Norms#

Collection of Ivy normalization functions.

ivy.layer_norm(x, normalized_idxs, /, *, scale=None, offset=None, eps=1e-05, new_std=1.0, out=None)[source]#

Apply Layer Normalization over a mini-batch of inputs.

Parameters:
  • x (Union[Array, NativeArray]) – Input array

  • normalized_idxs (List[int]) – Indices to apply the normalization to.

  • scale (Optional[Union[Array, NativeArray]], default: None) – Learnable gamma variables for elementwise post-multiplication, default is None.

  • offset (Optional[Union[Array, NativeArray]], default: None) – Learnable beta variables for elementwise post-addition, default is None.

  • eps (float, default: 1e-05) – small constant to add to the denominator. Default is 1e-05

  • new_std (float, default: 1.0) – The standard deviation of the new normalized values. Default is 1.

  • out (Optional[Array], default: None) – optional output array, for writing the result to. It must have a shape that the inputs broadcast to.

Return type:

Array

Returns:

ret

The layer after applying layer normalization.

Examples

With ivy.Array input: >>> x = ivy.array([[1.0, 2.0], [3.0, 4.0]]) >>> y = ivy.layer_norm(x, [0, 1], new_std=2.0) >>> print(y) ivy.array([[-2.68 , -0.894],

[ 0.894, 2.68 ]])

>>> x = ivy.array([[1., 2., 3.], [4., 5., 6.]])
>>> y = ivy.zeros((2, 3))
>>> ivy.layer_norm(x, [0], out=y)
>>> print(y)
ivy.array([[-1., -1., -1.],
           [ 1.,  1.,  1.]])
>>> x = ivy.array([[0.0976, -0.3452,  1.2740],
...                [0.1047,  0.5886,  1.2732],
...                [0.7696, -1.7024, -2.2518]])
>>> y = ivy.layer_norm(x, [0, 1], eps=0.001,
...                       new_std=1.5, scale=0.5, offset=[0.5, 0.02, 0.1])
>>> print(y)
ivy.array([[ 0.826, -0.178, 0.981 ],
           [ 0.831,  0.421, 0.981 ],
           [ 1.26 , -1.05 , -1.28 ]])
With a mix of :class:`ivy.Array` and :class:`ivy.Container` inputs:
>>> x = ivy.array([[1., 2., 3.], [4., 5., 6.]])
>>> normalized_idxs = ivy.Container({'a': [0], 'b': [1]})
>>> y = ivy.layer_norm(x, normalized_idxs, new_std=1.25, offset=0.2)
>>> print(y)
{
    a: ivy.array([[-1.25, -1.25, -1.25],
                  [1.25, 1.25, 1.25]]),
    b: ivy.array([[-1.53, 0., 1.53],
                  [-1.53, 0., 1.53]])
}
With one :class:`ivy.Container` input:
>>> x = ivy.Container({'a': ivy.array([7., 10., 12.]),
...                    'b': ivy.array([[1., 2., 3.], [4., 5., 6.]])})
>>> normalized_idxs = [0]
>>> y = ivy.layer_norm(x, normalized_idxs, eps=1.25, scale=0.3)
>>> print(y)
{
    a: ivy.array([-0.34198591, 0.04274819, 0.29923761]),
    b: ivy.array([[-0.24053511, -0.24053511, -0.24053511],
                  [0.24053511, 0.24053511, 0.24053511]])
}

With multiple ivy.Container inputs:

>>> x = ivy.Container(a=ivy.array([7.0, 10.0, 12.0]),
...                   b=ivy.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]))
>>> normalized_idxs = ivy.Container(a=[0], b=[1])
>>> new_std = ivy.Container(a=1.25, b=1.5)
>>> bias = ivy.Container(a=[0.2, 0.5, 0.7], b=0.3)
>>> y = ivy.layer_norm(x, normalized_idxs, new_std=new_std, offset=0.2)
>>> print(y)
{
    a: ivy.array([-1.62, 0.203, 1.42]),
    b: ivy.array([[-1.84, 0., 1.84],
                  [-1.84, 0., 1.84]])
}
# Both the description and the type hints above assumes an array input for
simplicity, but this function is *nestable*, and therefore also accepts
:class:`ivy.Container` instances in place of any of the arguments.

This should have hopefully given you an overview of the norms submodule, if you have any questions, please feel free to reach out on our discord in the norms channel!