BoxTransformerModule

Box Transformer architecture combining a multiple BoxTransformerLayer modules. It is mainly used in BoxTransformer.

Parameters

PARAMETER	DESCRIPTION
input_size	Input embedding size TYPE: Optional[int] DEFAULT: None
num_heads	Number of attention heads in the attention layers TYPE: int DEFAULT: 2
n_relative_positions	Maximum range of embeddable relative positions between boxes (further distances are capped to ±n_relative_positions // 2) TYPE: Optional[int] DEFAULT: None
dropout_p	Dropout probability both for the attention layers and embedding projections TYPE: float DEFAULT: 0.0
head_size	Head sizes of the attention layers TYPE: Optional[int] DEFAULT: None
activation	Activation function used in the linear->activation->linear transformations TYPE: ActivationFunction DEFAULT: 'gelu'
init_resweight	Initial weight of the residual gates. At 0, the layer acts (initially) as an identity function, and at 1 as a standard Transformer layer. Initializing with a value close to 0 can help the training converge. TYPE: float DEFAULT: 0.0
attention_mode	Mode of relative position infused attention layer. See the relative attention documentation for more information. TYPE: Sequence[Literal['c2c', 'c2p', 'p2c']] DEFAULT: ('c2c', 'c2p', 'p2c')
n_layers	Number of layers in the Transformer TYPE: int DEFAULT: 2

forward

Forward pass of the BoxTransformer

PARAMETER	DESCRIPTION
embeds	Embeddings to contextualize Shape: n_samples * n_keys * input_size TYPE: FoldedTensor
boxes	Layout features of the input elements TYPE: Dict

RETURNS	DESCRIPTION
Tuple[FloatTensor, List[FloatTensor]]	Output of the last BoxTransformerLayer Shape: n_samples * n_queries * n_keys Attention logits of all layers Shape: n_samples * n_queries * n_keys * n_heads

---