Juan M. Cruz-Martinez

I don't understand what you mean. The easiest way looks much more complex to me since you need to filter out things and any bug there will "break" the validation.

Also, I'm not completely sure you can achieve your goal here? You need to compute everything twice for every epoch just the same.

> where the tuple are two separate layers containing the respective losses Yes, but your goal is to reduce the number of calls. However, you will need to call the...

It is a big deal because that tiny change can move you from a physical to an unphysical situation by means of positivity, integrability and probably even normalisation. But also,...

Check the speed up you would get in a fit and whether it doesn't become much more complicated. The forward pass is not the heaviest part of the calculation and...

If the final code is not very complex I'd be happy with it. From what you explain in the comments it looks complicated. Specially the idea of the internal filtering...

If possible it'd be better if the stopping remains outside of Keras since it is not a given that the conditions for stopping would always be machine-learning-compatible. Moreover, it would...

It was done with a much older version of the code (https://inspirehep.net/literature/1834151) but the point is that doing that kind of thing becomes much more difficult as more things become...

> which is a pain to fix It's ok, but I'm waiting for #1597

It does look wrong, specially since there's nothing that would justify a 10^-7, isn't there? The data is all > 1 so it cannot be a add vs mult problem,...