Awesome
Deep Relational Metric Learning
This repository is the official PyTorch implementation of Deep Relational Metric Learning, which is based on another repository GeDML.
News
- [2021-12-14]: Update gedml to the newest version 2.0.1.
Framework
Datasets
CUB-200-2011
Download from here.
Organize the dataset as follows:
- cub200
|- train
| |- class0
| | |- image0_1
| | |- ...
| |- ...
|- test
|- class100
| |- image100_1
| |- ...
|- ...
Cars196
Download from here.
Organize the dataset as follows:
- cars196
|- train
| |- class0
| | |- image0_1
| | |- ...
| |- ...
|- test
|- class98
| |- image98_1
| |- ...
|- ...
Requirements
To install requirements:
pip install -r requirements.txt
Training
Baseline models
To train the baseline model with the ProxyAnchor loss on CUB200, run this command:
CUDA_VISIBLE_DEVICES=0 python examples/demo.py --data_path /home/zbr/Workspace/datasets --save_path /home/zbr/Workspace/exp/DRML --device 0 --batch_size 180 --test_batch_size 180 --setting proxy_anchor --embeddings_dim 512 --proxyanchor_margin 0.1 --proxyanchor_alpha 32 --num_classes 100 --wd 0.0001 --gamma 0.5 --step 10 --lr_trunk 0.0001 --lr_embedder 0.0001 --lr_collector 0.01 --dataset cub200 --delete_old --warm_up 1 --warm_up_list embedder collector
DRML models
To train the proposed DRML model using the ProxyAnchor loss on CUB200 in the paper, run this command:
CUDA_VISIBLE_DEVICES=0 python examples/demo.py --data_path /home/zbr/Workspace/datasets --save_path /home/zbr/Workspace/exp/DRML --device 0 --batch_size 180 --test_batch_size 180 --setting proxy_anchor_drml --embeddings_dim 512 --features_dim 1024 --branch_num 4 --proxyanchor_margin 0.2 --proxyanchor_alpha 64 --num_classes 100 --wd 0.0001 --gamma 0.5 --step 10 --lr_trunk 0.00005 --lr_embedder 0.001 --lr_collector 0.01 --weight_recon_loss 1 --weight_repre_loss 10 --dataset cub200 --delete_old --warm_up 1 --warm_up_list embedder ensemble repre
Device
We tested our code on a linux machine with an Nvidia RTX 3090 GPU card. We recommend using a GPU card with a memory > 8GB (BN-Inception + batch-size of 120 ).
Results
The baseline models achieve the following performances:
| Model name | Recall @ 1 | Recall @ 2 | Recall @ 4 | Recall @ 8 | NMI |
|---|---|---|---|---|---|
| cub200-ProxyAnchor-baseline | 67.3 | 77.7 | 85.7 | 91.4 | 68.7 |
| cars196-ProxyAnchor-baseline | 84.4 | 90.7 | 94.3 | 96.8 | 69.7 |
Our models achieve the following performances:
| Model name | Recall @ 1 | Recall @ 2 | Recall @ 4 | Recall @ 8 | NMI |
|---|---|---|---|---|---|
| cub200-ProxyAnchor-ours | 68.7 | 78.6 | 86.3 | 91.6 | 69.3 |
| cars196-ProxyAnchor-ours | 86.9 | 92.1 | 95.2 | 97.4 | 72.1 |
COMING SOON
- We will upload the code for cross-validation setting soon.
Citation
If you find our work useful in your research, please consider citing:
@inproceedings{zheng2021deep,
title={Deep Relational Metric Learning},
author={Zheng, Wenzhao and Zhang, Borui and Lu, Jiwen and Zhou, Jie},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={12065--12074},
year={2021}
}