No worries about the protocol -- this is a welcome contribution. Thanks @robwhess!

Note that this needs to be coordinated with #522 which splits the accuracy from loss in AccuracyLayer. We'll merge these PRs in order. You might need to do a slight adjustment once #522 is in but it should be easy (if there are conflicts at all).

Now that #522 is merged, could you adapt the top_k code? and rebase against dev again?

If you don't want to do it I can do it myself.

@sguada, I rebased on dev, though something seems to have gone wrong, since each commit here is now replicated twice in the log. Anyway, the status of the code is correct. Let me know if there's an easy way to fix the log.

Oh, it looks like you did a merge instead of a rebase or something to that effect. Try this

# just to be sure you're up-to-date
git checkout dev
git pull
# file this weird branch as "broken" for reference
git branch -m dev-top-k-accuracy broken
# branch off the original end of your feature branch, then rebase to dev
git checkout -b dev-top-k-accuracy 5fc2682
git rebase dev

# ...fix any conflicts by editing, git add, and git rebase --continue

# overwite your current PR with the correct rebase
git push -f flickr dev-top-k-accuracy

and if that doesn't work give a shout.

EDIT: swapped the order of the treeish and new branch name for correctness.

Thanks @shelhamer. It would have taken some work to figure that all out on my own. The rebase should be done correctly now and the PR ready to go.

Thanks!