[Update on 31/03/2020] This repository has been merged to Xfer.

[ICLR 2020] Synthetic information bottleneck for transductive meta-learning

This repo contains the implementation of the synthetic information bottleneck algorithm for few-shot classification on Mini-ImageNet, which is used in our ICLR 2020 paper Empirical Bayes Transductive Meta-Learning with Synthetic Gradients.

If our code is helpful for your research, please consider citing:

@inproceedings{
    Hu2020Empirical,
    title={Empirical Bayes Transductive Meta-Learning with Synthetic Gradients},
    author={Shell Xu Hu and Pablo Moreno and Yang Xiao and Xi Shen and Guillaume Obozinski and Neil Lawrence and Andreas Damianou},
    booktitle={International Conference on Learning Representations (ICLR)},
    year={2020},
    url={https://openreview.net/forum?id=Hkg-xgrYvH}
}

Authors of the code

Shell Xu Hu, Xi Shen and Yang Xiao

Dependencies

The code is tested under Pytorch > 1.0 + Python 3.6 environment with extra packages:

pip install -r requirements.txt

How to use the code on Mini-ImageNet?

Step 0: Download Mini-ImageNet dataset

cd data
bash download_miniimagenet.sh 
cd ..

Step 1 (optional): train a WRN-28-10 feature network (aka backbone)

The weights of the feature network is downloaded in step 0, but you may also train from scracth by running

python main_feat.py --outDir miniImageNet_WRN_60Epoch --cuda --dataset miniImageNet --nbEpoch 60

Step 2: Meta-training on Mini-ImageNet, e.g., 5-way-1-shot:

python main.py --config config/miniImageNet_1shot.yaml --seed 100 --gpu 0

Step 3: Meta-testing on Mini-ImageNet with a checkpoint:

python main.py --config config/miniImageNet_1shot.yaml --seed 100 --gpu 0 --ckpt cache/miniImageNet_1shot_K3_seed100/outputs_xx.xxx/netSIBBestxx.xxx.pth

Mini-ImageNet Results (LAST ckpt)

CIFAR-FS Results (LAST ckpt)