mindspore.ops.LayerNorm

class mindspore.ops.LayerNorm(begin_norm_axis=1, begin_params_axis=1, epsilon=1e-7)

Applies the Layer Normalization to the input tensor.

This operator will normalize the input tensor on given axis. LayerNorm is described in the paper Layer Normalization.

$$y=\frac{x-mean}{\sqrt{variance+ϵ}}\ast \gamma +\beta$$

where $\gamma$ is scale, $\beta$ is bias, $ϵ$ is epsilon.

Parameters

• begin_norm_axis (int) – The begin axis of the input_x to apply LayerNorm, the value must be in [-1, rank(input_x)). Default: 1 .

• begin_params_axis (int) – The begin axis of the parameter input (gamma, beta) to apply LayerNorm, the value must be in [-1, rank(input_x)). Default: 1 . Note: On the Ascend platform, the value of begin_params_axis needs to be equal to the value of begin_norm_axis .

• epsilon (float) – A value added to the denominator for numerical stability($ϵ$). Default: 1e-7 .

Inputs:

• input_x (Tensor) - Tensor of shape $(N,\dots )$. The input of LayerNorm. Supported dtypes: float16, float32, float64.

• gamma (Tensor) - Learnable parameter $\gamma$ . Tensor of shape input_x_shape[begin_params_axis:]. Supported dtypes: float16, float32, float64.

• beta (Tensor) - Learnable parameter $\beta$ . Tensor of shape input_x_shape[begin_params_axis:]. Supported dtypes: float16, float32, float64.

Outputs:

tuple[Tensor], tuple of 3 tensors, the normalized input and the updated parameters.

• output_x (Tensor) - The normalized input, has the same type and shape as the input_x.

• mean (Tensor) - The first begin_norm_axis dimensions of mean shape is the same as input_x, and the remaining dimensions are 1. Suppose the shape of the input_x is $(x_1,x_2,\dots ,x_R)$, the shape of the mean is $(x_1,\dots ,x_{begin\mathrm{\_}norm\mathrm{\_}axis},1,\dots ,1)$ (when begin_norm_axis=0, the shape of mean is $(1,\dots ,1)$ ).

• rstd (Tensor) - The reciprocal of the input standard deviation. Shape is the same as mean .

Raises

• TypeError – If begin_norm_axis or begin_params_axis is not an int.

• TypeError – If epsilon is not a float.

• TypeError – If input_x, gamma or beta is not a Tensor.

Supported Platforms:

Ascend GPU CPU

Examples

>>> import mindspore
>>> import numpy as np
>>> from mindspore import Tensor, ops
>>> input_x = Tensor(np.array([[1, 2, 3], [1, 2, 3]]), mindspore.float32)
>>> gamma = Tensor(np.ones([3]), mindspore.float32)
>>> beta = Tensor(np.ones([3]), mindspore.float32)
>>> layer_norm = ops.LayerNorm()
>>> output, _, _ = layer_norm(input_x, gamma, beta)
>>> print(output)
[[-0.2247448  1.         2.2247448]
 [-0.2247448  1.         2.2247448]]