Continual Meta Learning for AAAI-2020

A PyTorch implementation or our paper "Learning from the Past: Continual Meta-Learning with Bayesian Graph Neural Networks"

Updated: April 22, 2020

Log

• Update README
• Fix the query shuffle

Requirements

• Python 3.6
• Pytorch 1.1
• TensorboardX

Datasets

The links of datasets will be released afterwards,

• MiniImageNet (1.1 GB) Link
• TieredImageNet (12.9 GB) Link

The data folder should be organized as,

/data
/data/mini-imagenet
/data/tiered-imagenet

• Download 'mini_imagenet_train/val/test.pickle', and put them in the path 'tt.arg.dataset_root/mini-imagenet/compacted_dataset/'

• After data preparation, please change the arg.dataset_root in train.py

Training

The general command for training is,

python3 train.py

Change arguments for different experiments:

• dataset: "mini" / "tiered"
• um_unlabeled: for semi-supervised learning
• meta_batch_size: mini_batch size
• num_layers: GNN's depth
• num_cell: number of hidden states
• num_ways: N-way
• num_shots: K-shot
• seed: we select 111, 222, 333 for reproducibility

Remember to change dataset_root to suit your own case

The training loss and validation accuracy will be automatically saved in './asset/logs/', which can be visualized with tensorboard. The model weights will be saved in './asset/checkpoints'

Evaluation

For testing the trained model, you can use the command as

python3 eval.py -test_model "THE_MODEL_NAME"

Citation

Please cite the following paper in your publications if it helps your research

@inproceedings{DBLP:conf/aaai/LuoHZWBY20,
  author    = {Yadan Luo and
               Zi Huang and
               Zheng Zhang and
               Ziwei Wang and
               Mahsa Baktashmotlagh and
               Yang Yang},
  title     = {Learning from the Past: Continual Meta-Learning with Bayesian Graph
               Neural Networks},
  booktitle = {The Thirty-Fourth {AAAI} Conference on Artificial Intelligence, {AAAI}
               2020, The Thirty-Second Innovative Applications of Artificial Intelligence
               Conference, {IAAI} 2020, The Tenth {AAAI} Symposium on Educational
               Advances in Artificial Intelligence, {EAAI} 2020, New York, NY, USA,
               February 7-12, 2020},
  pages     = {5021--5028},
  year      = {2020},
  crossref  = {DBLP:conf/aaai/2020},
  url       = {https://aaai.org/ojs/index.php/AAAI/article/view/5942},
  timestamp = {Thu, 04 Jun 2020 16:49:55 +0200},
  biburl    = {https://dblp.org/rec/conf/aaai/LuoHZWBY20.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgement

Part of code is built on https://github.com/renmengye/few-shot-ssl-public and https://github.com/khy0809/fewshot-egnn

Contact

To report issues for this code, please open an issue on the issue tracker. If you have any further questions, please contact me via lyadanluol@gmail.com