torch_concepts.nn.functional.soft_select

soft_select(values, temperature, dim=1) → Tensor

Soft selection function, a special activation function for a network rescaling the output such that, if they are uniformly distributed, then we will select only half of them. A higher temperature will select more concepts, a lower temperature will select fewer concepts.

Parameters:

• values – Output of the network.

• temperature – Temperature for the softmax function [-inf, +inf].

• dim – dimension to apply the softmax function. Default is 1.

Returns:

Soft selection scores.

Return type:

Tensor