paper and code implementation are different

[chenslcool]

Thank your for nice job!
I notice that you use dot-product to measure simlarity between prediction in paper while use KL_div in code.

paper:

code:

NRC_SFDA/office-home/train_tar.py

Lines 315 to 317 in 1c26160

	loss += torch.mean(
	(F.kl_div(softmax_out_un, score_near, reduction='none').sum(-1) *
	weight.cuda()).sum(1)) #

I wonder why do you do this?
Thanks!

[Albert0147]

Hi, if you check pytorch document of kl_div, you will find they are indentical, without log here is just dot product.

[chenslcool]

...thanks 🤣