This repository contains a template base code for a complete pytorch pipeline.
This is a template because it works on fake data but aims to illustrate some pythonic syntax allowing the pipeline to be modular.
More specifically, this template base code aims to target :
- modularity : allowing to change models/optimizers/ .. and their hyperparameters easily
- reproducibility : saving the commit id, ensuring every run saves its assets in a different directory, recording a summary card for every experiment, building a virtual environnement
For the last point, if you ever got a good model as an orphane pytorch tensor whithout being able to remember in which conditions, with which parameters and so on you got, you see what I mean.
For a local experimentation, you start by setting up the environment :
python3 -m virtualenv venv
source venv/bin/activate
python -m pip install .
Then you can run a training, by editing the yaml file, then
python -m torchtmpl.main config.yml train
And for testing (not yet implemented)
python main.py path/to/your/run test
For running the code on a cluster, we provide an example script for starting an experimentation on a SLURM based cluster.
The script we provide is dedicated to a use on our clusters and you may need to adapt it to your setting.
Then running the simulation can be as simple as :
python3 submit.py
Every module/script is equiped with some test functions. Although these are not unitary tests per se, they nonetheless illustrate how to test the provided functions.
For example, you can call :
python3 -m virtualenv venv
source venv/bin/activate
python -m pip install .
python -m torchtmpl.models
and this will call the test functions in the torchtmpl/models/__main__.py script.