Closed peterzpy closed 3 years ago
Thank you for your attention.
I guess you see this in our code for imagenet & segmentation. As we state in the paper, we approximate the covariance matrices by their diagonals (see sec. 6.1) to save GPU memory (reduce the covariance tensor from 1000x2048x2048 to 1000x2048). You may check our code on cifar for the Vanilla ISDA.
Thanks for your reply. I have no more questions and you could close this issue. : )
It seems that you only use the diagonal elements of the covariance matrix. I wonder why not use the complete matrix to augment the feature? Looking forward to your reply.