Python v0.3.0

LayerNorm

class dragon.vm.torch.nn.LayerNorm(
  normalized_shape,
  eps=1e-05,
  elementwise_affine=True
)[source]

Apply the layer normalization. [Ba et.al, 2016]

The normalization is defined as:

\[y = \frac{x - \mathrm{E}[x]} {\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta \]

Examples:

x = torch.randn(2, 3, 4)
m = torch.nn.LayerNorm(x.size()[1:])
y = m(x)

See also

torch.nn.functional.layer_norm(…)

__init__

LayerNorm.__init__(
  normalized_shape,
  eps=1e-05,
  elementwise_affine=True
)[source]

Create a LayerNorm module.

Parameters:
  • normalized_shape (Union[int, Sequence[int]]) – The size normalized over the last dimensions.
  • eps (float, optional, default=1e-5) – The value to \(\epsilon\).
  • elementwise_affine (bool, optional, default=True) – True to apply an affine transformation.