/CVPR2020-Shoestring

Official code for the CVPR 2020 paper "Shoestring: Graph-Based Semi-Supervised Classification with Severely Limited Labeled Data."

Primary LanguagePython

Shoestring

This repo covers the implementation for our paper Shoestring.

Wanyu Lin, Zhaolin Gao, and Baochun Li. "Shoestring: Graph-Based Semi-Supervised Classification with Severely Limited Labeled Data" In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2020.

Table of Contents

Installation

1. Install anaconda (https://www.anaconda.com/download/)

2. Create a new environment and install tensorflow.

Create a new environment with python=3.7.

conda create --name NAME_OF_YOUR_ENVIRONMENT python=3.7.3

Activate environment.

conda activate NAME_OF_YOUR_ENVIRONMENT

If you have a CUDA-enabled GPU (check https://developer.nvidia.com/cuda-gpus for detail), install tensorflow GPU:

conda install -c anaconda tensorflow-gpu=1.13.1

If not, install tensorflow:

conda install -c conda-forge tensorflow=1.13.1

3. Install other packages

conda install -c anaconda networkx=2.3
conda install -c anaconda scikit-learn=0.21.1
conda install -c conda-forge texttable

Run demos

Run code with parameters to reproduce the results in our paper

# Cora
python train.py --pset config_citation.one_label_set --dataset cora --method l1 l2 cos
python train.py --pset config_citation.two_label_set --dataset cora --method l1 l2 cos
python train.py --pset config_citation.five_label_set --dataset cora --method l1 l2 cos
# Citeseer
python train.py --pset config_citation.one_label_set --dataset citeseer --method l1 l2 cos
python train.py --pset config_citation.two_label_set --dataset citeseer --method l1 l2 cos
python train.py --pset config_citation.five_label_set --dataset citeseer --method l1 l2 cos
# Pubmed
python train.py --pset config_citation.one_label_set --dataset pubmed --method l1 l2 cos
python train.py --pset config_citation.two_label_set --dataset pubmed --method l1 l2 cos
python train.py --pset config_citation.five_label_set --dataset pubmed --method l1 l2 cos
# Large Cora
python train.py --pset config_citation.one_label_set --dataset large_cora --method l1 l2 cos
python train.py --pset config_citation.two_label_set --dataset large_cora --method l1 l2 cos
python train.py --pset config_citation.five_label_set --dataset large_cora --method l1 l2 cos

Parameters

  • k Select the top k probs for each class as the unlabeled data. Between 1 and 0. Default is 0.
  • lam Weight for similarity calculated using distance. Default is 0.01.
  • pset Train size and parameters. Options are: config_citation.one_label_set, config_citation.two_label_set, config_citation.five_label_set. Default is config_citation.one_label_set.
  • dataset Dataset to train. Options are: cora, large_cora, citeseer, pubmed. Default is cora.
  • method Method to calculate the distance. Options are l1, l2, cos. Default is cos.

Cite

Please cite our paper if you use this code in your own work:

@inproceedings{Lin_2020_CVPR,
	author = {Lin, Wanyu and Gao, Zhaolin and Li, Baochun},
	title = {Shoestring: Graph-Based Semi-Supervised Classification With Severely Limited Labeled Data},
	booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
	month = {June},
	year = {2020}
}

Acknowledgments

Thanks for Kipf's implementation of GCN and Li's implementation of GLP and IGCN, on which this repository is initially based.