Model calibration has helped in developing reliable deep learning models. In this pull request, I have added a new loss function NACL (https://arxiv.org/abs/2303.06268, https://arxiv.org/abs/2401.14487) which has shown promising results for both discriminative and calibration in segmentation.
Description
Model calibration has helped in developing reliable deep learning models. In this pull request, I have added a new loss function NACL (https://arxiv.org/abs/2303.06268, https://arxiv.org/abs/2401.14487) which has shown promising results for both discriminative and calibration in segmentation.
Future Plans: Currently, MONAI has some of the alternative loss functions (Label Smoothing, and Focal Loss), but it doesn't have the calibration specific loss functions (https://arxiv.org/abs/2111.15430, https://arxiv.org/abs/2209.09641). Besides, these methods are better evaluated with calibration metrics, Expected Calibration Error (https://lightning.ai/docs/torchmetrics/stable/classification/calibration_error.html).
Types of changes
./runtests.sh -f -u --net --coverage
../runtests.sh --quick --unittests --disttests
.make html
command in thedocs/
folder.