starling.inference.constraints.ReConstraint

class ReConstraint

Bases: Constraint

Methods

__init__	Create constraint for end-to-end distance.
apply	Apply the constraint to the given latents.
bell_shaped_schedule	Bell-shaped schedule for time-dependent guidance strength.
compute_loss	Compute the loss for this constraint without applying gradients.
cosine_weight	Cosine schedule for time-dependent guidance strength.
get_adaptive_clip_threshold	Get an adaptive clipping threshold that follows a cosine schedule.
get_time_scale	Get the time-dependent scaling factor.
initialize	Called by the sampler to set model parameters.
should_apply_guidance	Check if guidance should be applied at the current timestep.

__init__(target, tolerance=0.0, force_constant=2.0, **kwargs)

Create constraint for end-to-end distance.

compute_loss(distance_maps: Tensor) → Tuple[Tensor, Tensor]

Compute the loss for this constraint without applying gradients.

Parameters:

distance_maps (torch.Tensor) – Pre-computed distance maps from the latents

Returns:

(per_batch_loss, total_loss) - Individual sample losses and mean loss

Return type:

tuple[torch.Tensor, torch.Tensor]