Multi-scale Adaptive Task Attention Network for Few-Shot Learning

This code implements the Multi-scale Adaptive Task Attention Network (MATANet).

Our code will be released soon.

Citation

If you find our work useful, please consider citing our work using the bibtex:

@Article{chen2020multi,
	author  = {Chen, Haoxing and Li, Huaxiong and Li, Yaohui and Chen, Chunlin},
	title   = {Multi-scale Adaptive Task Attention Network for Few-Shot Learning},
	journal = {arXiv preprint arXiv:2011.14479},
	year    = {2020},
}

Prerequisites

• Linux
• Python 3.6
• Pytorch 1.0+
• GPU + CUDA CuDNN
• pillow, torchvision, scipy, numpy

Datasets

Dataset download link:

• miniImageNet
• CUB-200-2011
• Stanford Dog
• Stanford Car

Note: You need to manually change the dataset directory.

miniImageNet Few-shot Classification

• Train a 5-way 1-shot model based on Conv-64F:

python MATA_Train_5way1shot.py --dataset_dir ./datasets/miniImageNet --data_name miniImageNet

Test model on the test set:

python MATA_Test_5way1shot.py --dataset_dir ./datasets/miniImageNet --data_name miniImageNet --resume ./results/MATA_miniImageNet_MATA64_5Way_1Shot/model_best_test.pth.tar 

Fine-grained Few-shot Classification

• Data preprocessing (e.g., CUB-200-2011).
• Run the preprocessing script.

python ./dataset/CUB_200_2011_preprocessing.py

• Train a 5-way 1-shot model based on Conv-64F:

python MATA_Train_5way1shot.py --dataset_dir ./datasets/CUB_200_2011 --data_name CUBBirds

Test model on the test set:

python MATA_Test_5way1shot.py --dataset_dir ./datasets/CUB_200_2011 --data_name CUBBirds --resume ./results/MATA_CUBBirds_MATA64_5Way_1Shot/model_best_test.pth.tar 

Pretrained models

We also provide some of the pre-trained models. You can run the following command to evaluate the model

python MATA_Test_5way1shot_fg.py --dataset_dir ./datasets/CUB_200_2011 --data_name CUBBirds --resume ./results/MATA_CUBBirds_MATA64_5Way_1Shot/model_best_test.pth.tar 

Contacts

Please feel free to contact us if you have any problems.

Email: haoxingchen@smail.nju.edu.cn