cambridge-iccs / fortran-ml-bridge-hackathon

Code and examples on directly calling Tensorflow / Pytorch / Keras ML models from Fortran.

fortran-ml-bridge

Code and examples on directly calling Tensorflow / Pytorch ML models from Fortran.

Problem Statement

We want to be able to run ML models directly in Fortran. Initially let's assume that the model has been trained in some other language (say Python) and saved (e.g. as a SavedModel). We want to run inference on this model without having to call the Python executable. This should be possible by using the existing ML C/C++ interfaces.

PyTorch

See the fortran-pytorch-lib directory.

Tensorflow

See the fortran-tf-lib directory.

fortran-ml-bridge-hackathon's Issues

Test for process_model output generating expected Fortran

In a sense writing this test is completing the project, as 'generating the expected Fortran' is the point. However, this test should check for different options for process_model and unexpected model types, etc.

Break down design of the Fortran template

Need:

• init (mostly written)
• finish (mostly written)
• calc
• load tensors

Need test that generated Fortran object can be linked with the TF-fortran library

This is a bigger issue that might be addressed in steps. The generated code needs to be able to allow a simple program to load and infer from a model.

Need test to check process_model output compiles

Since we are generating Fortran code we should have an easy way to check the validity (from a compile point of view) of the generated code.

Port to Julia

This project definitely needs more Julia