Adversarial Continual Learning

This is the official PyTorch implementation of the Adversarial Continual Learning at ECCV 2020. .

Abstract

Continual learning aims to learn new tasks without forgetting previously learned ones. We hypothesize that representations learned to solve each task in a sequence have shared structure while containing some task-specific properties. We show that shared features are significantly less prone to forgetting and propose a novel hybrid continual learning framework that learns a disjoint representation for task-invariant and task-specific features required to solve a sequence of tasks. Our model combines architecture growth to prevent forgetting of task-specific skills and an experience replay approach to preserve shared skills. We demonstrate our hybrid approach is effective in avoiding forgetting and show it is superior to both architecture-based and memory-based approaches on class incrementally learning of a single dataset as well as a sequence of multiple datasets in image classification.

Authors:

Sayna Ebrahimi (UC Berkeley, FAIR), Franziska Meier (FAIR), Roberto Calandra (FAIR), Trevor Darrell (UC Berkeley), Marcus Rohrbach (FAIR)

Citation

If using this code, parts of it, or developments from it, please cite our paper:

@article{ebrahimi2020adversarial,
  title={Adversarial Continual Learning},
  author={Ebrahimi, Sayna and Meier, Franziska and Calandra, Roberto and Darrell, Trevor and Rohrbach, Marcus},
  journal={arXiv preprint arXiv:2003.09553},
  year={2020}
}

Prerequisites:

• Linux-64
• Python 3.6
• PyTorch 1.3.1
• CPU or NVIDIA GPU + CUDA10 CuDNN7.5

Installation

• Create a conda environment and install required packages:

  bash
  conda create -n  python=3.6
  conda activate 
  pip install -r requirements.txt
  

• Clone this repo:

  bash
  mkdir ACL
  cd ACL
  git clone [email protected]:facebookresearch/Adversarial-Continual-Learning.git
  

• The following structure is expected in the main directory:

./src                     : main directory where all scripts are placed in
./data                    : data directory
./src/checkpoints         : results are saved in here

For each datasest run the following commands from

src

directory. Config file for each experiment contains the hyperparameters we used in the paper:

Split MNIST (5 Tasks):

python main.py --config ./configs/config_mnist5.yml

Permuted MNIST (10 Tasks):

python main.py --config ./configs/config_pmnist.yml

Split CIFAR100 (20 Tasks):

python main.py --config ./configs/config_cifar100.yml

Split MiniImageNet (20 Tasks):

python main.py --config ./configs/config_miniimagenet.yml

Sequence of 5 Tasks (CIFAR10, MNIST, notMNIST, Fashion MNIST, SVHN)

python main.py --config ./configs/config_multidatasets.yml

ACL with ResNet18 backbone

See here.

Datasets

miniImageNet data should be downloaded and pickled as a dictionary (

data.pkl

images

labels

ags.data_dir

miniimagenet

data.pkl

data/

notMNIST dataset is included here in

./data/notMNIST

Other datasets will be automatically downloaded and extracted to

./data

Questions/ Bugs

• For questions/bugs, contact the author Sayna Ebrahimi via email [email protected]

License

This source code is released under The MIT License found in the LICENSE file in the root directory of this source tree.

Acknowledgements

Our code structure is inspired by HAT.