Why cw_loss has a 50 intercept?

[zzzace2000]

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

advertorch/advertorch/loss.py

Line 76 in 98f9b41

loss = clamp(elementwise_margin(input, target) + 50, 0.)

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

[gwding]

@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.