This PR refactors classifier, such that all tensors are moved to annotated methods. Then it introduces flag for adversarial classification. In this setup, the classifier is optimized for better classification, however, the input to the classifier is optimized to worsen the representation for the classifier via the gradient reversal method.
This PR refactors classifier, such that all tensors are moved to annotated methods. Then it introduces flag for adversarial classification. In this setup, the classifier is optimized for better classification, however, the input to the classifier is optimized to worsen the representation for the classifier via the gradient reversal method.