David Widmann comments

Results 1467 comments of


                                            David Widmann

Integrating out variables via Laplace approximation (or other methods)

And the proposal in https://github.com/TuringLang/DynamicPPL.jl/issues/94

Gibbs gives different sampling results with fixed rng

Which version of Turing do you use? IIRC in some version particle filter methods did not respect the user-provided RNG but this should be fixed.

Gibbs gives different sampling results with fixed rng

Hmm this is unfortunate, you use the latest version :smile: So it seems there's a bug and at some point we don't pass around the user-provided RNG but fall back...

Gibbs gives different sampling results with fixed rng

Sure, then we can check easily if we managed to fix the problem.

Gibbs gives different sampling results with fixed rng

Yep, this seems to confirm my assumption that the problem is the particle filter method. I suspected the recent change in AdvancedPS to `TracedRNG` (and some default initialization there which...

Refactoring Glue Code

VarInfo isn't removed, so we still need all VarInfo code, don't we? We rather have to add some SimpleVarInfo code as well if it is added, or preferably generalize the...

Question: Which package is the MCMCserial() part of? been working through the notebook and have had problems with it not being defined hence that part of the notebook does not run

`MCMCSerial` (with capital `S`) is defined in AbstractMCMC. But users don't have to know this since it is re-exported by Turing. That is, if you run `using Turing` it will...

Question: Which package is the MCMCserial() part of? been working through the notebook and have had problems with it not being defined hence that part of the notebook does not run

Did you manage to fix the problem?

Question: Which package is the MCMCserial() part of? been working through the notebook and have had problems with it not being defined hence that part of the notebook does not run

Can you post the output of `] st --manifest`?

Errors in Sampling When Parameter Bounds Depend on Parameters

You have to define a custom bijector since the support of the distribution of `y` is stochastic. The discussion in https://github.com/TuringLang/Turing.jl/issues/1270 shows how this can be done.