Awesome

Lifelong Graph Learning

This repo is for the application in paper "Lifelong Graph Learning", CVPR, 2022.

Temporal and distributed pattern recognition using the Wearable Action Recognition Dataset (WARD).

Training and Testing

Note that MLP, AFGN and GAT perform the best with Adam, while the others perform the best with SGD.

For feature graph network (FGN):

python regular.py --model FGN --optim SGD
python lifelong.py --model FGN --optim SGD

For attention feature graph network (AFGN):

python regular.py --model AFGN --optim Adam
python lifelong.py --model AFGN --optim Adam

For multi-layer perceptron (MLP):

python regular.py --model MLP --optim Adam
python lifelong.py --model MLP --optim Adam

For graph attention network (GAT):

python regular.py --model GAT --optim Adam
python lifelong.py --model GAT --optim Adam

For grach convolutional network (GCN):

python regular.py --model GCN --optim SGD
python lifelong.py --model GCN --optim SGD

For approximated personalized propagation of neural predictions (APPNP):

python regular.py --model APPNP --optim SGD
python lifelong.py --model APPNP --optim SGD

You can also specify the dataset location to be downloaded (Default: /data/datasets). For example:

python regular.py --data-root ./ --model FGN --optim SGD

Reproduce results in the paper

Download pre-trained models (v2.0) and extract. Then run:

python evaluation.py --load saves/lifelong-fgn-s0.model
python evaluation.py --load saves/lifelong-afgn-s0.model
python evaluation.py --load saves/lifelong-appnp-s0.model
python evaluation.py --load saves/lifelong-gcn-s0.model
python evaluation.py --load saves/lifelong-gat-s0.model

We provide all snapshot models during training, which is named as "[task]-[model]-s[seed]-it[iteration].model".

For example, "lifelong-fgn-s0-it3000.model"

Citation

@inproceedings{wang2022lifelong,
  title={Lifelong graph learning},
  author={Wang, Chen and Qiu, Yuheng and Gao, Dasong and Scherer, Sebastian},
  booktitle={2022 Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2022}
}