ML Community Day is November 9! Join us for updates from TensorFlow, JAX, and more Learn more

tfp.experimental.mcmc.SequentialMonteCarlo

Sequential Monte Carlo transition kernel.

Inherits From: TransitionKernel

Sequential Monte Carlo maintains a population of weighted particles representing samples from a sequence of target distributions. It is not a calibrated MCMC kernel: the transitions step through a sequence of target distributions, rather than trying to maintain a stationary distribution.

propose_and_update_log_weights_fn Python callable with signature new_weighted_particles = propose_and_update_log_weights_fn(step, weighted_particles, seed=None). Its input is a tfp.experimental.mcmc.WeightedParticles structure representing weighted samples (with normalized weights) from the stepth target distribution, and it returns another such structure representing unnormalized weighted samples from the next (step + 1th) target distribution. This will typically include particles sampled from a proposal distribution q(x[step + 1] | x[step]), and weights that account for some or all of: the proposal density, a transition density p(x[step + 1] | x[step]), observation weightsp(y[step + 1] | x[step + 1]), and/or a backwards or 'L'-kernelL(x[step] | x[step + 1]). The (log) normalization constant of the weights is interpreted as the incremental (log) marginal likelihood. </td> </tr><tr> <td>resample_fn</td> <td> Resampling scheme specified as acallablewith signatureindices = resample_fn(log_probs, event_size, sample_shape, seed), wherelog_probsis aTensorof the same shape asstate.log_weightscontaining a normalized log-probability for every current particle,event_sizeis the number of new particle indices to generate,sample_shapeis the number of independent index sets to return, and the return valueindicesis anintTensor of shapeconcat([sample_shape, [event_size, B1, ..., BN]). Typically one of <a href="../../../tfp/experimental/mcmc/resample_deterministic_minimum_error"><code>tfp.experimental.mcmc.resample_deterministic_minimum_error</code></a>, <a href="../../../tfp/experimental/mcmc/resample_independent"><code>tfp.experimental.mcmc.resample_independent</code></a>, <a href="../../../tfp/experimental/mcmc/resample_stratified"><code>tfp.experimental.mcmc.resample_stratified</code></a>, or <a href="../../../tfp/experimental/mcmc/resample_systematic"><code>tfp.experimental.mcmc.resample_systematic</code></a>. Default value: <a href="../../../tfp/experimental/mcmc/resample_systematic"><code>tfp.experimental.mcmc.resample_systematic</code></a>. </td> </tr><tr> <td>resample_criterion_fn</td> <td> optional Pythoncallablewith signaturedo_resample = resample_criterion_fn(weighted_particles), passed an instance of <a href="../../../tfp/experimental/mcmc/WeightedParticles"><code>tfp.experimental.mcmc.WeightedParticles</code></a>. The return valuedo_resampledetermines whether particles are resampled at the current step. The default behavior is to resample particles when the effective sample size falls below half of the total number of particles. Default value: <a href="../../../tfp/experimental/mcmc/ess_below_threshold"><code>tfp.experimental.mcmc.ess_below_threshold</code></a>. </td> </tr><tr> <td>unbiased_gradients</td> <td> IfTrue, use the stop-gradient resampling trick of Scibior, Masrani, and Wood [{scibor_ref_idx}] to correct for gradient bias introduced by the discrete resampling step. This will generally increase the variance of stochastic gradients. Default value:True. </td> </tr><tr> <td>name</td> <td> Pythonstr` name for ops created by this kernel.

experimental_shard_axis_names The shard axis names for members of the state.
is_calibrated Returns True if Markov chain converges to specified distribution.

TransitionKernels which are "uncalibrated" are often calibrated by composing them with the tfp.mcmc.MetropolisHastings TransitionKernel.

name

propose_and_update_log_weights_fn

resample_criterion_fn

resample_fn

unbiased_gradients

Methods

bootstrap_results

View source

Returns an object with the same type as returned by one_step(...)[1].

Args
init_state Tensor or Python list of Tensors representing the initial state(s) of the Markov chain(s).

Returns
kernel_results A (possibly nested) tuple, namedtuple or list of Tensors representing internal calculations made within this function.

copy

View source

Non-destructively creates a deep copy of the kernel.

Args
**override_parameter_kwargs Python String/value dictionary of initialization arguments to override with new values.

Returns
new_kernel TransitionKernel object of same type as self, initialized with the union of self.parameters and override_parameter_kwargs, with any shared keys overridden by the value of override_parameter_kwargs, i.e., dict(self.parameters, **override_parameters_kwargs).

experimental_with_shard_axes

View source

Returns a copy of the kernel with the provided shard axis names.

Args
shard_axis_names a structure of strings indicating the shard axis names for each component of this kernel's state.

Returns
A copy of the current kernel with the shard axis information.

one_step

View source

Takes one Sequential Monte Carlo inference step.

Args
state instance of tfp.experimental.mcmc.WeightedParticles representing the current particles with (log) weights. The log_weights must be a float Tensor of shape [num_particles, b1, ..., bN]. The particles may be any structure of Tensors, each of which must have shape concat([log_weights.shape, event_shape]) for some event_shape, which may vary across components.
kernel_results instance of tfp.experimental.mcmc.SequentialMonteCarloResults representing results from a previous step.
seed PRNG seed; see tfp.random.sanitize_seed for details.

Returns
state instance of tfp.experimental.mcmc.WeightedParticles representing new particles with (log) weights.
kernel_results instance of tfp.experimental.mcmc.SequentialMonteCarloResults.