FTA is a Fortran-Torch-Adapter aimed for integrating deep learning model into Fortran environment. It has been used in replacing parameterization schemes in WRF with a published paper

! fortran program main
! 
! use PyTorch modle resnet.pt in a Fortran program with FTA library torch_wrapper
!

program  main
    use torch_wrapper
    implicit none

    CHARACTER(100), TARGET :: model_loc;
    TYPE(ftorchmodel) :: model;
    INTEGER :: res
    REAL(C_float) :: input(1, 3, 224, 224) = 1.0
    REAL :: output(1, 1000)
    INTEGER(C_INT) :: use_gpu = 1                  ! should be compatible with your script module device


    model_loc = "/home/dl/luc/FTB/demo/lib/resnet.pt"//CHAR(0)  ! CHAR(0) is necessary for C string termination 

    print *, "Torch Start"

    model = resnet18_new(model_loc, use_gpu)        ! initialize the model
    res = resnet18_forward(model, input, output)    ! use the model to perform reasoning task
    call resnet18_delete(model)                     ! delete this model

    print *, output(1, 1)
    print *, "Torch End"

end program  main

Dependencies

• CMake (>=3.0)
• gcc or icc compiler
• libtorch
• Docker & nvidia-container-toolkit (OPTIONAL)

How to Use

Notice: we assume you already have a machine with Nvidia Driver installed properly.

Manually setup

1. download libtorch
   1. match your local CUDA version
   2. C++11 ABI (choose this version if your compiler is icc)
2. define your model in the configuration file
   • data type supported: float, int, double
   • data size format: a, b, c
   • model name: will be used for interface name
3. run ./build.sh
4. use the torch_wrapper library in your Fortran program

Setup with Docker

1. install nvidia-container-toolkit & Docker
2. get image
   1. build locally docker build -t fortran-torch-adapter .
   2. Or you can pull from dockerhub docker pull 1813927768/fortran-torch-adapter:latest
3. start the container docker run -it --rm --gpus all fortran-torch-adapter /bin/bash
4. run ./build.sh in the container