search

BorealisAI / advertorch

A Toolbox for Adversarial Robustness Research
GNU Lesser General Public License v3.0
1.31k stars 198 forks source link

Why cw_loss has a 50 intercept? #72

Open zzzace2000 opened 4 years ago

zzzace2000 commented 4 years ago

In the code, I find the cw_loss has an intercept of 50.

https://github.com/BorealisAI/advertorch/blob/98f9b41d7c263a47b23c66f5a576fa7969b82c9b/advertorch/loss.py#L76

Is there a particular reason to shift the offset to 50? That basically makes the clamp function useless?

gwding commented 4 years ago

@zzzace2000 Thanks for the observation! I agree with you on this. This could be due to my carelessness in distinguishing cw adversarial example and "high-confidence" cw adversarial example, and the offset should at least be an adjustable parameter. I'll work on fixing this.