mindspore.mint.nn.functional.gelu

mindspore.mint.nn.functional.gelu(input, *, approximate='none') → Tensor

Gaussian Error Linear Units activation function.

GeLU is described in the paper Gaussian Error Linear Units (GELUs). And also please refer to BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.

When approximate argument is none, GELU is defined as follows:

$$GELU(x_i)=x_i\ast P(X<x_i),$$

where $P$ is the cumulative distribution function of the standard Gaussian distribution, $x_i$ is the input element.

When approximate argument is tanh, GELU is estimated with:

$$GELU(x_i)=0.5\ast x_i\ast (1+\tanh (\sqrt{(}2/\pi )\ast (x_i+0.044715\ast x_i^3)))$$

GELU Activation Function Graph:

Note

On the Ascend platform, when input is -inf, its gradient is 0, and when input is inf, its gradient is dout.

Parameters

input (Tensor) – The input of the activation function GeLU, the data type is float16, float32 or float64.

Keyword Arguments

approximate (str, optional) – the gelu approximation algorithm to use. Acceptable vaslues are 'none' and 'tanh' . Default: 'none' .

Returns

Tensor, with the same type and shape as input.

Raises

• TypeError – If input is not a Tensor.

• TypeError – If dtype of input is not bfloat16, float16, float32 or float64.

• ValueError – If approximate value is neither none nor tanh.

Supported Platforms:

Ascend

Examples

>>> import mindspore
>>> import numpy as np
>>> from mindspore import Tensor, mint
>>> input = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]]), mindspore.float32)
>>> result = mint.nn.functional.gelu(input)
>>> print(result)
[[-1.58655241e-01  3.99987316e+00 -0.00000000e+00]
 [ 1.95449972e+00 -1.41860323e-06  9.0000000e+00]]
>>> result = mint.nn.functional.gelu(input, approximate="tanh")
>>> print(result)
[[-1.58808023e-01  3.99992990e+00 -3.10779147e-21]
 [ 1.95459759e+00 -2.29180174e-07  9.0000000e+00]]