Project Overview

CSSL is a PyTorch Lightning codebase for continual self-supervised learning.

Supported SSL methods

The repository includes model configurations for:

• Barlow Twins

• BYOL

• DINO

• MoCo v2+

• SimSiam

• SwAV

• VICReg

Continual-learning plugins

Available plugin configurations include:

• experience replay

• dark experience replay

Repository layout

Key paths in the repository:

• cssl/: main package

• cssl/dataset/: datasets and continual scenarios

• cssl/framework/: framework-specific modules

• cssl/loss/: loss functions

• cssl/metrics/: evaluation and logging

• cssl/models/: model components

• cssl/plugins/: continual-learning plugins

• cssl/utils/: utilities, factories, callbacks, data manager

• config/model/: model YAML configuration files

• config/plugin/: plugin YAML configuration files

• train.py: main training script

• pretrain.py: alternative pretraining/evaluation flow

• tune.py: Optuna hyperparameter tuning