LibAUC 1.4.0

Introducing LibAUC 1.4.0

We are releasing a new version of LibAUC at long last!

New Features

• Added support for optimizing mAP at top-k positions by setting mAPLoss(top_k=k).
• Integrated 8 commonly used graph neural networks into the library, i.e., GCN, DeeperGCN, GIN, GINE, GAT, MPNN, GraphSAGE and PNA.
• Introduced cosine gamma schedule in self-supervised contrastive learning, which has been demonstrated in our recent work. To use cosine gamma schedule, we can set GCLoss('unimodal', gamma_schedule = 'cosine') and GCLoss('bimodal', gamma_schedule = 'cosine').
• Added datasets.webdataset, which supports loading image-text dataset stored in WebDataset format.

Tutorial Update

• Provided a tutorial to show how to train a GNN model by optimizing AUPRC with our novel APLoss and SOAP optimizer on a binary molecule classification task from ogbg-molpcba dataset.
• Provided a tutorial on optimizing global contrastive loss with SogCLR and cosine gamma schedule.
• Updated previous tutorials by adding a recommended pretraining part to significantly boost models’ performance.

What's Changed

• A new page was added on LibAUC website to show our active users.
• Added return_index argument to the class datasets.ImageFolder and added the explanation of return_index to our documentation.
• Changed the default eta value to 0.1 in pAUC_CVaR_Loss. Previously, the value is inconsistent with the documentation.
• Fixed the learning rate setting in SogCLR tutorial. Previously, the learning rate is incorrectly passed into the function.
• Fixed the circular import error when importing from metrics #57.

Acknowledgment

Team: Gang Li, Xiyuan Wei, Siqi Guo, Zihao Qiu, Tianbao Yang (Advisor)