tf1997 / MATANet

This repository is the code of paper "Multi-scale Adaptive Task Attention Network for Few-Shot Learning".

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

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

Our code is based on CovaMNet.

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.

Few-shot Classification

• Train a 5-way 1-shot model based on Conv-128F (on miniImageNet dataset):

python MATA_Train.py --dataset_dir ./datasets/miniImageNet --data_name miniImageNet --way_num 5 --shot_num 1

Test model on the test set:

python MATA_Test.py --dataset_dir ./datasets/miniImageNet --data_name miniImageNet --way_num 5 --shot_num 1
./results/MATA_miniImageNet_MATA_5Way_1Shot_K5/model_best.pth.tar --basemodel MATA

Contacts

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

Email: [e-mail]