About Loss Function

[Oguzhanercan]

Hi, I have a question about the loss function choice. As I see, you have used dot product for similarity of features. Did you experiment using L1, L2 MSE, others or their combination. I am asking because I have no enough gpu memory to use more than 1 reward function at this pipeline. If the feature space is orthogonal (At high dimensional space, this is expected), in my experiments, it is hard to optimize dot product - cosine similarity. I test it with rectifid, similar work to yours. When I used more than 1 face recognition network, it fails to generate images with same identity. If the feature space orthogonality is not a problem, is there a key to solve this problem. I thought that regularization in your work might solve this problem, but if I understand correctly, you did not used it for this purpose. @sgk98