neuralop.layers.embeddings.SinusoidalEmbedding

class neuralop.layers.embeddings.SinusoidalEmbedding(in_channels: int, num_frequencies: int | None = None, embedding_type: str = 'transformer', max_positions: int = 10000)[source]

SinusoidalEmbedding provides a unified sinusoidal positional embedding in the styles of Transformers George, R., Zhao, J., Kossaifi, J., Li, Z., and Anandkumar, A. (2024) and Neural Radiance Fields (NERFs) Mildenhall, B. et al (2020).

Parameters:

in_channelsint

Number of input channels to embed

num_freqsint, optional

Number of frequencies in positional embedding. By default, set to the number of input channels

embedding{‘transformer’, ‘nerf’}

Type of embedding to apply. For a function with N input channels, each channel value p is embedded via a function g with 2L channels such that g(p) is a 2L-dim vector. For 0 <= k < L:

• ‘transformer’ for transformer-style encoding.

  g(p)_k = sin((p / max_positions) ^ {k / N})

  g(p)_{k+1} = cos((p / max_positions) ^ {k / N})

• ‘nerf’ : NERF-style encoding.

  g(p)_k = sin(2^(k) * Pi * p)

  g(p)_{k+1} = cos(2^(k) * Pi * p)

max_positionsint, optional

Maximum number of positions for the encoding, default 10000 Only used if embedding == transformer.

Attributes:

out_channels

out_channels: required property for linking/composing model layers

Methods

forward(x)

References

Vaswani, A. et al (2017)

“Attention Is All You Need”. NeurIPS 2017, https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf.

Mildenhall, B. et al (2020)

“NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis”. ArXiv, https://arxiv.org/pdf/2003.08934.

property out_channels

out_channels: required property for linking/composing model layers

forward(x)[source]

Parameters:

x: torch.Tensor, shape (n_in, self.in_channels) or (batch, n_in, self.in_channels)