panda-peter / visda2019-multisource Goto Github PK

Hi @Panda-Peter, thank you for open sourcing this implementation. I was looking through the MMDLoss class in mmd.py and saw these two lines in the forward() function:

line 78:

xyk_c_antidiag = (1 - torch.eye(n_labels, n_labels).cuda()) * xyk_c

line 81:

mmd -= cfg.LOSSES.MMD_NEG_WEIGHT * ( (n_labels - 1) * (xk_c_sum + yk_c_sum) - 2 * xyk_c_antidiag.sum() )  / neg_num

May I know what they are used for? According to my understanding of your class-level MMD loss, the gaussian kernel should be computed using instances from within the same class, and this is computed in line 77 with xyk_c_diag:

xyk_c_diag = torch.eye(n_labels, n_labels).cuda() * xyk_c

xyk_c_antidiag stores the computed gaussian kernel between different classes. Why is this needed? Thank you!

Hi @Panda-Peter thank you for sharing this code and your group's great approach to the VisDA challenge

I'm trying to determine what hyperparameters you all used for training CycleGAN.
From the text file you shared it looks like those are all the defaults?

Can you please confirm whether this is correct?
Just trying to make sure we understand your approach