Skip to content

Latest commit

 

History

History
19 lines (11 loc) · 1.1 KB

README.md

File metadata and controls

19 lines (11 loc) · 1.1 KB
Raw

Asymmetric Polynomial Loss for Multi-Label Classification

Source code for ICASSP 2023 paper: Asymmetric Polynomial Loss for Multi-Label Classification.

APL is one step further beyond BCE-loss, ASL-loss, and PolyLoss. Surprisingly effective. Theoretically and experimentally validated.

Our code is based on ASL.

Quick Usage:

  1. Replace your torch.nn.BCELoss() or ASL-loss with APLloss.py.

  2. Adjust the parameters according to the discussion in the paper.

  3. See the stable performance improvement.

Experiment results:

We conduct experiments on Text Classification, Relation Extraction, and Image Classification.

Please click the above links and follow the steps in Quick Usage.