dime-sampler 1.0.0

Installation

Installing the repository version from PyPi is as simple as typing

pip install dime_sampler

in your terminal or Anaconda Prompt.

Usage

The package provides a direct drop-in replacement for emcee:

import emcee
from dime_sampler import DIMEMove

move = DIMEMove()

...
def log_prob(x):
  ...
# define your density function, the number of chains `nchain` etc...
...

sampler = emcee.EnsembleSampler(nchain, ndim, log_prob, moves=move)
...
# off you go sampling

The rest of the usage is analog to emcee. See below for getting a quickstart or have a look this tutorial for details. The parameters specific to the DIMEMove are documented here.

Tutorial

Lets look at an example. Let’s define a nice and challenging distribution (it’s the distribution from the figure above):

# some import
import emcee
import numpy as np
import scipy.stats as ss
from dime_sampler import DIMEMove
from dime_sampler.test_all import _create_test_func, _marginal_pdf_test_func

# make it reproducible
np.random.seed(0)

# define distribution
m = 2
cov_scale = 0.05
weight = (0.33, .1)
ndim = 35
initvar = np.sqrt(2)

log_prob = _create_test_func(ndim, weight, m, cov_scale)

log_prob will now return the log-PDF of a 35-dimensional Gaussian mixture with three separate modes.

Next, define the initial ensemble. In a Bayesian setup, a good initial ensemble would be a sample from the prior distribution. Here, we will go for a sample from a rather flat Gaussian distribution.

# number of chains and number of iterations
nchain = ndim * 5
niter = 5000

# initial ensemble
initmean = np.zeros(ndim)
initcov = np.eye(ndim) * np.sqrt(2)
initchain = ss.multivariate_normal(mean=initmean, cov=initcov).rvs(nchain)

Setting the number of parallel chains to 5*ndim is a sane default. For highly irregular distributions with several modes you should use more chains. Very simple distributions can go with less.

Now let the sampler run for 5000 iterations.

move = DIMEMove(aimh_prob=0.1, df_proposal_dist=10)
sampler = emcee.EnsembleSampler(nchain, ndim, log_prob, moves=move)
sampler.run_mcmc(initchain, int(niter), progress=True)

The setting of aimh_prob is the actual default value. For less complex distributions (e.g. distributions closer to Gaussian) a higher value can be chosen, which accelerates burn-in. The value df_proposal_dist sets the degrees of freedom for the proposal distribution of the independence move. 10 is a sane default and it is rather unlikely that this value must be changed.

The following code creates the figure above, which is a plot of the marginal distribution along the first dimension (remember that this actually is a 35-dimensional distribution).

# import matplotlib
import matplotlib.pyplot as plt

# get elements
chain = sampler.get_chain()
lprob = sampler.get_log_prob()

# plotting
fig, ax = plt.subplots(figsize=(9,6))
ax.hist(chain[-niter//2 :, :, 0].flatten(), bins=50, density=True, alpha=0.2, label="Sample")
xlim = ax.get_xlim()
x = np.linspace(xlim[0], xlim[1], 100)
ax.plot(x, ss.norm(scale=np.sqrt(initvar)).pdf(x), "--", label="Initialization")
ax.plot(x, ss.t(df=10, loc=move.prop_mean[0], scale=move.prop_cov[0, 0] ** 0.5).pdf(x), ":", label="Final proposals")
ax.plot(x, _marginal_pdf_test_func(x, cov_scale, m, weight), label="Target")
ax.legend()

To ensure proper mixing, let us also have a look at the MCMC traces, again focussing on the first dimension.

fig, ax = plt.subplots(figsize=(9,6))
ax.plot(chain[:, :, 0], alpha=0.05, c="C0")

Note how chains are also switching between the three modes because of the global proposal kernel.

While DIME is an MCMC sampler, it can straightforwardly be used as a global optimization routine. To this end, specify some broad starting region (in a non-Bayesian setup there is no prior) and let the sampler run for an extended number of iterations. Finally, assess whether the maximum value per ensemble did not change much in the last few hundred iterations. In a normal Bayesian setup, plotting the associated log-likelihood over time also helps to assess convergence to the posterior distribution.

fig, ax = plt.subplots(figsize=(9,6))
ax.plot(lprob, alpha=0.05, c="C0")
ax.plot(np.arange(niter), np.max(lprob) * np.ones(niter), "--", c="C1")

References

If you are using this software in your research, please cite

@techreport{boehl2022mcmc,
author={Gregor Boehl},
title={Ensemble MCMC Sampling for Robust Bayesian Inference},
journal={Available at SSRN 4250395},
year={2022}
}