Source code for espnet.nets.pytorch_backend.transformer.layer_norm

import torch


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

    :param int nout: output dim size
    :param int dim: dimension to be normalized
    """

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

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

        :param torch.Tensor x: input tensor
        :return: layer normalized tensor
        :rtype torch.Tensor
        """
        if self.dim == -1:
            return super(LayerNorm, self).forward(x)
        return super(LayerNorm, self).forward(x.transpose(1, -1)).transpose(1, -1)