henuliulei / M2L

This repository is the code of paper "Multi-level Metric Learning for Few-shot Image Recognition".

Multi-level Metric Learning for Few-Shot Image Recognition

This code implements the Multi-level Metric Learning for Few-shot Image Recognition (M2L).

Citation

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

@Article{chen2021multi,
	author  = {Chen, Haoxing and Li, Huaxiong and Li, Yaohui and Chen, Chunlin},
	title   = {Multi-level Metric Learning for Few-Shot Image Recognition},
	journal = {arXiv preprint arXiv:2103.11383},
	year    = {2021},
}

Prerequisites

• Linux
• Python 3.7
• Pytorch 1.0+
• GPU + CUDA CuDNN
• pillow, torchvision, scipy, numpy

Datasets

Dataset download link:

• miniImageNet
• tieredImageNet
• 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 MML(KL) model based on Conv-64F (on miniImageNet dataset):

python MML_Train_1shot.py --method_name KL --dataset_dir ./datasets/miniImageNet --data_name miniImageNet

Test model on the test set:

python Test_Batch.py --method_name KL --dataset_dir ./datasets/miniImageNet --data_name miniImageNet --resume ./results/miniImageNet_KL_1shot\KL_BatchSize_4_Conv64F_miniImageNet_5Way_1Shot/model_best_test.pth.tar 

Pre-trained models

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

python Test_Batch.py --method_name Wass --dataset_dir ./datasets/tieredImageNet --data_name tieredImageNet --resume ./results/tieredImageNet_Wass_1shot\Wass_BatchSize_4_Conv64F_tieredImageNet_5Way_1Shot/model_best_test.pth.tar 

Contacts

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

Email: [e-mail]