1 vin.models.target_myopic

vin.models.target_myopic

Target-conditioned myopic VIN scorer family.

This module owns the one-step architecture-family name that should score each candidate from actor-visible scene evidence plus an actor-visible target descriptor. The zero-descriptor configuration is runnable today as a named myopic baseline backed by aria_nbv.vin.models.scene_myopic.VinModelV3; nonzero target descriptors remain blocked until the actor-visible target-token contract is implemented.

1.1 Classes

Name Description
TargetConditionedMyopicScorer Runnable zero-target myopic scorer plus blocked target-conditioned scaffold.
TargetConditionedMyopicScorerConfig Config-as-factory for the one-step target-scorer family.

1.1.1 TargetConditionedMyopicScorer

vin.models.target_myopic.TargetConditionedMyopicScorer(config)

Runnable zero-target myopic scorer plus blocked target-conditioned scaffold.

With target_descriptor_dim == 0 this module delegates to VinModelV3, preserving the current CORAL VinPrediction training contract under a named myopic architecture family. Positive descriptor widths fail during construction because target-token ownership, actor-input visibility, and feature fusion are not implemented yet.

1.1.1.1 Attributes

Name Description
head_coral CORAL head delegated to the underlying v3 myopic scorer.

1.1.1.2 Methods

Name Description
init_bin_values Initialize delegated v3 CORAL bin representatives.
forward Score PoseTW["B N_q 12"] candidates through VIN v3.
summarize_vin Return the underlying v3 debug summary for the myopic baseline.
1.1.1.2.1 init_bin_values
vin.models.target_myopic.TargetConditionedMyopicScorer.init_bin_values(
    values,
    *,
    overwrite=False,
)

Initialize delegated v3 CORAL bin representatives.

1.1.1.2.2 forward
vin.models.target_myopic.TargetConditionedMyopicScorer.forward(
    efm,
    *,
    candidate_poses_world_cam,
    reference_pose_world_rig,
    p3d_cameras,
    backbone_out=None,
)

Score PoseTW["B N_q 12"] candidates through VIN v3.

With the only runnable zero-descriptor configuration, this returns the same B/N_q-aligned CORAL VinPrediction as VinModelV3; it does not consume a target token.

1.1.1.2.3 summarize_vin
vin.models.target_myopic.TargetConditionedMyopicScorer.summarize_vin(
    batch,
    *,
    include_torchsummary=True,
    torchsummary_depth=3,
)

Return the underlying v3 debug summary for the myopic baseline.

1.1.2 TargetConditionedMyopicScorerConfig

vin.models.target_myopic.TargetConditionedMyopicScorerConfig()

Config-as-factory for the one-step target-scorer family.

1.1.2.1 Attributes

Name Type Description
num_classes int Number of ordinal output classes for CORAL-style row scoring if the scorer uses the current VIN Lightning objective.
target_descriptor_dim int Dimension of the actor-visible target token or descriptor that conditions candidate scoring. 0 selects the runnable v3-backed myopic baseline.
candidate_token_dim int Internal candidate token width reserved for the first target-conditioned implementation.
base_scorer VinModelV3Config Architecture config for the v3-backed zero-descriptor myopic baseline. The top-level num_classes field remains authoritative for Lightning/binner compatibility.

Additional attribute identifiers: target_type

Back to top