d-li14 / dhm Goto Github PK

thanks for the amazing paper. the git said : while no computatonal overheads are introduced after discarding those auxiliary branches during inference.
do you mean we can pick out the corresponding weights(main branch) from the trained weight???
Hope your replay! Thanks

I see in the code that head1 has a channel number of 1024, so why not stick with the main branch, which is 512.

To stabilize the training process and avoid further regularization
effect, we fork auxiliary classifiers from each node of
the backbone network with probabilities toggling between
zero and one in our main experiments, following a binary
sampling strategy along the axis of network forward propagation.

I am not sure what these words means. From the code https://github.com/d-li14/DHM/blob/master/imagenet.py#L326, it seems you just calculate all the combinations of mimicking losses.

Hi there, great paper! I had a question regarding this line in the paper:

Locations of such intermediate layers are dynamically drawn from a given discrete probability distribution at each training epoch.

I am a bit confused, how do you actually find out which intermediate layers to add auxiliary classifiers to? I also don't see a code snippet for that, I only see auxiliary classifiers from head_3 and head_4 hardcoded. Can you please help understand this line?

Thank You again!