proposal

samplermcmc.proposal

class m3c2.proposal.AdaptiveKDE(names, **kwargs)

Adaptive KDE proposal base on hypKDE library

kde_jump(x, chain, ims=0, **kwargs)

Return new sample draw from KDE distribution.

class m3c2.proposal.CovProposal(*args, **kwargs)

Propose jumps out of the gaussian distibution defined by the covariance matrix and mode built from chain. Besides the standard arguments of the Proposal: class it accepts the following parameters:

Parameters:

• adapt (int) – frequency (in steps) of the covariance update.

• ini_mode (np.array) – initial mode of the proposal distribution. Default: null vector.

• ini_cov (np.array) – initial covariance of the proposal dist. Default: diagonal 0.1 matrix.

• get_qxy (bool) – return the ration of the proposal probabilities. Default: True.

Returns:

propsed point, O, None, qxy

Return type:

list

get_proposal(y, chain=None, chains=None)

use update_cov=True in sampler options

update_cov(chain)

Update the covariance matrix

class m3c2.proposal.DE(names, DE_skip=1000, **kwargs)

A differential evolution sampler.

DE(x, chain=None, **kwargs)

Return actual sample using differential evolution of a given chain.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

DE_all(x, chains=None, chain=None, **kwargs)

Return actual sample using differential evolution of all chains.

Args: x: the current position chains, chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

class m3c2.proposal.Gibbs(names, subset_names=[], cov=None, **kwargs)

Proposal with jump in a subset of parameters only

SCAM(x)

A Single Component Adaptive Metropolis sampler.

get_cov(chain=None)

Return covariance SVD decomposition.

get_direction(x, U, S)

Design the direction in which we move.

slice(x, chain=None)

Return actual sample.

update_cov(sub_chain)

Updates covariance of a subset of chain

class m3c2.proposal.HyperModelWhat(names, nindex=0, bins=[0, 0.5, 1], model_names=[], update_step=1000, choose_model=<built-in method random of numpy.random.mtrand.RandomState object>, **kwargs)

A proposal handling hyper models, where one parameter controls model selection.

HMSCAM(x, chain=None, **kwargs)

Return actual sample by jumping in one component direction.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

HMslice(x, chain=None, **kwargs)

Return actual sample by jumping in one component direction. Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

check_update(chain, force=False)

Check if update of cov is needed, and do it.

model_selection()

Random model selection

class m3c2.proposal.ParSplitProposal(*args, **kwargs)

Parameter splitting (and multi trial) proposal.

Params int ntrials:

number of fast trials. Default 1000.

Parameters:

• slowp (function) –

  proposition for slow pars. Signature slowp(old_slow, chain=None, chains=None)

  returns: new_point, <ignored>, <ignored>, proposal_prob

  it will be accessible via the proposal slow method call.

• fastp (function) –

  proposition for fast pars. Signature fastp(old_fast, chain=None, chains=None, **extra)

  returns: new_point, <ignored>, <ignored>, <ignored>

  where extra are the parameters returned by the slow_fun in ChainMultiTrial. it is wrapped by the proposal fast method to generate ntrial fast points.

• fastlq (function) –

  proposition probability for fast pars. Signature fastlq(old_fast, chain=None, chains=None, **extra)

  returns the proposal probability (float)

  where extra are the parameters returned by the slow_fun in ChainMultiTrial. it will be accessible via the proposal lq method call.

fast(x, chain=None, chains=None, **extra)

Returns ntrial fast points.

class m3c2.proposal.Prior(names, prior, **kwargs)

A simple sampler based on priors.

sample_prior(x, chain=None, **kwargs)

Return actual sample.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

class m3c2.proposal.Proposal(names, logger=None, sort_name=None, sort_range=None, **kwargs)

A proposal function. We use a class to stored meta-data and ease the customization.

get_cov(chain)

Return covariance SVD decomposition.

class m3c2.proposal.ReMHA(names, **kwargs)

A Regional Metropolis Hastings Algorithm sampler.

ReMHA(x, chain=None, **kwargs)

Return actual sample by jumping in one component direction.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

class m3c2.proposal.SCAM(names, get_cov=None, **kwargs)

A Single Component Adaptive Metropolis sampler.

SCAM(x, chain=None, **kwargs)

Return actual sample by jumping in one component direction.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

class m3c2.proposal.Slice(names, get_cov=None, get_dir=None, **kwargs)

A slice sampler.

Any proposed points is accepted here.

get_direction_(x, U, S)

Design the direction in which we move.

slice(x, chain=None, **kwargs)

Return actual sample.

Args: x: the current position chain: all additional runtime data needed to propose a new point.

Returns: y: the new position status: 0 (do MH decision), 1 (accept) or -1 (reject) logy: the log-likelihood at y position if already computed. qxy: 0 for symmetric proposal distribution

m3c2.proposal.init_logger()

Default logger is stdout.