Rémi Louf
Rémi Louf
> Personally, I wonder if aapl/aesara would be the appropriate abstraction layer where one would implement marinalization of discrete variables in PyMC models. As discussed above, implementing marinalization in aesara...
I think marginalisation is actually out of scope for AePPL and a better fit for AeMCMC: I don't have an example in mind, but marginalising greedily might block some important...
Well, it looks like we have different views of what AePPL, and in that sense we're both right within our view. I don't know what you mean by PPL, so...
> Yes, and I see that very differently from: give me a logprob expression for a graph of RVs `x` and `y`, where I condition on `x`, and `y`, or...
> Sure, that's the question being discussed in #85 It's just a matter of giving preference to pure RV nodes or marginalized nodes, one of them has to be the...
> Definitely. Note that all the rewrites that have been implemented so far are distribution agnostic, but we do want to include distribution aware rewrites for the cases where closed...
A GP with the Ornstein-Uhlenbeck kernel is equivalent to an AR(1) process (cf Rasmussen & Williams p.212), so it is definitely worth trying if we have time. I'd need to...
Pinging @lapp0
What happens when the model is loaded on several GPUs? I think that was the rationale behind pinning the logits to device 0.
Note that this is going to become irrelevant once #926 is merged.