bayesbay.BayesianInversion

class bayesbay.BayesianInversion(parameterization, log_likelihood, n_chains=10, walkers_starting_states=None, save_dpred=True)

A high-level class for Bayesian inversion using Markov chain Monte Carlo.

This is a subclass of BaseBayesianInversion that combines perturbations from parameterization and log_likelihood and initializes a bayesbay.MarkovChain for each walker.

Parameters:

  • parameterization (bayesbay.parameterization.Parameterization) – Pre-configured parameterization describing model structure, priors, and parameter-space perturbations.

  • log_likelihood (bayesbay.likelihood.LogLikelihood) – Log-likelihood object used to score proposed states. Its perturbation functions (e.g., hierarchical target updates) are merged with those from parameterization.

  • n_chains (int, optional) – Number of Markov chains to run. Default is 10.

  • walkers_starting_states (List[bayesbay.State], optional) – Optional starting state for each chain. If provided, the list length must be equal to n_chains. If None, each chain initializes from parameterization.

  • save_dpred (bool, optional) – Whether to store predicted data in saved chain states. Default is True.

Raises:

  • TypeError – If log_likelihood is not an instance of bayesbay.likelihood.LogLikelihood.

  • ValueError – If walkers_starting_states is provided and its length does not match n_chains.

Reference Details

chains

The MarkovChain instances of the current Bayesian inversion

get_results(keys=None, concatenate_chains=True)

To get the saved states from current inversion

Parameters:

  • keys (Union[str, List[str]]) – key(s) to retrieve from the saved states. This will be ignored when states are not of type State or dict

  • concatenate_chains (bool, optional) – whether to aggregate samples from all the Markov chains or to keep them seperate, by default True

Returns:

a dictionary from name of the attribute stored to the values, or a list of saved states (if the base level API is used, and states are not of type State or dict)

Return type:

Union[Dict[str, list], list]

static get_results_from_chains(chains, keys=None, concatenate_chains=True)

To get the saved states from a list of given Markov chains

Parameters:

  • chains (Union[BaseMarkovChain, List[BaseMarkovChain]]) – Markov chain(s) that the results are going to be extracted from

  • keys (Union[str, List[str]]) – key(s) to retrieve from the saved states. This will be ignored when states are not of type State or dict

  • concatenate_chains (bool, optional) – whether to aggregate samples from all the Markov chains or to keep them seperate, by default True

Returns:

a dictionary from name of the attribute stored to the values, or a list of saved states (if the base level API is used, and states are not of type State or dict)

Return type:

Union[Dict[str, list], list]

run(sampler=None, n_iterations=1000, burnin_iterations=0, save_every=100, verbose=True, print_every=100, parallel_config=None)

To run the inversion

Parameters:

  • sampler (bayesbay.samplers.Sampler, optional) – a sampler instance describing how chains intereact or modify their properties during sampling. This could be a sampler from the module bayesbay.samplers such as bayesbay.samplers.VanillaSampler (default), bayesbay.samplers.ParallelTempering, or bayesbay.samplers.SimulatedAnnealing, or a customised sampler instance.

  • n_iterations (int, optional) – total number of iterations to run, by default 1000

  • burnin_iterations (int, optional) – the iteration number from which we start to save samples, by default 0

  • save_every (int, optional) – the frequency with which we save the samples. By default a state is saved every after 100 iterations after the burn-in phase

  • verbose (bool, optional) – whether to print the progress during sampling or not, by default True

  • print_every (int, optional) – the frequency with which we print the progress and information during the sampling, by default 100 iterations

  • parallel_config (dict, optional) – keyword arguments passed to joblib.Parallel. Ignored when len(self.chains) is 1

set_perturbation_funcs(perturbation_funcs, perturbation_weights=None)
update_log_likelihood_targets(targets)

function to be called by self.log_likelihood when there are more target(s) added to the inversion. This method updates the perturbation function list for the current inversion and its chains, and initializes the current_state for each chain when there are hierarchical targets added.

This method is called because we register self (i.e. the current BayesianInversion) as an observer to its log likelihood instance.

Parameters:

targets (List[Target]) – the list of targets added to self.log_likelihood

back to top