torch.nn.functional.softmax — PyTorch 2.7 documentation (original) (raw)

torch.nn.functional.softmax(input, dim=None, _stacklevel=3, dtype=None)[source][source]¶

Apply a softmax function.

Softmax is defined as:

Softmax(xi)=exp⁡(xi)∑jexp⁡(xj)\text{Softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_j \exp(x_j)}

It is applied to all slices along dim, and will re-scale them so that the elements lie in the range [0, 1] and sum to 1.

See Softmax for more details.

Parameters

• input (Tensor) – input
• dim (int) – A dimension along which softmax will be computed.
• dtype (torch.dtype, optional) – the desired data type of returned tensor. If specified, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows. Default: None.

Return type

Tensor

Note

This function doesn’t work directly with NLLLoss, which expects the Log to be computed between the Softmax and itself. Use log_softmax instead (it’s faster and has better numerical properties).