search

IssamLaradji / M-ADDA

Domain Adaptation Based on the Triplet Loss
113 stars 27 forks source link

Fatal error in your paper #7

Closed jiajinuiuc closed 4 years ago

jiajinuiuc commented 4 years ago

You are solving unsupervised domain adaptation, however, I find out you use target label in your training procedure. That is cheating ！！

IssamLaradji commented 4 years ago

which line?

IssamLaradji commented 4 years ago

Ah look carefully. I haven't used the target labels.

label_tgt is just the discriminator label which is all ones for training the encoder or all zeros for training the discriminator.

jiajinuiuc commented 4 years ago

Sorry, this line: fit_center( src_model, tgt_model, src_trainloader, tgt_trainloader, tgt_opt, epochs=1) And in your paper: image You are using target label to regularize your embedded target feature distribution. Please correct me if I am wrong.

IssamLaradji commented 4 years ago

That's a typo yah. In Equation 4, you will see that the labels are not used, and in the code https://github.com/IssamLaradji/M-ADDA/blob/master/losses.py#L27-L57, the labels are not used as well.

IssamLaradji commented 4 years ago

I am closing the issue as there is no cheating and the paper is doing proper unsupervised domain adaptation.