karalets

Hello there, Thank you for your note. This is based on a mis-communication on my behalf: TyXe moved a while ago to https://github.com/TyXe-BDL/TyXe but I did not get rid of...

Hey, Thanks for adding this. Can we unpack these results and contrast them with MSE? Also, if you ignore the variance from the neural network output and use the loglikelihood...

Great, that means we have numerical instabilities with learning sigma. One thing immediately comes to mind: sigma has to be positive, so typically we learn a transformed version of it...

You misunderstand the purpose of the noise model. It can only do this if there is no measurement error whatsoever and the mean can always be great. If, however, there...

Again, please point me to the code now, there are best practices here that I would like to encourage. This is almost certainly the source of the behavior you encounter.

Also if you look into my example above in this issue, I showed how to do this exactly in code and I never used an abs. reminder: ``` def condition(self,...

> Now I've constraint it to be positive by adding a parameter transformation function. > > https://github.com/choderalab/pinot/blob/2f2996311fd418f24b3f1a9c94aa654f39605d95/pinot/net.py#L23 Yeah abs is no good. Use Exp (you could also do softplus). Abs...

> Yes. The only difference is abs rather than exp. I'll change it to exp now. > > https://github.com/choderalab/pinot/blob/master/pinot/net.py#L23 Please retry training after this and compare to the likelihoods and...

That makes no sense... maybe play with hyperparameters and plot some loss curves to inspect what is actually happening? MSE is clearly the wrong model, it can't be that a...

ok then there is a numerical issue somewhere, that object should not even need the epsilon... especially since the epsilon changes the noise model to be 'at least epsilon' You...