Overhaul pt2ts

[joewallwork]

Closes #368

For a while I've thought that it'd be more appropriate to have pt2ts as a command-line script that reads in a PyTorch model and writes out a TorchScript model. That's done in this PR.

I feel that other "template" code that's currently in there would be better off existing in examples/integration tests. Some of the testing code can be included in the main pt2ts script. Further, I think we should drop the duplicated pt2ts.py files from the examples, move any specific testing part of them (under if __name__ == "__main__:") to the bottom of the corresponding net definition files (e.g., simplenet.py), and make use of the new pt2ts script in each example. (This has the bonus of testing it, which isn't currently done for the template in utils/pt2ts.py.)

Note that I've made a recommendation to switch to a .ts extension for TorchScript files, although this is just personal preference.

Proposed workplan

• Separate TorchScript handling functions into a new Python module called ftorch_utils
• Make use of the Python module in a new pt2ts.py script
• Update utils README
• Set ftorch_utils up as a proper Python module
• Set pt2ts up as a script that gets installed in the Python environment.
• Install ftorch_utils in the CI
• Move the testing parts of the pt2ts.py scripts found in the examples into the corresponding net definition files
  • SimpleNet
  • Batching
  • MultiIO
  • Looping
  • MultiGPU
  • MPI
• Make use of the main pt2ts.py script in the examples
  • SimpleNet
  • Batching
  • MultiIO
  • Looping
  • MultiGPU
  • MPI
• Update example READMEs
  • SimpleNet
  • Batching
  • MultiIO
  • Looping
  • MultiGPU
  • MPI
• Update documentation
• Changelog

I propose to defer the following to follow-up PRs:

• Writing unit tests for the ftorch_utils module for a follow-up PR - see Unittest ftorch_utils Python module #563
• Allowing pt2ts to load pre-trained PyTorch models (needed for ResNet example) - see Allow pre-trained models with pt2ts #565

[jatkinson1000]

It's come up a couple of times, but there are good reasons not to use .ts as it's picked up by many systems as TypeScript.

[joewallwork]

It's come up a couple of times, but there are good reasons not to use .ts as it's picked up by many systems as TypeScript.

Ah okay fair enough. I'll modify it to stick with .pt then.

[joewallwork]

@copilot Fix the issue with the Windows CI.

[Copilot]

@joewallwork I've opened a new pull request, #566, to work on those changes. Once the pull request is ready, I'll request review from you.

[joewallwork]

@jatkinson1000 this is now ready for re-review. I found an unrelated problem with the Windows CI that it wasn't actually running any tests because none of the --test-regex example1 etc. regexes were picking anything up after we changed the naming scheme. That has now been fixed.

[jatkinson1000]

Thanks @joewallwork this is amazing and will make life a lot easier going forwards!

A few small comments, some of which are nits so I think this is really close to merging.

[jatkinson1000]

PyTorch also explicitly required for examples (not just LibTorch)?
As is mpi4py, matplotlib and more (mentioned below)...

Perhaps we don't specify modules here, or just provide "Other Python dependencies are required for examples/tests" and link to the below section?

[joewallwork]

Addressing this in #570 by modernising the pip install approach.

[jatkinson1000]

pip has some odd things to say here: https://pip.pypa.io/en/stable/cli/pip_install/#finding-packages

But happy to merge this if it seems to be working correctly.

There is no priority in the locations that are searched. Rather they are all checked, and the “best” match for the requirements (in terms of version number - see the specification for details) is selected.

[jatkinson1000]

Thanks @joewallwork

LGTM pending merge of #570 and two typos, thanks!