constants

constants

Constants for encoder configurations, backbone types, and output keys.

SpatialBackboneType

Bases: StrEnum

Backbones producing (B, C, H, W) spatial feature maps via timm features_only.

FlatBackboneType

Bases: StrEnum

Backbones producing (B, S, D) token sequences via timm forward_features.

DinoV2SigLIPBackboneType

Bases: StrEnum

DINOv2+SigLIP paired RGB vision backbone identifiers.

ImageTextModelType

Bases: StrEnum

Available image+text multimodal encoders.

AttentionImplementation

Bases: StrEnum

Attention implementation types.

PoolingMethod

Bases: StrEnum

Feature pooling methods for spatial and flat RGB encoders.

supports_spatial property

supports_spatial

Whether this pooling method works with spatial (B, C, H, W) feature maps.

supports_sequential property

supports_sequential

Whether this pooling method works with sequential (B, S, D) token sequences.

BatchNormHandling

Bases: StrEnum

How to handle BatchNorm layers in spatial RGB backbones.

BatchNorm is problematic for temporal data: when reshaping (B,T,C,H,W) to (B*T,C,H,W), batch statistics mix frames across time, leaking future information into each frame's representation. This causes train/test mismatch since the signal vanishes at inference.

LanguageEncoderType

Bases: StrEnum

Available language encoders.

EncoderOutputKeys

Bases: StrEnum

Types of encoder output keys to use for extracting an output feature from an encoder.