MCEx is an experimental package designed to be allow experimentation with MCMC package design. It's goal is to be simple to use, understand, extend and improve, while still being fast. The hope is that some of the lessons learned in this experimental package lead to improvements in PyMC. People are encouraged to fork MCEx to try out their own designs and improvements as well as make criticisms.
For a tutorial on basic inference, see tutorial.py in the examples folder.
Some design decisions
| Design decision | Advantages | Disadvantages |
|---|---|---|
| Computational core outsourced to Theano |
|
|
| Random variables, distributions, chains, chain history, and model all distinct |
|
|
| Functional style design |
|
|
- Simplify standard usage.
- Build a GPU example
- Give sample a way of automatically choosing step methods.
- Do some profiling to see why likelihoods are slower in mcex than pymc
- Fix step_methods.gibbs.categorical so that it's faster, currently very slow.
- Implement a potential object which can take incomplete covariances and exploit the conditional independence of nodes to do the whole calculation
- Build examples showcasing different samplers
- Reconsider nphistory design
- Make HMC and related automatically choose a variance/covariance
- Eliminate model arguments to AddVar, AddData, step methods etc by making it hidden and default
- Eliminate test = argument to AddVar by allowing the user to specify a 'test point'.