[Edit] A brief description of what this project does and who it's for.

This Github repository was set up using the continualTemplate repository as a template. It spins up a PyTorch container (see .devcontainer/Dockerfile), with continualTrain mounted. The main idea is to use pluggy hooks to launch training logic that centrally exists in the continualTrain repository.

While a tad complex, setting up the repository like this:

• led to a reproducible project with a couple of simple installations
• cleanly self-documented how the CL models were trained (you just have to check out the training directory!)
• avoided a lot of the redundant code when setting up a new repository.

First, in .devcontainer/.json, set the correct mounts. These mounts will be binded to your devcontainer.

The required packages are defined in training.toml and installed in a virtual environment automatically by the Dockerfile.

Execute ./own.sh to quickly return ownership to your user (instead of root, thanks to Docker)

This script is automatically run by Docker to parse training.toml and install the needed tools into your virtual environment, inside the container.

Training is configured by a .toml file specifiying the model save path, dataset root path, directory of implementations, and WandB settings. .gitignore has already been configured to ignore training.toml because it will contain sensitive information (WandB API key, for example). Since you need continualTrain installed locally, you can use it initialize the file by running barracks initialize.

Please ensure that all paths provided in the training.toml file are valid directories!

Each file in the training_dir is a python file that implements the following hooks:

• get_dataset an Avalanche dataset
• get_model a continualUtils inherited torch model
• get_criterion a PyTorch criterion
• get_strategy an AValanche strategy
• get_metadata Must define strategy_name, dataset_name, model_name, wandb_entity, wandb_project_name for logging purposes

and more!

The way to launch training