blackjax.adaptation.laps_burn_in

blackjax.adaptation.laps_burn_in#

Classes#

`History`
`AdaptationState`
`Adaptation`

Functions#

`no_nans`(a)
`nan_reject`(nonans, old, new)	Equivalent to
`build_kernel`(logdensity_fn, ndims[, microcanonical])	MCLMC kernel (with nan rejection)
`initialize`(rng_key, logdensity_fn, microcanonical, ...)	initialize the chains based on the equipartition of the initial condition.
`update_history`(new_vals, history)
`update_history_scalar`(new_val, history)
`contract_history`(theta, weights)
`equipartition_diagonal`(state)	Ei = E_ensemble (- grad log p_i x_i ). Ei is 1 if we have converged.
`equipartition_fullrank`(state, rng_key)	loss = Tr[(1 - E)^T (1 - E)] / d^2
`equipartition_diagonal_loss`(Eii)
`equipartition_fullrank_loss`(delta_z)

Module Contents#

no_nans(a)[source]#

nan_reject(nonans, old, new)[source]#: Equivalent to return new if nonans else old

build_kernel(logdensity_fn, ndims, microcanonical=True)[source]#: MCLMC kernel (with nan rejection)

initialize(rng_key, logdensity_fn, microcanonical, sample_init, num_chains, mesh, superchain_size)[source]#: initialize the chains based on the equipartition of the initial condition. We initialize the velocity along grad log p if E_ii > 1 and along -grad log p if E_ii < 1.

update_history(new_vals, history)[source]#

update_history_scalar(new_val, history)[source]#

contract_history(theta, weights)[source]#

class History[source]#

observables: blackjax.types.Array[source]#

stopping: blackjax.types.Array[source]#

weights: blackjax.types.Array[source]#

class AdaptationState[source]#

L: float[source]#

inverse_mass_matrix: Any[source]#

step_size: float[source]#

step_count: int[source]#

EEVPD: float[source]#

EEVPD_wanted: float[source]#

history: Any[source]#

equipartition_diagonal(state)[source]#: Ei = E_ensemble (- grad log p_i x_i ). Ei is 1 if we have converged. equipartition_loss = average over parameters (Ei)

equipartition_fullrank(state, rng_key)[source]#: loss = Tr[(1 - E)^T (1 - E)] / d^2 where Eij = <xi gj> is the equipartition patrix. Loss is computed with the Hutchinson’s trick.

equipartition_diagonal_loss(Eii)[source]#

equipartition_fullrank_loss(delta_z)[source]#

class Adaptation(ndims, microcanonical, alpha=1.0, C=0.1, r_end=0.01, bias_type=0, save_num=10, observables=lambda x: ..., observables_for_bias=lambda x: ..., contract=lambda x: ...)[source]#

ndims[source]#

alpha = 1.0[source]#

C = 0.1[source]#

r_end = 0.01[source]#

observables[source]#

observables_for_bias[source]#

contract[source]#

bias_type = 0[source]#

save_num = 10[source]#

norm_factor[source]#

initial_state[source]#

summary_statistics_fn(state, info, rng_key)[source]#

update(adaptation_state, Etheta)[source]#

while_cond(info, counter)[source]#: determine if we want to switch to adjustment