sample_input_masks

bayesflow.utils.sample_input_masks(subnet: Layer, x: Tensor, conditions: Tensor | None, subnet_kwargs: dict, training: bool, fixed_target_prob: float, missing_target_prob: float, missing_conditions_prob: float, seed_generator: SeedGenerator = None) → tuple[Tensor | float, Tensor | float, dict]

Generate the target and condition masks and populate subnet_kwargs.

The diffusion type inference networks (diffusion, flow matching, consistency) share the same input-masking scheme: fixed_target_prob randomly fixes some targets to their known value (so the network can condition on them), missing_target_prob randomly marks fixed targets as missing, and missing_conditions_prob randomly marks conditions as missing, so the network learns to handle missing fields. The masks are forwarded to subnets that accept them (e.g. diffusion_transformer).

Returns:

mask_xTensor or float

The per-target inference mask (1 = inferred/noised, 0 = fixed).

loss_maskTensor or float

The mask the caller applies to the loss.

subnet_kwargsdict

The keyword arguments, with mask entries added for subnets that accept them.