TuringPosteriorDB
Ok, here's the current rundown:
Models
Models are in src/models. Each model has it's own manifest.toml and project.toml, but is not currently wrapped in
any sort of struct to handle conditioning. I don't think I fully understand the benefit of that yet.
Benchmarking
I wrote a quick wrapper function in src/helper-functions/benchmarking-functions.jl that feels way too hacky to keep. But it seems to work right now.
This function takes a posterior identifier, e.g. Rate_1_data-Rate_1_model, and then runs the TuringBenchmarking.jl suite for the corresponding Turing model.
Also:
get_turing_samples("PDB_Unique_Identifier")now works, e.g.:
get_turing_samples("Rate_1_data-Rate_1_model")will find the Turing model, find the data, fit the model with NUTS() (need to add some options here),
and then return the samples.
As does the equivalent Stan function:
get_stan_samples("Rate_1_data-Rate_1_model")Testing
Nothing yet
Notes on Models
| Model | Notes |
|---|---|
| Rate_5_model | AD Broken for posterior predictive checks |
| Rate_4_model | AD Broken for posterior predictive checks |
| Kilpisjarvi | Check for divergences and correctness -- large variance in sample mean for both Stan and Turing |
Actually all the transformed parameter models are broken.