LGL

Lifelong Graph Learning

Chen Wang, Yuheng Qiu, Dasong Gao, Sebastian Scherer. "Lifelong Graph Learning." Conference on Computer Vision and Pattern Recognition (CVPR), 2022.

Note

This repo only contains source code for citation graphs.

Please go to Human Action Recognition and Feature Matching for other experiments.

Dependencies

Python 3
PyTorch v1.5+
DGL v0.4 (Only used for downloading graph datasets)
```
  conda install dgl=0.4
```
Open Graph Benchmark

Training

Citation datasets (Cora, Citeseer, and Pubmed) are automatically downloaded before training.
Default dataset (download) location is '/data/datasets', you may change it via args '--data-root [data_location]'.
```
  python train.py --data-root [data_location] --config config/Regular/FGNRegularCoraPubmed.yaml
```

To save your model during training

  python lifelong_data.py --data-root [data_location] --config config/Regular/FGNRegularCoraPubmed.yaml --save [model_file_location]

To try some baselined model under --model with GCN, GAT, MLP, SAGE and APP

  python lifelong_data.py --data-root [data_location] --config config/Regular/BaselineRegularCoraPubmed.yaml --save [model_file_location] --model [modle_name]

LifeLong Learning

Class-incremental Tasks

  python lifelong.py --data-root [data_location] --config config/ClassIncremental/FGNClassIncrementalCoraPubmed.yaml --save [model_file_location]

Data-incremental Tasks

  python lifelong_data.py --data-root [data_location] --config config/ --config config/DataIncremental/FGNDatalifelongCoraPubmed.yaml --save [model_file_location]

Testing

Download pretrained models. Change the path of the pretrained model under --load, and change the --config.

  python train.py --data-root [data_location] --config config/Regular/FGNRegularOGB.yaml --load pretrained_model/Regular/nonlifelongFGNKTransCat_ogbn-arxiv.pt

Some API for usage

    usage: lifelong.py [-h] [-c CONFIG] [--device DEVICE] [--data-root DATA_ROOT] [--dataset DATASET] [--model MODEL] [--load LOAD]
                    [--save SAVE] [--optm OPTM] [--lr LR] [--batch-size BATCH_SIZE] [--jump JUMP] [--iteration ITERATION]
                    [--memory-size MEMORY_SIZE] [--seed SEED] [-p] [--eval EVAL] [--sample-rate SAMPLE_RATE] [--k K]
                    [--hidden HIDDEN [HIDDEN ...]] [--drop DROP [DROP ...]] [--merge MERGE]

    optional arguments:
    -h, --help            show this help message and exit
    -c CONFIG, --config CONFIG
                            config file path
    --device DEVICE       cuda or cpu
    --data-root DATA_ROOT
                            dataset location
    --dataset DATASET     cora, citeseer, or pubmed
    --model MODEL         LGL or SAGE
    --load LOAD           load pretrained model file
    --save SAVE           model file to save
    --optm OPTM           SGD or Adam
    --lr LR               learning rate
    --batch-size BATCH_SIZE
                            minibatch size
    --jump JUMP           reply samples
    --iteration ITERATION
                            number of training iteration
    --memory-size MEMORY_SIZE
                            number of samples
    --seed SEED           Random seed.
    -p, --plot            increase output verbosity
    --eval EVAL           the path to eval the acc
    --sample-rate SAMPLE_RATE
                            sampling rate for test acc, if ogb datasets please set it to 200
    --k K                 the level of k hop.
    --hidden HIDDEN [HIDDEN ...]
    --drop DROP [DROP ...]
    --merge MERGE         Merge some class if needed.

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}
    }

WhitePhosphorus4 / LGL

LGL