hard_swish¶

paddle.fluid.layers.nn. hard_swish ( x, threshold=6.0, scale=6.0, offset=3.0, name=None ) [source]

This operator implements the hard_swish activation function. Hard_swish is proposed in MobileNetV3, and performs better in computational stability and efficiency compared to swish function. For more details please refer to: https://arxiv.org/pdf/1905.02244.pdf

The formula is as follows:

\[\begin{split}out = \\frac{x * (min(max(0, x+offset), threshold))}{scale}\end{split}\]

In the above equation:

threshold and scale should be positive, offset can be positive or negative. It is recommended to use default parameters.

Parameters

x (Variable) – Input feature, multi-dimensional Tensor. The data type should be float32 or float64.
threshold (float, optional) – The threshold in Relu function. Default: 6.0
scale (float, optional) – The scale factor. Default: 6.0
offset (float, optional) – The offset factor. Default: 3.0
name (str, optional) – The default value is None. Normally there is no need for user to set this property. For more information, please refer to Name

Returns

The output tensor with the same shape and data type as input.

Return type

Variable

Examples:

           import paddle.fluid as fluid
import paddle
import numpy as np
paddle.enable_static()

DATATYPE='float32'

x_data = np.array([i for i in range(1,5)]).reshape([1,1,4]).astype(DATATYPE)

x = fluid.data(name="x", shape=[None,1,4], dtype=DATATYPE)
y = fluid.layers.hard_swish(x)

place = fluid.CPUPlace()
#place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
out, = exe.run(feed={'x':x_data}, fetch_list=[y.name])
print(out)  # [[0.66666667, 1.66666667,3., 4.]]

          