Features normalization computing centers

[joppichristian]

Thank you for sharing code!

I have a question about the function get_centers that computes the centers of clusters.
Did you forget to normalize features before multiply them per mask? Or simply the normalization is in the net forward?

Contrastive-Adaptation-Network-for-Unsupervised-Domain-Adaptation/solver/utils.py

Line 99 in bdea238

centers += torch.sum(feature * mask, dim=1)

[WDYIE]

I think that you can see the code normalizing features in clusteing.py :

def cos(self, pointA, pointB, cross):
    pointA = F.normalize(pointA, dim=1) #Cosine similarity only cares about direction
    pointB = F.normalize(pointB, dim=1)
    if not cross:
        return 0.5 * (1.0 - torch.sum(pointA * pointB, dim=1))
    else:
        NA = pointA.size(0)
        NB = pointB.size(0)
        assert(pointA.size(1) == pointB.size(1))
        return 0.5 * (1.0 - torch.matmul(pointA, pointB.transpose(0, 1)))

[HHHedo]

Hi, I have the same question. It seems that the center is the sum of features instead of mean.

[WDYIE]

Hi, There are normalize operation, Beacause Cosine similarity only cares about direction！

[HHHedo]

Hi, There are normalize operation, Beacause Cosine similarity only cares about direction！

Thanks for your reply!