Skip to content

config

Pydantic configuration models for Jim's samplers.

Integer counts require Python integers; booleans, floats and strings are rejected. Each sampler has its own *Config class discriminated by a type literal; SamplerConfig is the discriminated-union annotation a caller passes to Jim(..., sampler_config=...).

Attributes:

Name Type Description
FlowMCLocalKernelConfig

Discriminated union of every flowMC local-kernel config, by its own

SMCKernelConfig

Discriminated union of the BlackJAX SMC mutation kernels, by their own

SamplerConfig

Discriminated union of every concrete sampler config.

FlowMCLocalKernelConfig = Annotated[Union[FlowMCMALAConfig, FlowMCHMCConfig, FlowMCGRWConfig], Discriminator('kernel')] module-attribute ¤

Discriminated union of every flowMC local-kernel config, by its own kernel field -- the same pattern SamplerConfig itself uses (there via type), so the active kernel's settings live in exactly one place with no separate selector field to fall out of sync with it.

SMCKernelConfig = Annotated[Union[SMCTPCNConfig, SMCGRWConfig], Discriminator('kernel')] module-attribute ¤

Discriminated union of the BlackJAX SMC mutation kernels, by their own kernel field, as FlowMCLocalKernelConfig is for flowMC.

SamplerConfig = Annotated[Union[FlowMCConfig, BlackJAXNSAWConfig, BlackJAXNSSConfig, BlackJAXSwiGConfig, BlackJAXSMCConfig], Discriminator('type')] module-attribute ¤

Discriminated union of every concrete sampler config.

BaseSamplerConfig ¤

Bases: BaseModel

Fields shared by all sampler configs.

BlackJAXNSAWConfig ¤

Bases: BaseSamplerConfig[Literal['blackjax-ns-aw']], _CheckpointMixin, _LiveSetConfigMixin

Configuration for the BlackJAX acceptance-walk nested sampler.

Note

This sampler requires the sampling space to be the unit hypercube [0, 1]^n_dims. When using Jim, this means all sample_transforms must map the prior support onto the unit cube.

Note

Periodic parameters are not configured here. Pass a periodic argument to Jim instead. For NS AW, bounds are implicit as [0, 1]; just list the parameter names.

BlackJAXNSSConfig ¤

Bases: BaseSamplerConfig[Literal['blackjax-nss']], _CheckpointMixin, _LiveSetConfigMixin, _ShardingMixin

Configuration for the BlackJAX nested slice sampler.

Note

Periodic parameters are not configured here. Pass a periodic argument to Jim instead.

BlackJAXSMCConfig ¤

Bases: BaseSamplerConfig[Literal['blackjax-smc']], _CheckpointMixin

Configuration for the BlackJAX SMC sampler.

mcmc_kernel selects the kernel that moves the particles at each temperature: an SMCTPCNConfig (default) or an SMCGRWConfig, an equivalent dict with a "kernel" key, or the bare kernel name for its defaults (mcmc_kernel="GRW"). flow moves the particles in the latent space of a normalizing flow, with either kernel: an SMCFlowConfig (default), a dict of its fields, True for the defaults, or False/None for no flow.

Without target_ess or target_ess_fraction, the ESS target is 3.0 * n_particles with persistent sampling and 0.9 * n_particles without.

Parameters¤

batch_size : int, optional Number of particles to process per sequential batch during the MCMC update step. When batch_size > 0, the sampler uses jax.lax.map instead of jax.vmap, which reduces peak GPU memory at the cost of sequential execution. 0 (default) uses the original full jax.vmap behaviour.

Note

Periodic parameters are not configured here. Pass a periodic argument to Jim instead.

BlackJAXSwiGConfig ¤

Bases: BaseSamplerConfig[Literal['blackjax-swig']], _CheckpointMixin, _LiveSetConfigMixin, _ShardingMixin

Configuration for Nested Slice within Gibbs (SwiG) sampling.

blocks are expressed in sampling-space parameter names. A supported likelihood validates that they form an exact partition and identifies which blocks refresh its cache.

FlowMCConfig ¤

Bases: BaseSamplerConfig[Literal['flowmc']], _CheckpointMixin

Configuration for FlowMCSampler.

The local_kernel field selects the MCMC kernel used for local proposals -- pass a FlowMCMALAConfig (default), FlowMCHMCConfig, or FlowMCGRWConfig instance (or an equivalent dict with a "kernel" key, or just the bare kernel name as a string for its defaults, e.g. local_kernel="HMC"):

  • FlowMCMALAConfig — Metropolis-Adjusted Langevin; default.
  • FlowMCHMCConfig — Hamiltonian Monte Carlo.
  • FlowMCGRWConfig — Gaussian random walk.

Parallel tempering is off by default. To enable, pass a FlowMCParallelTemperingConfig, a dict of its fields, or simply True (uses all defaults). False disables it.

Note

Periodic parameters are not configured here. Pass a periodic argument to Jim instead; Jim resolves parameter names to dimension indices and passes them to the sampler.

FlowMCGRWConfig ¤

Bases: BaseModel

Gaussian random-walk local-kernel settings for the flowMC backend.

step_size may be a scalar float (applied uniformly to all dimensions) or a 1-D np.ndarray of length n_dims for per-dimension step sizes.

FlowMCHMCConfig ¤

Bases: BaseModel

HMC local-kernel settings for the flowMC backend.

step_size may be a scalar float (applied uniformly) or a 1-D np.ndarray of length n_dims for per-dimension sizes. condition_matrix may also be a scalar or 1-D array and is used as the mass matrix / preconditioning for HMC leapfrog steps.

FlowMCMALAConfig ¤

Bases: BaseModel

MALA local-kernel settings for the flowMC backend.

step_size may be a scalar float (applied uniformly to all dimensions) or a 1-D np.ndarray of length n_dims for per-dimension step sizes.

FlowMCParallelTemperingConfig ¤

Bases: BaseModel

Parallel-tempering settings for the flowMC backend.

Construct directly or pass a plain dict to FlowMCConfig.parallel_tempering. Use True to enable with all defaults, False / None to disable.

SMCFlowConfig ¤

Bases: BaseModel

Normalizing-flow settings for the BlackJAX SMC backend.

At every mutation a newly initialised masked-coupling rational-quadratic spline flow trains on the particles for n_epochs, and the kernel moves the particles in the flow's latent space.

SMCGRWConfig ¤

Bases: BaseModel

Gaussian random-walk kernel settings for the BlackJAX SMC backend.

Each mutation runs n_mcmc_steps_per_dim * n_dims steps with proposal covariance s * Cov(particles). After it, s is multiplied by exp(scale_adaptation_gain * (acceptance - target_acceptance_rate)); initial_cov_scale is the first s.

SMCTPCNConfig ¤

Bases: BaseModel

t-pCN kernel settings for the BlackJAX SMC backend.

A mutation stops once the particles' mean log posterior has not improved for n_patience_per_dim * n_dims steps, a count that grows as the step size shrinks, or after n_max_per_dim * n_dims steps. The step size adapts towards target_acceptance_rate.