beanmachine.ppl.inference.proposer.hmc_utils module

class beanmachine.ppl.inference.proposer.hmc_utils.DictTransform(transforms: Dict[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.distributions.transforms.Transform])

Bases: object

A general class for applying a dictionary of Transforms to a dictionary of Tensors


transforms – Dict of torch.distributions.Transform keyed by the RVIdentifier

inv(node_vals: Dict[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]) Dict[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]

Apply the inverse of each Transform to the corresponding Tensor in node_vals

log_abs_det_jacobian(untransformed_vals: Dict[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor], transformed_vals: Dict[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]) torch.Tensor

Computes the sum of log det jacobian log |dy/dx| on the pairs of Tensors

class beanmachine.ppl.inference.proposer.hmc_utils.DualAverageAdapter(initial_epsilon: torch.Tensor, delta: float = 0.8)

Bases: object

The dual averaging mechanism that’s introduced in [1] and was applied to HMC and NUTS for adapting step size in [2]. The implementation and notations follows [2].

[1] Yurii Nesterov. “Primal-dual subgradient methods for convex problems” (2009).

[2] Matthew Hoffman and Andrew Gelman. “The No-U-Turn Sampler: Adaptively

Setting Path Lengths in Hamiltonian Monte Carlo” (2014).

finalize() torch.Tensor
step(alpha: torch.Tensor) torch.Tensor
class beanmachine.ppl.inference.proposer.hmc_utils.MassMatrixAdapter(initial_positions: torch.Tensor, full_mass_matrix: bool = False)

Bases: object

Adapts the mass matrix. The (inverse) mass matrix is initialized to identity and will be updated during adaptation windows.

  • matrix_size – The size of the mass matrix. This value should be the same

  • tensor. (as the length of the flattened position) –


[1] “HMC algorithm parameters” from Stan Reference Manual

finalize() None
initialize_momentums(positions: torch.Tensor) torch.Tensor

Randomly draw momentum from MultivariateNormal(0, M). This momentum variable is denoted as p in [1] and r in [2].


positions – the positions of the energy function.

step(positions: torch.Tensor)
class beanmachine.ppl.inference.proposer.hmc_utils.RealSpaceTransform(world:, target_rvs: Set[beanmachine.ppl.model.rv_identifier.RVIdentifier])

Bases: beanmachine.ppl.inference.proposer.hmc_utils.DictTransform

Transform a dictionary of Tensor values from a constrained space to the unconstrained (real) space.

  • world – World which contains the random variables of interest.

  • target_rvs – Set of RVIdentifiers corresponding to the random variables of interest.

class beanmachine.ppl.inference.proposer.hmc_utils.WelfordCovariance(diagonal: bool = True)

Bases: object

An implementation of Welford’s online algorithm for estimating the (co)variance of samples.

[1] “Algorithms for calculating variance” on Wikipedia’s_online_algorithm

finalize(regularize: bool = True) torch.Tensor
step(sample: torch.Tensor) None
class beanmachine.ppl.inference.proposer.hmc_utils.WindowScheme(num_adaptive_samples: int)

Bases: object

Spliting adaptation iterations into a series of monotonically increasing windows, which can be used to learn the mass matrices in HMC.

[1] “HMC algorithm parameters” from Stan Reference Manual

property is_end_window
property is_in_window