Question about angular center loss about mean-shifted-anomaly-detection HOT 4 CLOSED

I think there were two losses in the paper right?
The authors did use the angular loss in the next line in the code. I might be missing something though.

from mean-shifted-anomaly-detection.

I misunderstood. Actually it is the angular loss.

from mean-shifted-anomaly-detection.

Hi, I see the angular center loss is expressed as follow:

I'm confused why it is expressed as such math formual?
and i see the corresponding code is:
out_1 = model(img1)
out_2 = model(img2)
out_1 = out_1 - center
out_2 = out_2 - center
center_loss = ((out_1 ** 2).sum(dim=1).mean() + (out_2 ** 2).sum(dim=1).mean())
which seems like the traidional center loss. Could you please explain?

I also think about this, and steal can not understand. I think the right code about the angular center loss should be write like this:

def angular_loss(out_1, out_2, center):
        out_1 = F.normalize(out_1, dim=-1)
        out_2 = F.normalize(out_2, dim=-1)

        center_loss = -((out_1 * center).sum(dim=1).mean() + (out_2 * center).sum(dim=1).mean())

        return center_loss

out_1 = model(img1)
out_2 = model(img2)

center_loss = angular_loss(out_1, out_2, center)

out_1 = out_1 - center
out_2 = out_2 - center

loss = contrastive_loss(out_1, out_2) + center_loss

from mean-shifted-anomaly-detection.

Hi
Please note that optimizing the Euclidean metric when dealing with unit vectors is proportional to optimizing the angular distance multiplied by a constant factor. Therefore, this optimization process is equivalent.
Hope this clarifies your questions.

from mean-shifted-anomaly-detection.