This is the Pytorch implementation for the paper "Meta-GMVAE: Mixture of Gaussian VAE for Unsupervised Meta-Learning", in ICLR 2021.

Contribution of this work

• We propose a novel unsupervised meta-learning model, namely Meta-GMVAE, which meta-learns the set-conditioned prior and posterior network for a VAE. Our Meta-GMVAE is a principled unsupervised meta-learning method, unlike existing methods on unsupervised meta-learning that combines heuristic pseudo-labeling with supervised meta-learning.
• We propose to learn the multi-modal structure of a given dataset with the Gaussian mixture prior, such that it can adapt to a novel dataset via the EM algorithm. This flexible adaptation to a new task, is not possible with existing methods that propose VAEs with Gaussian mixture priors for single task learning.
• We show that Meta-GMVAE largely outperforms relevant unsupervised meta-learning baselines on two benchmark datasets, while obtaining even better performance than a supervised meta-learning model under a specific setting. We further show that Meta-GMVAE can generalize to classification tasks with different number of ways (classes).

This code is written in Python. Dependencies include

• python >= 3.6
• pytorch = 1.4 or 1.7
• tqdm

• Download Omniglot data from here.
• Download pretrained features for Mini-ImageNet from here.
• (Optional) If you want to train SimCLR from scratch, download images for ImageNet from here.

data directory should be look like this:

data/

├── omiglot/
  ├── train.npy
  ├── val.npy
  └── test.npy
  
├── mimgnet/
  ├── train_features.npy
  ├── val_features.npy
  └── test_features.npy
  
└── imgnet/ -> (optional) if you want to train SimCLR from scratch
  ├── images/
    ├── n0210891500001298.jpg  
    ├── n0287152500001298.jpg 
	       ...
    └── n0236282200001298.jpg 
  ├── train.csv
  ├── val.csv
  └── test.csv

To reproduce Omniglot 5-way experiment for Meta-GMVAE, run the following code:

cd omniglot
python main.py --data-dir DATA DIRECTORY (e.g. /home/dongbok/data/omniglot/) --save-dir SAVE DIRECTORY (e.g. /home/dongbok/omniglot-5way-experiment) --way 5 --sample-size 200

To reproduce Omniglot 20-way experiment for Meta-GMVAE, run the following code:

cd omniglot
python main.py --data-dir DATA DIRECTORY (e.g. /home/dongbok/data/omniglot/) --save-dir SAVE DIRECTORY (e.g. /home/dongbok/omniglot-20way-experiment) --way 20 --sample-size 300

To reproduce Mini-ImageNet 5-way experiment for Meta-GMVAE, run the following code:

cd mimgnet
python main.py --data-dir DATA DIRECTORY (e.g. /home/dongbok/data/mimgnet/) --save-dir SAVE DIRECTORY (e.g. /home/dongbok/mimgnet-5way-experiment)

(Optional) To reproduce SimCLR features for Mini-ImageNet, run the following code:

cd simclr
python main.py --data-dir DATA DIRECTORY (e.g. /home/dongbok/data/imgnet/) --save-dir SAVE DIRECTORY (e.g. /home/dongbok/simclr-experiment) --feature-save-dir FEATURE SAVE DIRECTORY (e.g. /home/dongbok/data/mimgnet)

To cite the paper, please use this BibTex

@inproceedings{lee2021metagmvae,
  title={Meta-GMVAE: Mixture of Gaussians VAEs for Unsupervised Meta-Learning},
  author={Dong Bok Lee and Dongchan Min and Seanie Lee and Sung Ju Hwang},
  booktitle={Proceedings of the 9th International Conference on Learning Representations},
  year={2021}
}