CBM

中文版说明请点击这里

Introduction

This repository contains the code for the paper:
Cooperative Bi-path Metric for Few-shot Learning
Zeyuan Wang, Yifan Zhao, Jia Li, Yonghong Tian
ACM Conference on Multimedia (ACM MM), 2020

Environments

• python = 3.6

• pytorch = 1.6

• scikit-learn = 0.23

• python-lmdb = 0.96

Datasets:

• miniImageNet (in the form of .pickle):

  Link: https://github.com/gidariss/FewShotWithoutForgetting

• tieredImageNet (in the form of .jpg):

  Link: https://github.com/renmengye/few-shot-ssl-public

Steps

1. Set the Paths

• Change the variable dataset_dir in configuration file ./torchFewShot/datasets/miniImageNet_load.py to the correct path to miniImageNet.
• Change the variable dataset_dir in configuration file ./torchFewShot/datasets/tieredImageNet.py to the correct path to tieredImageNet.
• Change the variable file in save_base_proto.py.py to the correct path to the train set file of miniImageNet.

2. Train Models

train baseline++ on miniImageNet for 5-shot

python train.py mini --nExemplars 5

train baseline++ on miniImageNet for 1-shot

python train.py mini --nExemplars 1

train baseline++ on tieredImageNet for 5-shot

python train.py tiered --nExemplars 5

train baseline++ on tieredImageNet for 1-shot

python train.py tiered --nExemplars 1

3. Save Feature Vectors of Base Classes

save feature vectors of base classes of miniImageNet for 5-shot

python save_base_proto.py mini --nExemplars 5

save feature vectors of base classes of miniImageNet for 1-shot

python save_base_proto.py mini --nExemplars 1

4. Test Methods

test baseline++ on miniImageNet for 5-shot

python test.py mini --nExemplars 5

test baseline++ on miniImageNet for 1-shot

python test.py mini --nExemplars 1

test baseline++ on tieredImageNet for 5-shot

python test.py tiered --nExemplars 5

test baseline++ on tieredImageNet for 1-shot

python test.py tiered --nExemplars 1

test CBM on miniImageNet for 5-shot

python test.py CBM_5_shot

test CBM on miniImageNet for 1-shot

python test.py CBM_1_shot

test CBM_LLE on miniImageNet for 5-shot

python test.py CBM_LLE_5_shot

test CBM_LLE on miniImageNet for 1-shot

python test.py CBM_LLE_1_shot

Citation

If you use this code for your research, please cite our paper:

@inproceedings{DBLP:conf/mm/WangZ0020,
  author    = {Zeyuan Wang and
               Yifan Zhao and
               Jia Li and
               Yonghong Tian},
  editor    = {Chang Wen Chen and
               Rita Cucchiara and
               Xian{-}Sheng Hua and
               Guo{-}Jun Qi and
               Elisa Ricci and
               Zhengyou Zhang and
               Roger Zimmermann},
  title     = {Cooperative Bi-path Metric for Few-shot Learning},
  booktitle = {{MM} '20: The 28th {ACM} International Conference on Multimedia, Virtual
               Event / Seattle, WA, USA, October 12-16, 2020},
  pages     = {1524--1532},
  publisher = {{ACM}},
  year      = {2020},
  url       = {https://doi.org/10.1145/3394171.3413946},
  doi       = {10.1145/3394171.3413946},
  timestamp = {Thu, 15 Oct 2020 16:32:08 +0200},
  biburl    = {https://dblp.org/rec/conf/mm/WangZ0020.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgments

This code is based on the implementations of Cross Attention Network for Few-shot Classification.

---

简介

本代码仓库是对以下论文的实现:
Cooperative Bi-path Metric for Few-shot Learning
Zeyuan Wang, Yifan Zhao, Jia Li, Yonghong Tian
ACM Conference on Multimedia (ACM MM), 2020

环境

• python = 3.6

• pytorch = 1.6

• scikit-learn = 0.23

• python-lmdb = 0.96

数据集:

• miniImageNet (以.pickle格式):

  地址: https://github.com/gidariss/FewShotWithoutForgetting

• tieredImageNet (以.jpg格式):

  地址: https://github.com/renmengye/few-shot-ssl-public

流程

1. 设置路径

• 改变文件 ./torchFewShot/datasets/miniImageNet_load.py 中的变量 dataset_dir，指向miniImageNet。
• 改变文件 ./torchFewShot/datasets/tieredImageNet.py 中的变量 dataset_dir，指向tieredImageNet。
• 改变文件 save_base_proto.py.py 中的变量 file，指向miniImageNet的训练集的pickle文件。

2. 训练模型

train baseline++ on miniImageNet for 5-shot

python train.py mini --nExemplars 5

train baseline++ on miniImageNet for 1-shot

python train.py mini --nExemplars 1

train baseline++ on tieredImageNet for 5-shot

python train.py tiered --nExemplars 5

train baseline++ on tieredImageNet for 1-shot

python train.py tiered --nExemplars 1

3. 保存基础类别的特征向量

save feature vectors of base classes of miniImageNet for 5-shot

python save_base_proto.py mini --nExemplars 5

save feature vectors of base classes of miniImageNet for 1-shot

python save_base_proto.py mini --nExemplars 1

4. 测试不同的方法

test baseline++ on miniImageNet for 5-shot

python test.py mini --nExemplars 5

test baseline++ on miniImageNet for 1-shot

python test.py mini --nExemplars 1

test baseline++ on tieredImageNet for 5-shot

python test.py tiered --nExemplars 5

test baseline++ on tieredImageNet for 1-shot

python test.py tiered --nExemplars 1

test CBM on miniImageNet for 5-shot

python test.py CBM_5_shot

test CBM on miniImageNet for 1-shot

python test.py CBM_1_shot

test CBM_LLE on miniImageNet for 5-shot

python test.py CBM_LLE_5_shot

test CBM_LLE on miniImageNet for 1-shot

python test.py CBM_LLE_1_shot

引用

如果你使用了该代码，请以下列各式引用我们的论文：

@inproceedings{DBLP:conf/mm/WangZ0020,
  author    = {Zeyuan Wang and
               Yifan Zhao and
               Jia Li and
               Yonghong Tian},
  editor    = {Chang Wen Chen and
               Rita Cucchiara and
               Xian{-}Sheng Hua and
               Guo{-}Jun Qi and
               Elisa Ricci and
               Zhengyou Zhang and
               Roger Zimmermann},
  title     = {Cooperative Bi-path Metric for Few-shot Learning},
  booktitle = {{MM} '20: The 28th {ACM} International Conference on Multimedia, Virtual
               Event / Seattle, WA, USA, October 12-16, 2020},
  pages     = {1524--1532},
  publisher = {{ACM}},
  year      = {2020},
  url       = {https://doi.org/10.1145/3394171.3413946},
  doi       = {10.1145/3394171.3413946},
  timestamp = {Thu, 15 Oct 2020 16:32:08 +0200},
  biburl    = {https://dblp.org/rec/conf/mm/WangZ0020.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

致谢

该代码主要基于 Cross Attention Network for Few-shot Classification 实现，在此对原作者的工作表示衷心的感谢！