/TACO

Primary LanguagePython

TACO

The source code of our works on few-shot learning:

  • AAAI 2021 paper: Task Cooperation for Semi-Supervised Few-Shot Learning.

Content

  • Personal Homepage
  • Basic Introduction
  • Running Tips
  • Citation

Personal Homepage

Basic Introduction

  • Semi-Supervised Few-Shot Learning (SS-FSL) is proposed in this paper.
  • Inspired by the idea that unlabeled data can be utilized to smooth the model space in traditional semi-supervised learning, we propose TAsk COoperation (TACO) which takes advantage of unsupervised tasks to smooth the meta-model space.

Environment Dependencies

The code files are written in Python, and the utilized deep learning tool is PyTorch.

  • python: 3.7.3
  • numpy: 1.21.5
  • torch: 1.9.0
  • torchvision: 0.10.0
  • pillow: 8.3.1

Datasets

We provide the dataset including:

  • MiniImagenet, which has 60,000 images for few-shot classification. The dataset path should be configured in paths.py.

Running Tips

  • python train_semi_taco.py: running SS-FSL, the hyper-paramters should be configured in the code file.
  • python train_sup_taco.py: the proposed TACO could also be applied to Supervised Few-shot Learning, the hyper-paramters should be configured in the code file.

Citation

  • Han-Jia Ye, Xin-Chun Li , De-Chuan Zhan. Task Cooperation for Semi-Supervised Few-Shot Learning. In: Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI'21), online conference, 2021.
  • [BibTex]