FTorch_coupling_examples

Usage

These examples provide complete demonstrations of how to use FTorch to couple various neural network implementations in PyTorch using Fortran.

• CNN-simple: Train a simple CNN to predict y = x + 1 mapping where x is a 2D tensor of random values between -1 and 1 of size (bs, 1, 20, 20), where domain size is 20 * 20. The example runs inference in Python and Fortran and checks that error is within an acceptable tolerance.

• FNO-1D: Train a simple FNO-1D network to predict the sin wave sin(2πx) given an input of some dummy values and a uniform grid of x-positions on [0, 1], and the target sine wave. The example runs inference in Python and Fortran with the same dummy values and unform grid.

Installation

Install FTorch as described here, then compile the examples by following the README in each directory. For illustration, the cmake command will look something like the following (if using LibTorch - otherwise use PyTorch).

cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH="/path/to/libtorch/share/cmake/Torch/;/path/to/FTorch/lib/cmake/FTorch/"

Acknowledgements

This module includes an implementation of Fourier Neural Operators in FNO-1d/fno-1d.py adapted from Pahlavan et al. (2024), On the importance of learning non-local dynamics for stable data-driven climate modeling: A 1D gravity wave-QBO testbed, as made available in their accompanying repository HamidPahlavan/Nonlocality Portions of the code are used under the MIT License.