how to visualize the loss landscape?

[seyeeet]

in the paper you mentioned and showed an example of visualizing the loss landscape.
Can you please explain how to do it ?
I would really appreciate it if you can share the code for that since I am new to this topic and your implementation will be super useful!

[jbeomlee93]

Hi @seyeeet, the loss landscape was visualized through the following process.

1. For an image x, compute the classification loss L(x).
2. Compute vec{n} by backward L, and obtain any random vector vec{r}.
3. Compute L(x+alpha*vec{n}+beta*vec{r}). -1<beta<1, alpha>0: adversarial attack, alpha<0: adversarial climbing
4. Plot (alpha, beta, L(x+alpha*vec{n}+beta*vec{r}))

Thanks!

[seyeeet]

Thanks for your explanation, it is more clear now.
Can you please explain a little more how to compute vec{n} and vec{r} and what is the shape of vec{n} and vec{r}.
I mean what vec{n} means, I mean what is the dimension of vec and are n and r integers?
Sorry I am new to this and your explanation is very helpful

[jbeomlee93]

vec{n} and vec{r} are actually matrices with the same spatial size as the input image.

vec{n} is a gradient map of classification loss L with respect to an image, which can be obtained similar to the following:

AdvCAM/obtain_CAM_masking.py

Lines 138 to 142 in fa08f0a

	model.zero_grad()
	img_single.grad.zero_()
	loss.backward()
	data_grad = img_single.grad.data

vec{r} is literally a random matrix.

Thanks.

[seyeeet]

thank you