ssamot / dann-keras Goto Github PK

Hey, thanks for the simple and sound code. Very useful!

I was wondering which part is implementing the idea of this paper:
"Revisiting batch normalization for practical domain adaptation"

Hi,

Thanks so much for creating this notebook, it's a very nice simple example of domain adaptation =)

I just have a quick question regarding this line:
stats2 = domain_classification_model.train_on_batch(X_adv, [y_adversarial_2])

Why do you train the domain model on X0 (source) and X1 (target) and use

y_adversarial_2 = to_categorical(np.array(([0] * batch_size + [1] * batch_size)))

I'm trying to wrap my head around why it's trained on inverted labels, would really appreciate your thoughts =) Thanks so much for your help!

Since we can't use the class labels of Xt, how are you y_class as target labels for source classifier part of combined model. And while testing why are you using embedding model to predict instead of using source classification model directly???

Hi Thank you for sharing this code,