torch-btg 0.0.1

Bridging the Gap: Unifying the Training and Evaluation of Neural Network Binary Classifiers

Bridging the Gap: Unifying the Training and Evaluation of Neural Network Binary Classifiers

A PyTorch implementation of Bridging the Gap (BtG) losses including F-beta (F1), Accuracy, and AUROC.

Project Webpage: btg.yale.edu

PDF Paper

Citation:

@inproceedings{tsoi2022bridging,
  title         = {Bridging the Gap: Unifying the Training and Evaluation of Neural Network Binary Classifiers},
  author        = {Tsoi, Nathan and Candon, Kate and Li, Deyuan and Milkessa, Yofti and V{\'a}zquez, Marynel},
  booktitle     = {Advances in Neural Information Processing Systems},
  year          = {2022}
}

Usage

Install the torch-btg package:

pip install torch-btg

Use the desired loss in your code, for example,

• F1-loss:

from torch_btg.loss import f1_loss
...
criterion = fb_loss(beta=1.0)
...

• Accuracy loss:

from torch_btg.loss import accuracy_loss
...
criterion = accuracy_loss()
...

Development

Setup

python -m pip install --user tox

Then run tests with:

tox