A tool to centralize the execution of continual learning models in Avalanche, from anywhere, for any project, given you follow the interfacing instructions.
This repository doesn't claim to do too much aside from run the main training loop and log things. As such, it contains a training loop skeleton + a logging module.
It also includes continualUtils, a collection of common utilities and tools that can plug in with the main training loop.
Bring Your Own Batteries (BYOB): continualTrain is an interface that runs the main training loop inside an already defined docker container, along with a comprehensive logging module to WandB. You bring your own model and strategies. Or, pick them up from continualUtils and modify them.
Everything is a plugin. Some are required (e.g. model, strategy, dataset), and others are not. If you do not give continualTrain a plugin, it will use a default plugin.
Set up a .env file to do WandB logging
# This is secret and shouldn't be checked into version control
WANDB_API_KEY=$SECRETY_KEY
WANDB_DISABLE_GIT=true #annoying when false