fusion

fusion

Configuration for feature fusion modules.

FusionConfig dataclass

FusionConfig(_target_=MISSING, input_features=MISSING, output_name=MISSING, hidden_dimension=MISSING)

A fusion module that combines multiple input features into one representation.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
input_features	list[str]	Feature names consumed by the fusion stage.
output_name	str	Name registered for the fused feature.
hidden_dimension	int	Hidden layer width.

ConcatFusionConfig dataclass

ConcatFusionConfig(_target_='versatil.models.encoding.fusion.concat.ConcatFusion', input_features=MISSING, output_name=MISSING, hidden_dimension=MISSING)

Bases: FusionConfig

Configuration for feature concatenation fusion.

AttentionFusionConfig dataclass

AttentionFusionConfig(_target_='versatil.models.encoding.fusion.attention.AttentionFusion', input_features=MISSING, output_name=MISSING, hidden_dimension=MISSING, number_of_heads=8, dropout=0.1, input_feature_query=None)

Bases: FusionConfig

Configuration for cross-attention feature fusion.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
number_of_heads	int	Number of attention heads.
dropout	float	Dropout rate for attention weights.
input_feature_query	str | None	Name of the feature to use as query in cross-attention. If None, uses the first feature.

MLPFusionConfig dataclass

MLPFusionConfig(_target_='versatil.models.encoding.fusion.mlp.MLPFusion', input_features=MISSING, output_name=MISSING, hidden_dimension=MISSING, mlp_hidden_dims=MISSING, activation_name=value, dropout=0.1)

Bases: FusionConfig

Configuration for MLP-based feature fusion.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
mlp_hidden_dims	list[int]	List of hidden layer dimensions for the MLP.
activation_name	str	Name of the activation function to use in the MLP.
dropout	float	Dropout rate for the MLP.