Method forward

modules.py:616–646

r"""Apply positional encoding to the input. Args: tensor (torch.Tensor): Input tensor to be positionally encoded. encoding_size (optional, int): Number of encoding functions used to compute a positional encoding (default: 6). include_input

(self, tensor)

                self.normalization = torch.tensor(1/self.frequency_bands)

    def forward(self, tensor) -> torch.Tensor:
        r"""Apply positional encoding to the input.

        Args:
            tensor (torch.Tensor): Input tensor to be positionally encoded.
            encoding_size (optional, int): Number of encoding functions used to compute
                a positional encoding (default: 6).
            include_input (optional, bool): Whether or not to include the input in the
                positional encoding (default: True).

        Returns:
        (torch.Tensor): Positional encoding of the input tensor.
        """

        encoding = [tensor] if self.include_input else []
        if self.gaussian_pe:
            for func in [torch.sin, torch.cos]:
                encoding.append(func(torch.matmul(tensor, self.gaussian_weights.T)))
        else:
            for idx, freq in enumerate(self.frequency_bands):
                for func in [torch.sin, torch.cos]:
                    if self.normalization is not None:
                        encoding.append(self.normalization[idx]*func(tensor * freq))
                    else:
                        encoding.append(func(tensor * freq))

        # Special case, for no positional encoding
        if len(encoding) == 1:
            return encoding[0]
        else:
            return torch.cat(encoding, dim=-1)


def layer_factory(layer_type, w0=30):