Patrick Orlando
Patrick Orlando
Great, @VladLujerdeanu! Thanks for sharing.
I'd suggest using loss for early stopping. These curves looks like overfitting, but they also suggest that your train and val sets might not follow the same statistics. It could...
And this behaviour happens for different splits? It does look odd. It is possible for accuracy to increase whilst loss increases but I'm still not sure what is going on....
Hey @ydennisy, I'm not sure if you still need this, but I was working on it too and thought I'd share for you or anyone else who comes here looking...
Hey @kingdavescott, The most straightforward way to do this would be to create a Keras Layer that you can then make part of a model. Here's an example. ```python import...
Hi @xiaoyaoyang, It's still an identity matrix. There may be candidates that the user has interacted with which are used as negatives for this batch, even though they will be...
> Just want to make sure I understand it correctly. Let's say User had interacted with 2 candidates, the matrix is sparse and if I write down the interaction matrix,...
Hi @jiunsiew, Whoops, looks like I made some mistakes in the examples. 😅 You are correct. I'll edit my answer to not confuse others.
@xiaoyaoyang, correct. The optimisation is attempting to push all scores for postive items higher, and all scores for negative items lower. If there is an accidental hit, we artificially set...
Hey @nblasco, You would need to save the query and candidate towers, and the optimiser state if you are using anything other than SGD. You could then load these models...