Training a model in fortran

[NicoClinco]

Hello,
I was wondering if it is possible to enable the training of a model directly in fortran when imported from a torch-script.
Have you any plan in order to do so?

In particular:

• I want to use my optimizer in fortran, and in order to do so I would like to have a callback in fortran for getting the
  parameters of the model, then change them accordingly to my optimizer.

Any idea, or also a suggestion is highly appreciated.

Thank you,
Nicola

[jwallwork23]

Hi Nicola, thanks for your interest in FTorch. We have a plan to implement training in Fortran and the work in progress is documented here: https://cambridge-iccs.github.io/FTorch/page/online.html

We will hopefully soon merge #320, which will expose Optimizers in FTorch. So hopefully the functionality you seek will be available soon.

[jatkinson1000]

Hi @NicoClinco,

Thanks for asking, this is the biggest feature we are working on at the moment, and I'm tagging @jwallwork23 who has put a lot of work behind this.

We have exposed the autograd function to torch_tensors in Fortran and you can see an example of this in our Example 8.

We are actively working on bringing the optimisers over in the pull request #320 and this is almost ready to be merged.

The final step would be to enable optimisation over parameters of a net.
If this is something you are keen to work on we'd be more than happy to collaborate and develop in tandem with you.

You may also want to join our announcements mailing list where we will post updates and new features as they come in.

[NicoClinco]

Thank you very much. I will be glad to contribute, the problem is that I am not so familiar with Torch in C++. I am trying to creating a binding in C++ such that with a torch::jit::Model I extract the parameters, but I cannot find a good documentation for doing that.

The idea is to create a simple callback in C (...and the binding in fortran similar to your interface) such that we extract the parameters from a torch::jit::Model and use this parameters with an external optimizer.
Any suggestion is welcome
Thank you,
Nicola