Offline training

The established approach for designing and training a model for running inference in FTorch is the so-called 'offline' approach, which is described in the following. We are currently working on implementing an alternative online training approach - see the online training user guide page for details.

Workflow

Below we provide a schematic of the offline training workflow, which is broken down into separate tasks below.

1. Design the ML model in PyTorch

This task is done purely in Python and is not described here. See the PyTorch documentation for information on how to do this.

2. Data generation

This task is done purely in Fortran. Run the Fortran model to generate training data. Depending on how you write the Fortran model output to disk, there may be some work required to get this training data in an appropriate format for later reading into Python.

3. Training

This tasks is done purely in Python. Create a Python script that loads the training data that was saved to disk, splits it into training and validation sets, runs an optimizer, and then writes the trained model to a file with .pt extension. For more information on how to do this, see the PyTorch documentation.

4. pt2ts

Having written a trained model to a file with .pt extension, use the pt2ts.py utility Python script to convert it to TorchScript format. A template pt2ts.py script can be found in the utils subdirectory. See the README there for more details on how to use the script.

5. Hybrid model

In order to run inference with the trained ML model, you will need to modify your Fortran model so that it uses FTorch syntax to set up appropriate torch_tensor and torch_model objects and call the torch_tensor_forward subroutine to run the inference. For examples of how to do this, see the examples user guide page.