Repository for the paper : ME-D2N: Multi-Expert Domain Decompositional Network for Cross-Domain Few-Shot Learning （ACM MM 2022）
If you have any questions/advices/potential ideas, welcome to contact me by fuyq20@fudan.edu.cn.
A anaconda envs is recommended:
conda create --name py36 python=3.6
conda activate py36
conda install pytorch torchvision -c pytorch
pip3 install scipy>=1.3.2
pip3 install tensorboardX>=1.4
pip3 install h5py>=2.9.0
We evaluate our methods on five datasets: mini-Imagenet works as source dataset, cub, cars, places, and plantae serve as the target datasets, respectively.
- The datasets can be conviently downloaded and processed as in FWT.
- Remember to modify your own dataset dir in the 'options.py'.
- We follow the same the same auxiliary target images as in our previous work meta-FDMixup, and the used jsons have been provided in the output dir of this repo.
If you can't find the Plantae dataset, we provide it here, please cite its paper.
As in most of the previous CD-FSL methods, a pretrained feature extractor baseline is used.
- you can directly download it from this link, rename it as 399.tar, and put it to the
./output/checkpoints/baseline - or you can pretrain it as follows:
python3 train_metaTeacher.py --modelType pretrain --dataset miniImagenet --name baseline --train_aug
Our method is target set specific, and we take the cub target set under the 5-way 1-shot setting as an example.
- Training St-Net
python3 train_metaTeacher.py --modelType St-Net --dataset miniImagenet --name St-Net-1shot --train_aug --warmup baseline --n_shot 1
- Training Tt-Net
python3 train_metaTeacher.py --modelType Tt-Net --dataset cub --name Tt-Net-target-set-cub-1shot --train_aug --warmup baseline --n_shot 1 --stop_epoch 100
- note: as stated in paper, only Tt-Net under 1-shot setting is trained 100 epochs. In other cases, 400 epochs are adopted.
- Training the ME-D2N student model
python3 train_metaStudent.py --modelType Student --target_set cub --name ME-D2N-target-set-cub-1shot --train_aug --warmup baseline --n_shot 1 --ckp_S output/checkpoints/St-Net-1shot/399.tar --ckp_A output/checkpoints/Tt-Net-target-set-cub-1shot/99.tar
- testing for St-Net/Tt-Net
python test.py --name St-Net-1shot --dataset DATASET --save_epoch 399 --n_shot 1
- DATASET: miniImagenet/cub/cars/places/plantae
python test.py --name Tt-Net-target-set-cub-1shot --dataset DATASET --save_epoch 99 --n_shot 1
- DATASET: miniImagenet/cub
- testing for ME-D2N
python test_twoPaths.py --name ME-D2N-target-set-cub-1shot --target_set cub --dataset DATASET --save_epoch 399 --n_shot 1
- DATASET: miniImagenet/cub
We also provide our pretrained models as follows: (coming soon
- just take them in the right dir. Take ME-D2N for the 1-shot as an example, rename it as 399.tar, and move it to the
ouput/checkpoints/ME-D2N-target-set-cub-1shot/
If you find our work or codes useful, please consider citing our work ﾍ|･∀･|ﾉ*~●
@inproceedings{fu2022me,
title={ME-D2N: Multi-Expert Domain Decompositional Network for Cross-Domain Few-Shot Learning},
author={Fu, Yuqian and Xie, Yu and Fu, Yanwei and Chen, Jingjing and Jiang, Yu-Gang},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={6609--6617},
year={2022}
}