summer intern feedback and open issues #8
Description
I just copied his notes from a file on slack to discuss them here (similar to https://github.com/lacerbi/pyvbmc/issues/28).
Things to investigate:
The diagnostics computed in slice sampling differ slightly from what is in pymc, what is the cause of this?
- original MATLAB split slice sample chain into 4 parts for whatever reason, in gpyreg and pymc the split is to 2
- pymc and gpyreg/original MATLAB have slightly different formula
Improvements:
- Docstrings of slice sampling might need editing for terminology, e.g. use of "stochastic parameters"
- this was not originally written by me, but terminology from pymc
- Posterior object is slgithly unnecessary looking
Issues:
- Test for bayesian quadrature with noise fails sometimes after latest bugfix, look into that a bit
- Calling predict without giving hyperparameters crashes the library, implement an exception for this
- Calling quad also crashes without hyperparameters, and no data GP as well
- We can call update with too large (or too small I think) hyperparameter array, causing silent bugs
OpenBLAS:
- OpenBLAS needs to have threading set to 1 for it to function well.
- export OPENBLAS_NUM_THREADS=1 fixes the problem though it restricts parallelism a bit
- this can be done by the user, or within the program as well I believe
- another way to solve this is to use another BLAS library such as Intel MKL, which is the default for Anaconda
- however, Intel MKL has another issue, namely than on machines with an AMD processor performance tanks
- this can currently be fixed on AMD machines by export MKL_DEBUG_CPU_TYPE=5
- however, there is no guarantee that this fix will work since it is based on undocumented features that are subject to change