rockysj/wdgrl

wdgrl's Issues

Hyperparameters for reproducing results for baselines

Can you provide the hyperparameters for reproducing the results for the baselines?
For example, there is one for MMD in office-caltech folder but the other methods have zero coefficients and I could not find any reference in paper.

Pytorch implementation

Hi, I'm trying to implement your work with the Pytorch framework. But I can't get good accuracy.
The training code I use is mainly from jvanvugt/pytorch-domain-adaptation
Have you implemented your work with Pytorch? Or Could you give me some advice about my question?

Many thanks！

Trying to understand part of the code of office_wd.py ?

In line 61 when you take gradients -
gradients = tf.gradients(critic_out, [h2_whole])[0]

What is the role of [0] ?

Code For Office-Caltech

Can you provide the code for Office-Caltech dataset please?

Trials for the experiments ?

Are we using only a single trial for the experiments in the paper ? If so, why is that ? Is it because, we are using the whole dataset at once for the domain adaptation.

mnist code

Hi, I think you made a mistake in in mnist code. You put both source and target to train the classifier (use y_ instead of ys_true) so the result is 99.1%, which is too good to be true.

Hyperparameter optimization ?

In the experiments, do you report the best results obtained over the hyper-parameter range or do you do some kind of validation. Validation is generally not possible for unsupervised domain adaptation.