LyCIL

A clean, PyTorch Lightning (2.x) continual/incremental learning (CIL) library inspired by PyCIL, with up-to-date methods, NPU support, modern interface, and ready-to-run CLI examples.

Features

• Modular LightningModule implementations for iCaRL and LUCIR that inherit from a common BaseIncremental.
• ExemplarBuffer with herding selection, per-class quotas, and persistence.
• CosineClassifier with dynamic expansion (used by LUCIR), and a standard Linear head (used by iCaRL by default; switchable).
• CIFAR-100 and ImageNet LightningDataModules with task-by-task class splits.
• Custom LightningCLI subclass that orchestrates multi-task training loops.
• Concise, English docstrings and comments.

Quick Start

We highly recommend using uv to manage the environment.

# uv venv                           # create virtual env
uv sync --extra cuda                # for cuda lightning
uv sync --extra npu                 # for npu lightning

Alternatively, install via pip or conda:

# conda or pip, with optional dependencies for lightning
pip install -e ".[cuda]"            # 'cuda', or 'npu'

Use LightningCLI commands to train, validate, test, or predict with continual learning models. For example, to train learning without forgetting (LwF) on CIFAR-100:

# valid commands: {fit,validate,test,predict}
python examples/lwf-cli.py fit -c configs/smoketest_lwf.yml
# or pass individual config files
python examples/lwf-cli.py fit --trainer configs/trainer/smoketest.yml --model configs/model/lwf.yml --data configs/data/cifar100.yml