ProxyGML: A Simple yet Powerful Metric Learning Strategy

Official PyTorch code for the NeurIPS 2020 spotlight paper "Fewer is More: A Deep Graph Metric Learning Perspective Using Fewer Proxies". In this paper, we propose a novel graph-based deep metric learning loss, namely ProxyGML, which is simple to implement. The pipeline of ProxyGML is as shown below.

Slides&Poster&Video

Slides and poster of our NeurIPS 2020 10-min spotlight talk are available. Each page of the slides contains comments, which we assume can be of help to better understand our work. Our 3-min poster video presentation is also available. Feel free to use our slides if you want to share our work in your group meeting or introduce it to your friends!

Installation

We recommend the following dependencies.

• pytorch==1.2
• pillow==5.2.0
• tqdm==4.26.0
• matplotlib==2.2.2
• pandas==0.23.4
• scipy==1.2.1
• scikit-learn==0.20.3
• scikit-image==0.14.2
• h5py==2.9.0

Datasets

1. Download three public benchmarks for deep metric learning

   • CUB-200-2011
   • Cars-196 (Img, Annotation)
   • Stanford Online Products (Link)

2. We train our model in the paradigm of classification. So all datasets are preprocessed into train/test parts as follows ( take CUB200-2011 for example) and stored in .\data

cub200                                         
└───train
|    └───0
|           │   xxx.jpg
|           │   ...
|
|    ...
|
|    └───99
|           │   xxx.jpg
|           │   ...

└───test
|    └───100
|           │   xxx.jpg
|           │   ...
|
|    ...
|
|    └───199
|           │   xxx.jpg
|           │   ...
|    ...

Getting Started

Note that a fine-grained combination of parameter $N$ and parameter $r$ resultes in better overall performance than that described in the paper.

CUB-200-2011

python train.py -b 32 --gpu 2 --dataset cub200 --freeze_BN --epochs 50 --dim 512 --r 0.05 -C 100 --N 12 --weight_lambda 0.3 --centerlr 0.03 --rate 0.1 --new_epoch_to_decay 20 40

Cars-196

python train.py -b 32 --gpu 2 --dataset cars196 --freeze_BN --epochs 50 --dim 512 --r 0.05 -C 98 --N 12 --weight_lambda 0.3 --centerlr 0.03 --rate 0.1 --new_epoch_to_decay 20 40

Stanford Online Products

python train.py -b 32 --gpu 2 --dataset online_products --epochs 50 --dim 512 --r 0.05 -C 11318 --N 1 --weight_lambda 0.0 --centerlr 0.3 --rate 0.1 --new_epoch_to_decay 20 40

Citation

If you find our paper or this project helps your research, please kindly consider citing our work via:

@inproceedings{Zhu2020ProxyGML,
  title={Fewer is More: A Deep Graph Metric Learning Perspective Using Fewer Proxies},
  author={Yuehua Zhu and Muli Yang and Cheng Deng and Wei Liu},
  booktitle={NeurIPS},
  year={2020}
}