STUNT: Few-shot Tabular Learning with Self-generated Tasks from Unlabeled Tables (ICLR 2023 Spotlight)
Official PyTorch implementation of "STUNT: Few-shot Tabular Learning with Self-generated Tasks from Unlabeled Tables" by Jaehyun Nam, Jihoon Tack, Kyungmin Lee, Hankook Lee, Jinwoo Shin.
TL;DR: We propose a few-shot semi-supervised tabular learning framework that meta-learns over the self-generated tasks from unlabeled tables.
conda create -n stunt python=3.8 -y
conda activate stunt
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install torchmeta tensorboardX
conda install -c conda-forge faiss-gpu
Download the datasets and place at /data folder
The options for the training method are as follows:
<MODE>: {protonet}<MODEL>: {mlp}<DATASET>: {income}
cd data/<DATASET>
python generate_pseudo_val.py
python main.py --mode <MODE> --model <MODEL> --dataset <DATASET>
Place the labeled sample index of the test set in data/<DATASET>/index<SHOT> before evaluation.
python eval.py --data_name <DATASET> --shot_num <SHOT> --seed <SEED> --load_path <PATH>
@inproceedings{nam2023stunt,
title={{STUNT}: Few-shot Tabular Learning with Self-generated Tasks from Unlabeled Tables},
author={Jaehyun Nam and Jihoon Tack and Kyungmin Lee and Hankook Lee and Jinwoo Shin},
booktitle={International Conference on Learning Representations},
year={2023}
}