---

PyTorch Project Template

This repository offers a structured framework for PyTorch projects. It was highly influenced by Christof Henkel's public repositories.

The template contains example model, dataset etc. for the TLVMC-Freezing-of-Gait-Competition and can be run with the competition-data in place.

It is best used in combination with my automated Kaggle container setup and the included local MLflow instance.

Automated logging of models to the MLflow model registry is still a work in progress. By now, only state_dicts are saved as artifacts to MLflow. Some other details are also still being worked on to create the best training-experience.

Necessary Prerequisites

• You know how to use a Makefile
• Screen to keep your training running even when you disconnect from your SSH session.

Repository Structure

File/Folder Name	Description
configs	Configuration files for the training reside here
src	Contains all the python code from training, model, dataset to post-processing
main.py	Runs the actual training program
Makefile	Contains all commands to call the main.py file in different ways needed
preprocessing.ipynb	Jupyter notebook for preprocessing data to the right format needed here

Configuration / How to get started

Individual files added to configs can be used for training a model. All you have to do is change the paths and paramenters and add the file name to the top of the Makefile or your call to the Makefile.

The given config file will then be used when a training run is started.

Generally speaking, all you have to do is preprocess your data so that it can be loaded by a dataset that you provide, and pass it to a suitable model (take a look in into the examples to understand the specifics in this setup). Besides that, you can now automate the whole training process with the config.py file used.

Data should generally reside in a folder called data. When using my automated container setup, the raw competition data is downloaded and mounted under data automatically.

To start a training run, use the commands from the Makefile explained below.

Using the Makefile

1. Run training on all folds consecutively
   Command:

   make all_folds CONFIG=<your_config>

   This command trains the model on all data folds one by one, based on the specified configuration.

2. Run training on the first fold only
   Command:

   make one_fold CONFIG=<your_config>

   This command trains the model on the first fold using the specified configuration.

3. Run training on all folds combined, with validation on fold 0
   Command:

   make all_data CONFIG=<your_config>

   This command trains the model on the full dataset and validates using fold 0.

Notes

• Replace <your_config> with the name of the configuration file (without the .py extension) located in the configs directory. For example, if your configuration file is configs/config.py, use CONFIG=config.
• The Makefile dynamically determines whether to run distributed training or single-GPU training based on the DISTRIBUTED option in your configuration file.

This manual ensures a simplified approach to running various training scenarios with minimal setup effort.