ODIN#

class scio.scores.ODIN(*, act_norm=None, mode='raw', temperature=?, epsilon=0.0, fgm_norm=inf)[source]#

Bases: BaseScoreClassif

ODIN for classification.

Distilled softmax plus adversarial reinforcement.

Parameters:

temperature (float) – Temperature scaling factor.
epsilon (float) – Amplitude of the adversarial reinforcement. If zero, the method is only applying temperature scaling to logits. Defaults to 0.0.
fgm_norm (float) – Parameter \(p\) in the \(L^p\) adversarial reinforcement. Defaults to inf.
mode – See BaseScoreClassif.
act_norm – See BaseScore.

References

[LLS18]

Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of Out-of-Distribution image detection in Neural Networks. In International Conference on Learning Representations. 2018. URL: https://openreview.net/forum?id=H1VGkIxRZ.