Aron Jansen

Seems to work fine, and gives the same results as trvl-mask-layers.

Ok perfect, thanks :)

@scarlehoff @Radonirinaunimi Positivity is included in the validation model, I remember we discussed this before, and if I remember correctly there was some disagreement on whether this was necessary or...

What I mean is we would have one model of the form (say we only have the DEUTERON observable) `x -> pdf -> DEUTERON -> (DEUTERON_tr, DEUTERON_val)`, where the tuple...

In this PR I've already decoupled the computation of the observable from the masking+loss, that was quite simple and gives identical results. The tricky part is how to use that...

Ah I hadn't thought about that, you're right that conventionally the validation at step t is computed after training for t steps. My proposal would have a shift by one...

True, but that should be easy to solve. Just save the weights at every step and when the stopping condition hits, instead of just stopping, revert to the previous epoch....

This old tensorboard profile illustrates the speedup. The gaps will be mostly removed by epochs-to-batches. Of the rest the validation step is more than 50% of the training step. ...

Yes the experimental one should stay separate I agree, it also doesn't play any role in performance. And I agree the approach I proposed is definitely not the cleanest. What...

## Update I made some progress, and have good hopes that a much nicer implementation than I suggested above is possible. (Just hope still because I haven't looked at reverting...