Two Complementary Perspectives to Continual Learning

This repository contains the codebase that was used to produce the results for the paper: "Two Complementary Perspectives to Continual Learning:Ask Not Only What to Optimize, But Also How"

Setup

This code uses:

• Python 3.8
• Pytorch 2.1
• Avalanche 0.4.0a (The library code is included in this repository, https://github.com/ContinualAI/avalanche)

To setup your python environmet we provide a conda_environment.yml to mirrors our Anaconda packages. Please use the following bash command:

conda env create -n TwoCompCLEnv --file conda_environment.yml
conda activate TwoCompCLEnv

Reproducing results

To reporduce results of individual experiments, run the python code as below.

1. Select the benchmark configuration from reproduce/benchmark/. Optionally you can provide the --dset_rootpath argument to specify the locaiton of dataset. Otherwise the dataset are stored in ./data.
2. Select a strategy configuration from reproduce/strategy. Set the learning rate {lr} and for strategies involving experience replay, set the total size of the auxilliary buffer {mem_size}. For the online learning experiments also a batch_size {bs} needs to be specified.
3. Select a seed and a location to store the results {save_path}.

Generally:

python train.py --benchmark_config {path_to_config} --strategy_config {path_to_config} --mem_size {total_buffer_size} --lr {0.1, 0.01, 0.001} --bs {10, 64, 128} --seed {142, 152, 162, 172, 182} --save_path {path_to_storage}

• mem_size
  • Rotated MNIST - Offline: 3000 | Online: 60000
  • CIFAR-100 / MiniImageNet - Offline: 10000 | Online: 50000

Example for offline RotatedMNIST:

python train.py --benchmark_config ./reproduce/benchmark/rotmnist_offline.yml --strategy_config ./reproduce/strategy/er_gem.yml --mem_size 3000 --lr 0.1 --bs 128 --seed 142 --save_path ./results/

Note:

We didn't run with deterministic CUDNN backbone for computational efficiency, which might result in small deviations in results. We average all results over 5 initialization seeds {142, 152, 162, 172, 182}.

Access Results

The results are logged with Tensorboard. To view the results use:

tensorboard --logdir={path_to_storage}

Credit

We particularly thank the repositories containing the Avalanche continual learning library (https://github.com/ContinualAI/avalanche), and the codebase for continual evaluation (https://github.com/Mattdl/ContinualPrototypeEvolution)

License

Code is available under MIT license: A short and simple permissive license with conditions only requiring preservation of copyright and license notices.