Official implementation for NeurIPS 2020 paper Co-Tuning for Transfer Learning.

[News] 2022/04/27 Please refer to the Transfer Learning Library for a modular implementation.

[News] 2021/01/13 The COCO 70 dataset used in the paper is available for download!

COCO 70 dataset is a large-scale classification dataset (1000 images per class) created from COCO. It is used to explore the effect of fine-tuning with a large amount of data. Check our paper if you are interested in how it was created. Please respect the original license of COCO when you use it.

To download COCO 70, follow these steps:

1. download separate files here (the file is too large to upload, so I have to split it into chunks)

2. merge separate files into a single file by cat COCO70_splita* > COCO70.tar

3. extract the dataset from the file by tar -xf COCO70.tar

The directory architecture looks like the following:

├── classes.txt #per class name per name

├── dev

├── dev.txt # [filename, class_index] per line, 0 <= class_index <= 69

├── test

├── test.txt

├── train

└── train.txt

There are 100 images per class for validation (dev.txt) and test (test.txt) respectively, and 800 images per class for training (train.txt).

• python3
• torch == 1.1.0 (with suitable CUDA and CuDNN version)
• torchvision == 0.3.0
• scikit-learn
• numpy
• argparse
• tqdm

python --gpu [gpu_num] --data_path /path/to/dataset --class_num [class_num] --trade_off 2.3 train.py 

If you use our code or use the constructed COCO-70 dataset, please consider citing:

@article{you2020co,
  title={Co-Tuning for Transfer Learning},
  author={You, Kaichao and Kou, Zhi and Long, Mingsheng and Wang, Jianmin},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

If you have any problem about our code, feel free to contact ykc20@mails.tsinghua.edu.cn or kz19@mails.tsinghua.edu.cn.