irakorshunova / bruno Goto Github PK

Why is the trainable variable of the variance of the GP/TP kernel (v in Eq. 9 of the BRUNO paper, diagonal values on the kernel matrix) expressed as the inverse softplus of the square root of the actual variance?

And then actual variance is recovered like this:

Is it simply to ensure v is non-negative? But then taking the square of self.var_vbl would suffice for that. Is the softplus some kind of trick for stability or convergence (and if so is it documented anywhere)? Thank you!

P.S. What does the m1 in m1_shapenet stand for? 🤔

Hello, thank you for open-sourcing the code! I have a few high-level questions about the models:

1. Why is validation only done for RNN and not conditional?

In the original RNN version, there is validation done during training:

Whereas in the conditional version, eval_loss is never used:

2. Is conditional BRUNO not maximizing joint (conditional) log likelihood?

BRUNO is clearly maximizing the joint log likelihood:

However, conditional BRUNO does not seem to be maximizing the joint conditional log likelihood... or is it?

3. "Conditional de Finetti" is not guaranteed

Do you think this is this a problem, or not really since in practice it works nonetheless?

Thank you very much!