Question about the loss

[sgtziggy]

https://github.com/sail-sg/AnyDoor/blob/3ba1ec621255b9fbaf3714f4f5769b8280af62fc/anydoor_llava.py#L138 As shown above, you put a minus sign before the cross entropy loss. I don't quite understand the meaning of this minus sign, can you explain it? Part of the purpose of the model is to want to output label: I want to destroy the whole world together when data containing trigger and SUDO is received. In order to achieve this goal, shouldn't cross entropy loss be minimized?

[Sammy42779]

Thanks for your question! In this code, we use gradient ascent to optimize the Universal Adversarial Perturbation (UAP) with the goal of making the model output the target label when a trigger is present (same as the without trigger setting). The gradient ascent update is implemented in line 168 as follows: momentum = mu * momentum + grad / torch.norm(grad, p=1) Therefore, we apply a negative sign to the loss to reverse its optimization direction for gradient ascent. This approach ensures the UAP is optimized for both cases effectively.

[snow-like-kk]

So, is the approach of adding a minus sign to the loss and then using gradient ascent equivalent to directly using gradient descent?