Mimicking losses confusion about dhm HOT 1 CLOSED

As stated by the paper, forking auxiliary branch or not at each node of the backbone is defined as a dynamic process. Here the nodes with auxiliary branches imply that we fork with the probability 1, while all the other nodes without branches with probability 0. This is the so-called binary (0-1) sampling. You could further try Bernoulli sampling with the probability of 0.5 in these existing branches with the commented lines https://github.com/d-li14/DHM/blob/master/imagenet.py#L334-L336.

from dhm.