class

LogSoftmax

extendsModule

LogSoftmax(dim: int | None = None)

source edit

Implementing kernel

C++LogSoftmaxBackwardclass

Log-Softmax activation function.

Applies element-wise:

\text{LogSoftmax}(x_i) = \ln\!\left(\frac{e^{x_i}}{\sum_j e^{x_j}}\right) = x_i - \ln\!\sum_j e^{x_j}

Combines the softmax normalisation with a logarithm in a single, numerically stable pass. The output lives in $(-\infty, 0]$ and is intended to be paired with lucid.nn.NLLLoss for multi-class classification, or equivalently used directly with lucid.nn.functional.cross_entropy.

Parameters

dimint or None= None

The dimension along which log-softmax is computed. Must be specified explicitly for most use cases. Default: None.

Notes

Input: $(*)$ — any shape.
Output: $(*)$ — same shape as input; values along dim are non-positive and represent log-probabilities.

Examples

>>> import lucid
>>> import lucid.nn as nn
>>> m = nn.LogSoftmax(dim=-1)
>>> x = lucid.tensor([[1.0, 2.0, 3.0]])
>>> m(x)
tensor([[-2.4076, -1.4076, -0.4076]])
>>> # Classifier output layer — pair with NLLLoss
>>> logits = lucid.randn(32, 10)
>>> log_probs = nn.LogSoftmax(dim=-1)(logits)
>>> log_probs.shape
(32, 10)

Used by 1

lucid.nn.modules

Constructors

dunder

init

→None

__init__(dim: int | None = None)

source edit

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

Instance methods

extra_repr

→str

extra_repr()

source edit

Return a string representation of the layer's configuration.

forward

→Tensor

forward(x: Tensor)

source edit

Apply the activation function element-wise.

Parameters

inputTensor

Input tensor of arbitrary shape.

Returns

Tensor

Output tensor of the same shape as input.

>>> import lucid >>> import lucid.nn as nn >>> m = nn.LogSoftmax(dim=-1) >>> x = lucid.tensor([[1.0, 2.0, 3.0]]) >>> m(x) tensor([[-2.4076, -1.4076, -0.4076]]) >>> # Classifier output layer — pair with NLLLoss >>> logits = lucid.randn(32, 10) >>> log_probs = nn.LogSoftmax(dim=-1)(logits) >>> log_probs.shape (32, 10)