#sampler #mcmc #model #sampling #python #data #affine-invariant

emcee

Implementation of Python's emcee affine-invariant mcmc ensemble sampler

6 releases (2 breaking)

Uses old Rust 2015

1.0.0-alpha.2 Aug 2, 2017
1.0.0-alpha.1 Jun 6, 2017
0.3.0 Jun 2, 2017
0.2.0 Jun 1, 2017
0.1.1 May 20, 2017

#353 in Science

Download history 55/week @ 2024-03-13 94/week @ 2024-03-20 302/week @ 2024-03-27 120/week @ 2024-04-03 290/week @ 2024-04-10 275/week @ 2024-04-17 58/week @ 2024-04-24 166/week @ 2024-05-01 20/week @ 2024-05-08 47/week @ 2024-05-15 36/week @ 2024-05-22 269/week @ 2024-05-29 379/week @ 2024-06-05 210/week @ 2024-06-12 47/week @ 2024-06-19 252/week @ 2024-06-26

1,029 downloads per month
Used in light-curve-feature

MIT license

69KB
1K SLoC

Rust 1K SLoC // 0.0% comments Python 182 SLoC // 0.2% comments

emcee

Join the chat at https://gitter.im/rust-emcee/Lobby Build Status Crates version Docs

A re-implementation of Dan Foreman-Mackey's emcee in Rust.

See the hosted documentation here

The fitting_a_model_to_data example is a re-creation of the "fitting a model to data" example from the emcee documentation.

Attribution

If you make use of emcee in your work, please cite Dan's paper (arXiv, ADS, BibTeX).

A copy of the original MIT license is given under DFM-LICENSE.

Basic usage

Implementing models

The sampler requires a struct that implements emcee::Prob, for example:

use emcee::{Guess, Prob};

struct Model;

impl Prob for Model {
    fn lnlike(&self, params: &Guess) -> f32 {
        // Insert actual implementation here
        0f32
    }

    fn lnprior(&self, params: &Guess) -> f32 {
        // Insert actual implementation here
        0f32
    }
}

The trait has a default implementation for lnprob which computes the product of the likelihood and prior probability (sum in log space) as per Bayes' rule. Invalid prior values are marked by returning -std::f32::INFINITY from the priors function. Note your implementation is likely to need external data. This data should be included with your Model class, for example:

struct Model<'a> {
    x: &'a [f32],
    y: &'a [f32],
}

// Linear model y = m * x + c
impl<'a> Prob for Model<'a> {
    fn lnlike(&self, params: &Guess) -> f32 {
        let m = params[0];
        let c = params[1];

        -0.5 * self.x.iter().zip(self.y)
            .map(|(xval, yval)| {
                let model = m * xval + c;
                let residual = (yval - model).powf(2.0);
                residual
            }).sum::<f32>()
    }

    fn lnprior(&self, params: &Guess) -> f32 {
        // unimformative priors
        0.0f32
    }
}

Initial guess

Next, construct an initial guess. A Guess represents a proposal parameter vector:

use emcee::Guess;

let initial_guess = Guess::new(&[0.0f32, 0.0f32]);

The sampler implemented by this create uses multiple walkers, and as such the initial guess must be replicated once per walker, and typically dispersed from the initial position to aid exploration of the problem parameter space. This can be achieved with the create_initial_guess method:

let nwalkers = 100;
let perturbed_guess = initial_guess.create_initial_guess(nwalkers);
assert_eq!(perturbed_guess.len(), nwalkers);

Constructing a sampler

The sampler generates new parameter vectors, assess the probability using a user-supplied probability model, accepts more likely parameter vectors and iterates for a number of iterations.

The sampler needs to know the number of walkers to use, which must be an even number and at least twice the size of your parameter vector. It also needs the size of your parameter vector, and your probability struct (which implements Prob):

let nwalkers = 100;
let ndim = 2;  // m and c

// Build a linear model y = m * x + c (see above)

let initial_x = [0.0f32, 1.0f32, 2.0f32];
let initial_y = [5.0f32, 7.0f32, 9.0f32];

let model = Model {
    x: &initial_x,
    y: &initial_y,
};

let sampler = emcee::EnsembleSampler::new(nwalkers, ndim, &model)
    .expect("could not create sampler");

Then run the sampler:

let niterations = 100;
sampler.run_mcmc(&perturbed_guess, niterations).expect("error running sampler");

Iterative sampling

It is sometimes useful to get the internal values proposed and evaluated during each proposal step of the sampler. In the Python version, the method sample is a generator which can be iterated over to evaluate the sample steps.

In this Rust version, we provide this feature by exposing the sample method, which takes a callback, which is called once per iteration with a single Step object. For example:

sampler.sample(&perturbed_guess, niterations, |step| {
    println!("Current iteration: {}", step.iteration);
    println!("Current guess vectors: {:?}", step.pos);
    println!("Current log posterior probabilities: {:?}", step.lnprob);
});

Studying the results

The samples are stored in the sampler's flatchain which is constructed through the flatchain method on the sampler:

let flatchain = sampler.flatchain().unwrap();

for (i, guess) in flatchain.iter().enumerate() {
    // Skip possible "burn-in" phase
    if i < 50 * nwalkers {
        continue;
    }

    println!("Iteration {}; m={}, c={}", i, guess[0], guess[1]);
}

Dependencies

~315–540KB