LayerNorm.py

# Written by Shigeki Karita, 2019
# Published under Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
# Adapted by Florian Lux, 2021

import torch


class LayerNorm(torch.nn.LayerNorm):
    """
    Layer normalization module.

    Args:
        nout (int): Output dim size.
        dim (int): Dimension to be normalized.
    """

    def __init__(self, nout, dim=-1):
        """
        Construct an LayerNorm object.
        """
        super(LayerNorm, self).__init__(nout, eps=1e-12)
        self.dim = dim

    def forward(self, x):
        """
        Apply layer normalization.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Normalized tensor.
        """
        if self.dim == -1:
            return super(LayerNorm, self).forward(x)
        return super(LayerNorm, self).forward(x.transpose(1, -1)).transpose(1, -1)