gmumcl

A small, generic continual-learning framework for PyTorch: five incremental scenarios, a five-call plugin API, an accuracy-matrix trainer, and reference implementations of the standard CL baselines.

Install

activationscope builds a C++ extension with -std=c++17, so install torch 2.8 first (newer torch headers require C++20), then build it against that torch:

pip install "torch==2.8.*" --index-url https://download.pytorch.org/whl/cpu   # or /cu128
pip install setuptools wheel ninja numpy
pip install activationscope --no-build-isolation
pip install gmumcl              # or: pip install "gmumcl[datasets]"

From source: git clone https://github.com/gmum/gmumcl && cd gmumcl && pip install -e ".[test]". Full notes (why the pin, why --no-build-isolation): docs/usage.md.

Quickstart

python scripts/run.py --config configs/base.yaml -O method.name=ewc -O method.lambda=10

from gmumcl import train

metrics = train.run({
    "data": {"name": "toy", "num_classes": 4, "classes_per_task": 2},
    "model": {"name": "mlp", "hidden": 16},
    "trainer": {"epochs": 1, "lr": 0.01, "device": "cpu"},
    "method": {"name": "ewc", "lambda": 1.0},
})
print(metrics["accuracy_matrix"])

Scenarios (data.scenario)

Dataset names: toy (CPU clusters), mnist, cifar10, cifar100, plus rotated_mnist (domain-incremental) and permuted_mnist.

Methods (method.name)

Docs

• docs/framework.md — plugin lifecycle, trainer, metrics, adding a method
• docs/scenarios.md — the five scenarios and their protocols
• docs/methods.md — formulas, defaults and provenance per method
• docs/usage.md — install, CLI, configs, custom components
• docs/results.md — reproduced Split-MNIST table (from AdaReg) and deviations

License

MIT (GMUM). See LICENSE.