We provide a PyTorch implementation of CovaMNet for few-shot learning. The code was written by Wenbin Li [Homepage].

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

Distribution Consistency based Covariance Metric Networks for Few-shot Learning.
Wenbin Li, Jinglin Xu, Jing Huo, Lei Wang, Yang Gao and Jiebo Luo. In AAAI 2019.

• Linux
• Python 3
• Pytorch 0.4
• GPU + CUDA CuDNN

• Clone this repo:

git clone https://github.com/WenbinLee/CovaMNet
cd CovaMNet

• Install PyTorch 0.4 and other dependencies (e.g., torchvision).

• miniImageNet.
• StanfordDog.
• StanfordCar.
• CUB-200.
  Thanks Victor Garcia for providing the miniImageNet dataset. In our paper, we just used the CUB-200 dataset. In fact, there is a newer revision of this dataset with more images, see Caltech-UCSD Birds-200-2011. Note, if you use these datasets, please cite the corresponding papers.

• Train a 5-way 1-shot model:

python CovaMNet_Train_5way1shot.py --dataset_dir ./datasets/miniImageNet --data_name miniImageNet

• Test the model (specify the dataset_dir and data_name first):

python CovaMNet_Test_5way1shot.py --resume ./results/CovaMNet_miniImageNet_Conv64_5_Way_1_Shot/model_best.pth.tar

• The results on the miniImageNet dataset:

• Data prepocessing (e.g., StanfordDog)
• Specify the path of the dataset and the saving path.
• Run the preprocessing script.

#!./dataset/StanfordDog/StanfordDog_prepare_csv.py
python ./dataset/StanfordDog/StanfordDog_prepare_csv.py

• Train a 5-way 1-shot model:

python CovaMNet_Train_5way1shot.py --dataset_dir ./datasets/StanfordDog --data_name StanfordDog

• Test the model (specify the dataset_dir and data_name first):

python CovaMNet_Test_5way1shot.py --resume ./results/CovaMNet_StanfordDog_Conv64_5_Way_1_Shot/model_best.pth.tar

• The results on the fine-grained datasets:

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

@inproceedings{li2019CovaMNet,
  title={Distribution Consistency based Covariance Metric Networks for Few-shot Learning},
  author={Li, Wenbin and Xu, Jinglin and Huo, Jing and Wang, Lei and Gao Yang and Luo, Jiebo},
  booktitle={AAAI},
  year={2019}
}