src.model.sinusoidal_embedding module

class src.model.sinusoidal_embedding.SinusoidalEmbedding(*args: Any, **kwargs: Any)

Bases: Module

Sinusoidal embedding generator.

Commonly used in diffusion models to map discrete timesteps into high-dimensional continuous feature vectors. Includes a sinusoidal encoding stage followed by an MLP projection.

__init__(embed_dim)

Parameters:

embed_dim (int) – Output embedding dimension.

forward(t: torch.LongTensor)

Parameters:

t (torch.LongTensor) – Timestep indices. Shape: (batch_size, )

Returns:

Timestep embeddings.

Shape: (batch_size, embed_dim)

Return type:

torch.Tensor