pytorch-project-template

A template for building pytorch projects and packages.

pytorch_project_template
├── config.py            - File to define run parameters
├── data/                - Directory for data storage, data won't be tracked
├── dependencies/        - Directory for dependencies and submodules
├── docker-compose.yml   - Docker compose file
├── dockerfile           - Dockerfile with all dependencies
├── .dockerignore
├── environment.yml      - Conda environment file
├── .gitignore
├── __init__.py          - Defines project as a module
├── LICENSE              - MIT license
├── main.py              - Script that defines API routes (if using API)
├── models/              - Directory for pytorch modules
├── README.md
├── setup.py             - PyPI setup script, use for creating pip package
├── tests/
│   └── test.py          - Add some test code here, used by CI testing 
└── utils/                
    ├── datasets.py      - Custom datasets
    ├── samplers.py      - Custom samplers and batch samplers
    └── training.py      - Functions for training and evaluation

Quick setup

Requires git, python, and conda.

1. Clone this project:

   git clone https://github.com/timesler/pytorch-project-template

2. Create a new project (see create_project.py --help):

   create_project.py <new_project_path> <new_project_owner>

   Optionally, the project can also include:

   • git: --git
   • docker and docker-compose to develop models in a container: --docker
   • a dockerfile and docker-compose service to serve models with a REST API: --api
   • a setup.py file for building a pip/PyPI package: --pypi
   • all of the above: -full

   E.g., for building a python package, you probably want:

   create_project.py <new_project_path> <new_project_owner> --git --pypi

3. Create development environment using conda or docker:

   • conda:

     conda create -n pytorch-py37 -f environment.yml
     conda activate pytorch-py37

   • docker/nvidia-docker:

     docker-compose run --rm train bash

4. (Optional) Deploy API for serving model:

   docker-compose up -d api

Use as Github template

Alternatively, this repo can be used as a template for creating new repos on Github. Simply click "Use this template" at the top of the main repo page.

Following this, you should:

• Overwrite this README
• Remove the create_project.py utility
• Remove all unwanted files and folders