search

TuringLang / TuringGLM.jl

Bayesian Generalized Linear models using `@formula` syntax.
https://turinglang.org/TuringGLM.jl/dev
MIT License
70 stars 7 forks source link

Getting more numerical errors when setting `standardize=true` #37

Closed rikhuijzer closed 2 years ago

rikhuijzer commented 2 years ago

With an unstandardized dataset and standardize=false (the default), I got 3 numerical errors on a Logistic model. With standardize=true, I got hundreds of numerical errors. This shouldn't happen, right? With standardization, the sampler already starts roughly in the right place meaning that less numerical errors should occur.

storopoli commented 2 years ago

Is this during warmup or sampling?

rikhuijzer commented 2 years ago

No idea. I had to cancel the run because my whole system got unresponsive and I haven't touched any Turing model since then

storopoli commented 2 years ago

This is important. numerical errors on a warmup are pretty common.