discussion above and the availability of different Numpy for .NET libraries.

CC @NiklasGustafsson

[GeorgeS2019]

Related discussion

[NiklasGustafsson]

Is there a plan to support Torch and numpy interoperability in .NET?

As far as TorchSharp is concerned, the answer is no. We're trying to keep the scope very narrow and clean, so that we can manage it without too many dependencies and with minimal staff. I also would like to see a de facto standard .NET numerics library, but we will not be using TorchSharp as a forcing function.

[GeorgeS2019]

@NiklasGustafsson

Now that we have tensor.from_array(), what is your recommended way to do tensor.to_array

Typically, in pytorch, this is tensor.numpy()

Any suggestion?

Ideally there is a small documentation to guide how best to do tensor.from_array and tensor.to_array in terms of performance and best align with .NET best practice.

[NiklasGustafsson]

Tensor.data<T> is what you're looking for.

[GeorgeS2019]

[NiklasGustafsson]

Tensor.data is used in 158 unit tests... :-)

[GeorgeS2019]

FYI:

Bridge with NumPy

A short documentation section on .NET arrays and tensors.?

• System.Array to Tensor
• Tensor to System.Array

[NiklasGustafsson]

Let's not describe from_array and data as having anything to do with numpy. They do not -- they bridge between .NET arrays and tensors. As I have pointed out elsewhere, numpy is a Python library that TorchSharp has no plans to integrate with.

[GeorgeS2019]

Agree,!

What u mean by this?

I also would like to see a de facto standard .NET numerics library,

[NiklasGustafsson]

That it would be great if there were something in the .NET world that was like numpy in capabilities and broad acceptance / deployment. I am not aware of any such library existing.