Home

Awesome

PL-CFE

Implementation code for

Rethinking Clustering-Based Pseudo-Labeling for Unsupervised Meta-Learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.

By Xingping Dong, Jianbing Shen, and Ling Shao.

Any comments, please email: xingping.dong@gmail.com, shenjianbingcg@gmail.com

If you use this software for academic research, please consider to cite the following paper:

@inproceedings{dong2022rethinking,
  title={Rethinking Clustering-Based Pseudo-Labeling for Unsupervised Meta-Learning},
  author={Dong, Xingping and Shen, Jianbing and Shao, Ling},
  booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XX},
  pages={169--186},
  year={2022},
  organization={Springer}
}

Getting started

Requirements

To avoid any conflict with your existing Python setup, it is suggested to work in a virtual environment with Miniconda.

Create a new environment (PL-CFE) with conda

bash creat_env.sh

Datasets

ImageNet

  1. Download the Imagenet dataset into $imagenet_dir.
  2. Copy extract_ILSVRC.sh to $imagenet_dir
  3. Go to $imagenet_dir and run extract_ILSVRC.sh
cd $imagenet_dir
bash extract_ILSVRC.sh

Omniglot

We have provide the cache files for Omniglot dataset in './data/omniglot' Please extract it on './data' and make sure the structure of files is the same to 'data/omniglot/omniglot_cache_train.npz'

cd data
tar -xf omniglot.tar.xz 
cd ..

miniImageNet

We download miniImageNet (Google Drive) from few-shot-ssl. Then place 'mini-imagenet.tar.gz' into ./data/miniimagenet/

tieredImageNet

You can use 'cfe/download_datasets' to download tieredImageNet.

cd cfe
python download_datasets.py -d tieredimagenet
cd ..

Usage

We provide shell scripts in 'scripts to reproduce our results. Here, we provide the procedure to train a model on omniglot. (Similar precedure for other datasets)

Step 1: Unsupervised training model with our CFE (multiple GPUs)

Notice that this step requires multiple GPUs. If you use one GPU for training, you may suffer from performance degradation

cd scripts
bash train_cfe_omni.sh
# For miniImagenet, please download the ImageNet and then modify the dir_imagenet in train_cfe_mini
# bash train_cfe_imagenet

Step 2: Extract the embeddings for the few shot dataset

After training, you can use this model to extract the embeddings and make clustering by using cfe/cfe_encoding.py, and save the results as new datasets. Please do data transform in this step. In the following steps, we would not do any transform.

We also provide a script as following. Please modify $pretrain_path to select the embedding model. We provide the pre-trained model in Google Drive

bash cfe_encoding.sh

Step 3: Run the supervised method, such as MAML, on the pseudo labeling dataset based on clustering in our CFE embedding.

bash omni-maml-pem.sh

Please refer 'EP/scripts' to evaluate the EP method.

Acknowledgement

Some codes are borrowed from the following open-source codes. Thanks for the contributions of these authors.

pytorch-maml, MoCo, EP