I'm not sure what's the conclusion over DeViSE and older approaches (or extreme classification folks).
My paper on label embeddings is on Arxiv. Many Nvidia K80s worked hard to bring you these numbershttps://arxiv.org/abs/1607.05691
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I think you could unleash the full potential of embedding regression vs classification on the Flickr dataset BTW (100k+ labels)
-
if you want to have a look at Flickr, also look up http://arxiv.org/abs/1511.02251 by
@armandjoulin and colleagues.
End of conversation
New conversation -
-
-
Congrats! By the way - would 1 K80 be detected as 1 or 2 GPU? - what did you use for parallelism? Keras w/ tf.device()?
-
1 K80 would be 1 GPU. It's data parallelism (asynchronous SGD). It's pure TensorFlow (although I have the same setup using Keras now)
End of conversation
New conversation -
-
-
how do you decide which architecture you should invest time on?
-
just throw spaghetti at the wall, see what sticks
End of conversation
New conversation -
-
-
that's awesome that it's both faster *and* more accurate!
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
is test time slower? (ie because you have to do NN) have you thought about doing a discrete ver ala learned hash reprs?
-
yes, inference is incrementally slower (essentially it's like having one more FC layer).
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.