continualTrain
Github Repositories
Overview
continualTrain allows you to launch and track containerized continual learning runs with minimal code overhead. Built on Pytorch, Avalanche, Pluggy, Weights & Biases, Poetry, and Docker/Singularity, it’s goal is to make your research life easier.
Templating continualTemplate helps organize a repository that works with continualTrain. While optional, it makes it a bit more convenient to get started.
Command Line Interface (CLI) tool
Once you install continualTrain, you will have access to barracks on the command line. barracks is the built-in CLI, which sets up your repository and launches containerized training.
barracks interfaceWhat’s up with continualUtils?
Bring your own batteries
continualTrain is infrastructure. So, you still need to bring your own models, strategies, optimizers, etc. Alternatively, if you find what you need in continualUtils or avalanche-lib, pick it up from there!
Why bother?
My research focuses on continual learning approaches. Developing a tool that quickly launches Avalanche training runs on any machine helps improve my productivity and research quality. It allows me to study the differences between continual learning approaches with the tweaks I want!
More importantly, continual learning research needs baselines. Comparing different approaches is difficult without standardized benchmarks. Our community is still in its early stages, making this an ideal time to adopt a consistent method for evaluating solutions. continualTrain is a step in that direction.
Getting started
continualTrain is a Python package you install, which gives you access to continualUtils and barracks.
Installation
Optional: Follow these instructions provided by Github to use the continualTemplate repository.
Then, install the continualTrain package from PyPi. We recommend doing it inside a virtual environment, but that is optional.
pip install continualtrainUsage
Once continualTrain is installed, you have access to barracks on the command line and continualUtils as a Python package import.
Citation
@online{aswani,
author = {Aswani, Nishant},
title = {continualTrain},
url = {https://nishantaswani.com/projects/barracks.html},
langid = {en}
}