class

Hardtanh

extendsModule

Hardtanh(min_val: float = -1.0, max_val: float = 1.0, inplace: bool = False)

source

Hardtanh activation function.

Applies element-wise:

\text{Hardtanh}(x) = \text{clamp}(x,\; \text{min\_val},\; \text{max\_val}) = \begin{cases} \text{min\_val} & \text{if } x < \text{min\_val} \\ x & \text{if } \text{min\_val} \leq x \leq \text{max\_val} \\ \text{max\_val} & \text{if } x > \text{max\_val} \end{cases}

A piecewise-linear activation that clips the input to a fixed interval. With the default [-1, 1] range it approximates the tanh function using only comparisons and clamps, making it suitable for quantised models.

Parameters

min_valfloat= -1.0

Lower bound of the clamping range. Default: -1.0.

max_valfloat= 1.0

Upper bound of the clamping range. Default: 1.0.

inplacebool= False

Accepted for API compatibility; currently unused. Default: False.

Notes

Input: $(*)$ — any shape.
Output: $(*)$ — same shape as input, values in $[\text{min\_val},\; \text{max\_val}]$ .

ReLU6 is a special case of Hardtanh with min_val=0 and max_val=6.

Examples

>>> import lucid
>>> import lucid.nn as nn
>>> m = nn.Hardtanh(min_val=-1.0, max_val=1.0)
>>> x = lucid.tensor([-3.0, -0.5, 0.0, 0.5, 3.0])
>>> m(x)
tensor([-1. , -0.5,  0. ,  0.5,  1. ])
>>> # Custom range for output normalisation
>>> m = nn.Hardtanh(min_val=0.0, max_val=6.0)
>>> x = lucid.randn(4, 64)
>>> out = m(x)
>>> out.shape
(4, 64)

Methods (3)

dunder

init

→None

__init__(min_val: float = -1.0, max_val: float = 1.0, inplace: bool = False)

source

Initialise the Hardtanh module. See the class docstring for parameter semantics.

forward

→Tensor

forward(x: Tensor)

source

Apply the activation function element-wise.

Parameters

inputTensor

Input tensor of arbitrary shape.

Returns

Tensor

Output tensor of the same shape as input.

extra_repr

→str

extra_repr()

source

Return a string representation of the layer's configuration.