LearnableInversePositionalEncoding

class rectools.models.nn.transformers.net_blocks.LearnableInversePositionalEncoding(use_pos_emb: bool, session_max_len: int, n_factors: int, use_scale_factor: bool = False, **kwargs: Any)

Bases: PositionalEncodingBase

Class to introduce learnable positional embeddings.

Parameters

• use_pos_emb (bool) – If True, learnable positional encoding will be added to session item embeddings.

• session_max_len (int) – Maximum length of user sequence.

• n_factors (int) – Latent embeddings size.

• use_scale_factor (int) – Use multiplication embedding on the root of the dimension embedding

• kwargs (Any) –

Methods

forward(sessions)

Forward pass to add learnable positional encoding to sessions and mask padding elements.

Attributes

forward(sessions: Tensor) → Tensor

Forward pass to add learnable positional encoding to sessions and mask padding elements.

Parameters

sessions (torch.Tensor) – User sessions in the form of sequences of items ids.

Returns

Encoded user sessions with added positional encoding if use_pos_emb is True.

Return type

torch.Tensor