Hong Ge
Hong Ge
For efficiency reasons, there are cases where we want to handle model parameters differently in logdensity and sampling functions. This might be of general interest to Turing users, so would...
See e.g. https://github.com/TuringLang/Turing.jl/issues/1851#issuecomment-1173963099
- Customised `sample` function. https://gist.github.com/torfjelde/cc5c41e97eb4c97e22a19b8440f6d506 - Benchmarking AD backends. https://gist.github.com/torfjelde/cc5c41e97eb4c97e22a19b8440f6d506
Hidden Markov Models are quite common in time series analysis. Since they involve discrete variables, HMC is not always appropriate, although we can still apply HMC for continuous parameters of...
The following is a simple benchmarking for loops, on 3 different reverse-mode AD implementations in Julia. At the moment, in terms of efficiency: `ReverseDiff` > `Tracker` > `Zygote`. ```julia julia>...
The current figure for MCMC on the landing page looks like: where MCMC sampling is depicted by loss surface with scattering points. This is not very representative of MCMC IMO....
The benchmark numbers on the wiki are seriously out-of-date, and probably misleading about Turing's performance. Better to update the numbers using the current releases. https://github.com/TuringLang/Turing.jl/wiki
The current `resample!` function requires an additional parameter to distinguish between normal particles and reference particle in conditional SMC. Perhaps a cleaner solution is to add a member field (e.g....
We did a lot of benchmarking in the StanCon [AdvancedHMC](https://xuk.ai/software/StanCon-AHMC.pdf) submission. It would be helpful to make it more visible on our website.
Thermodynamic methods provide a way to estimate marginal likelihoods (aka model evidence) using MCMC, serving as an alternative to nested sampling methods. Some related papers: 1. Friel, Nial, and Anthony...