BoxLayoutEmbedding

This component encodes the geometrical features of a box, as extracted by the BoxLayoutPreprocessor module, into an embedding. For position modes, use:

• "sin" to embed positions with a fixed SinusoidalEmbedding
• "learned" to embed positions using a learned standard pytorch embedding layer

Each produces embedding is the concatenation of the box width, height and the top, left, bottom and right coordinates, each embedded depending on the *_mode param.

Parameters

PARAMETER	DESCRIPTION
size	Size of the output box embedding TYPE: int
n_positions	Number of position embeddings stored in the PositionEmbedding module TYPE: int
x_mode	Position embedding mode of the x coordinates TYPE: Literal['sin', 'learned'] DEFAULT: 'sin'
y_mode	Position embedding mode of the x coordinates TYPE: Literal['sin', 'learned'] DEFAULT: 'sin'
w_mode	Position embedding mode of the width features TYPE: Literal['sin', 'learned'] DEFAULT: 'sin'
h_mode	Position embedding mode of the height features TYPE: Literal['sin', 'learned'] DEFAULT: 'sin'

---