tycki 0.0.3.dev1

Fast MCMC in Python

tycki

A simple Python tool for MCMC sampling using the amazing MCMClib.

Installation

tycki is on PyPI, so just

pip install tycki

Usage

tycki provides four samples from MCMClib:

• Random-walk Metropolis-Hastings tycki.RWMH
• Metropolis-adjusted Langevin algorithm tycki.MALA
• Hamiltonian Monte Carlo tycki.HMC
• No-U-Turn sampler tycki.NUTS

For sampling, the algorithms just require a start point, a log density function and - depending on the algorithm - the gradient of the log density function:

import tycki

rwmh = tycki.RWMH()
samples = rwmh.sample(log_density=lambda x: -(x**2).sum(), x0=[0, 0])

mala = tycki.MALA(x0=[0, 0]) # starting point can also be given in constructor
samples = mala.sample(log_density=lambda x: -(x**2).sum(), grad_log_density=lambda x: -2*x)

If you're too lazy to derive the gradient of your log density function manually, you can also use AD tools like jax:

import tycki
import jax
from jax import grad

@jax.jit
def logp(x, s=1):
        return -.5*(x**2 / s**2).sum()

rwmh = tycki.RWMH(x0=[0, 0])
samples = rwmh.sample(logp, grad(logp))

mala = tycki.MALA(x0=[0, 0])
samples = mala.sample(logp, grad(logp))

Licensing

This project is based on the amazing MCMClib by Keith O'Hara, licensed under the Apache License 2.0. Modifications and additions in order to make the by Richard D. Paul are licensed under MIT License.

See the LICENSE file for the full text of the MIT License.