Robust And Decomposable Average Precision for Image Retrieval (NeurIPS 2021)

This repository contains the source code for our ROADMAP paper (NeurIPS 2021).

Use ROADMAP

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Datasets

We use the following datasets for our submission

• CUB-200-2011 (download link available on this website : http://www.vision.caltech.edu/visipedia/CUB-200.html)
• Stanford Online Products (you can download it here : https://cvgl.stanford.edu/projects/lifted_struct/)
• INaturalist-2018 (obtained from here https://github.com/visipedia/inat_comp/tree/master/2018#Data)

Run the code

SOP

The following command reproduce our results for Table 4.

CUDA_VISIBLE_DEVICES=0 python roadmap/single_experiment_runner.py \ 'experience.experiment_name=sop_ROADMAP_${dataset.sampler.kwargs.batch_size}_sota' \ experience.seed=333 \ experience.max_iter=100 \ 'experience.log_dir=${env:HOME}/experiments/ROADMAP' \ optimizer=sop \ model=resnet \ transform=sop_big \ dataset=sop \ dataset.sampler.kwargs.batch_size=128 \ dataset.sampler.kwargs.batches_per_super_pair=10 \ loss=roadmap

With the transformer backbone :

CUDA_VISIBLE_DEVICES=0 python roadmap/single_experiment_runner.py \ 'experience.experiment_name=sop_ROADMAP_${dataset.sampler.kwargs.batch_size}_DeiT' \ experience.seed=333 \ experience.max_iter=75 \ 'experience.log_dir=${env:HOME}/experiments/ROADMAP' \ optimizer=sop_deit \ model=deit \ transform=sop \ dataset=sop \ dataset.sampler.kwargs.batch_size=128 \ dataset.sampler.kwargs.batches_per_super_pair=10 \ loss=roadmap

INaturalist

For ROADMAP sota results:

CUDA_VISIBLE_DEVICES='0,1,2' python roadmap/single_experiment_runner.py \ --multirun \ 'experience.experiment_name=inat_ROADMAP_${dataset.sampler.kwargs.batch_size}_sota' \ experience.seed=333 \ experience.max_iter=90 \ 'experience.log_dir=${env:HOME}/experiments/ROADMAP' \ optimizer=inaturalist \ model=resnet \ transform=inaturalist \ dataset=inaturalist \ dataset.sampler.kwargs.batch_size=384 \ loss=roadmap_inat

CUB-200-2011

For ROADMAP sota results:

CUDA_VISIBLE_DEVICES=0 python roadmap/single_experiment_runner.py \ 'experience.experiment_name=cub_ROADMAP_${dataset.sampler.kwargs.batch_size}_sota' \ experience.seed=333 \ experience.max_iter=200 \ 'experience.log_dir=${env:HOME}/experiments/ROADMAP' \ optimizer=cub \ model=resnet_max_ln \ transform=cub_big \ dataset=cub \ dataset.sampler.kwargs.batch_size=128 \ loss=roadmap

CUDA_VISIBLE_DEVICES=0 python roadmap/single_experiment_runner.py \ 'experience.experiment_name=cub_ROADMAP_${dataset.sampler.kwargs.batch_size}_sota_DeiT' \ experience.seed=333 \ experience.max_iter=150 \ 'experience.log_dir=${env:HOME}/experiments/ROADMAP' \ optimizer=cub_deit \ model=deit \ transform=cub \ dataset=cub \ dataset.sampler.kwargs.batch_size=128 \ loss=roadmap

The results are not exactly the same as my code changed a bit (for instance the random seed are not the same).

Contacts

If you have any questions don't hesitate to create an issue on this repository. Or send me an email at elias.ramzi@lecnam.net.

Don't hesitate to cite our work:

@inproceedings{
ramzi2021robust,
title={Robust and Decomposable Average Precision for Image Retrieval},
author={Elias Ramzi and Nicolas THOME and Cl{\'e}ment Rambour and Nicolas Audebert and Xavier Bitot},
booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
year={2021},
url={https://openreview.net/forum?id=VjQw3v3FpJx}
}

Resources

• Pytorch Metric Learning (PML): https://github.com/KevinMusgrave/pytorch-metric-learning
• SmoothAP: https://github.com/Andrew-Brown1/Smooth_AP
• Blackbox: https://github.com/martius-lab/blackbox-backprop
• FastAP: https://github.com/kunhe/FastAP-metric-learning
• SoftBinAP: https://github.com/naver/deep-image-retrieval
• timm: https://github.com/rwightman/pytorch-image-models
• PyTorch: https://github.com/pytorch/pytorch
• Hydra: https://github.com/facebookresearch/hydra
• Faiss: https://github.com/facebookresearch/faiss
• Ray: https://github.com/ray-project/ray