README.md

BayesMixPy: a Python interface to BayesMix

Installation

We suggest working in a virtual environment using, e.g., conda:

conda create --name bayesmix
conda activate bayesmix
conda install pip

After you have cloned the bayesmix github directory, navigate to the Python subfolder and install bayesmixpy using pip:

cd python
pip3 install -r requirements.txt
pip3 install -e .

Usage

bayesmixpy provides two functions: build_bayesmix and run_mcmc. The first one installs bayesmix and its executables for you, while the second one calls the executable that runs the MCMC sampler from Python.

Building bayesmix

To build bayesmix, in a Python shell or a notebook write

from bayesmixpy import build_bayesmix

n_proc = 4 # number of processors for building in parallel
build_bayesmix(n_proc)

this will print out the installation log.

Running bayesmix

To run_mcmc must define the model and the algorithm in some configuration files or text strings. See the documentation for more details.

For instance, to fit a Dirichlet Process Mixture on univariate data using a Normal-Normal-InverseGamma hierarchy using Neal's Algorithm 3, we use the following

eval_dens, n_clus_chain, best_clus, chains = run_mcmc(
    "NNIG",
    "DP",
    data,
    nnig_params,
    dp_params,
    algo_params,
    dens_grid,
    return_chains=True)

where data is a np.array of data points, dens_grid is a np.array of points where to evaluate the density, and nnig_params, dp_params and algo_params are defined as follows.

nnig_params="""
ngg_prior {
    mean_prior {
        mean: 5.5
        var: 2.25
    }
    var_scaling_prior {
        shape: 0.2
        rate: 0.6
    }
    shape: 1.5
    scale_prior {
        shape: 4.0
        rate: 2.0
        }
}
"""

This specifies that the base (centering) measure is a Normal-InverseGamma with parameters $(\mu_0, \lambda_0, a_0, b_0)$. Further, $\mu_0 \sim N(5.5, 2.25)$, $\lambda_0 ~ Gamma(0.2, 0.6)$, $a_0 = 1.5$, $b0 \sim Gamma(4.0, 2.0)$ [See the messages NNIGPrior and NNIGPrior::NGGPrior in the file hierarchy_prior.proto for further reference].

dp_params = """
gamma_prior {
    totalmass_prior {
        shape: 4.0
        rate: 2.0
    }
}
"""

This specifies that the concentration parameter of the DP has an hyperprior which is a Gamma distribution with parameters (4, 2).

algo_params = """
    algo_id: "Neal3"
    rng_seed: 20201124
    iterations: 2000
    burnin: 1000
    init_num_clusters: 3
"""

See the notebook in notebooks/gaussian_mix_uni.ipynb for a concrete usage example