An implement of the NeurIPS 2022 paper: Exploring the Algorithm-Dependent Generalization of AUPRC Optimization with List Stability.

• Ubuntu 16.04
• CUDA 11.1
• Python 3.8.10
• Pytorch 1.8.2+cu11

See requirement.txt for others.

Download SOP, PKU VehicleID, and iNaturalist. Unzip these files and place then in ./data/[dataset]/images.

Download the pretrained model of ResNet-50 in ./pretrained_models.

Run the following command for training & validation

bash scripts/run.sh config/$DATASET/$CONFIG $gpu_id

For example,

bash scripts/run.sh config/iNaturalist/soprc_sgd.yaml 0

bash scripts/test.sh config/$DATASET/$CONFIG $gpu_id

For example,

bash scripts/test.sh config/iNaturalist/soprc_sgd.yaml 0

The following methods are provided in this repository (see Appendix in our paper):

• Pairwise Losses, including Contrastive Loss, Triplet Loss, Multi-Similarity (MS) Loss, Cross-Batch Memory (XBM)
• Ranking-Based Losses, including SmoothAP, FastAP, DIR, BlackBox, Area Under the ROC Curve Loss (AUROC), and SOPRC (Ours)

See losses/loss_warpper.py for usage.

By default, these losses take a dict with keys "feat" and "target" as input. Here the feature is an $(N\times M) \times D$ tensor, where $N$ is the number of ids, $M$ is the number of positive examples for each id and $D$ is feature dimension. The target is an $(N\times M) \times 1$ tensor, where the first $M$ examples belong to the same id, and so on. See config/demo.yaml for more details on configures. For example, by setting batchsize = 224, num_sample_per_id = 4, output_channels = 512, we have $N = 56, M = 4, D = 512$, and the target could be $[2,2,2,2,1,1,1,1,4,4,4,4,...]$.

If this code is helpful to you, please consider citing our paper:

@inproceedings{wen2022exploring,
  title={Exploring the Algorithm-Dependent Generalization of AUPRC Optimization with List Stability},
  author={Wen, Peisong and Xu, Qianqian and Yang, Zhiyong and He, Yuan and Huang, Qingming},
  booktitle={Annual Conference on Neural Information Processing Systems},
  year={2022}
}