nn.softmax_zero

   fn softmax_zero(tensor: @Tensor<T>, axis: usize) -> Tensor<T>;

Applies the Softmax zero function to an n-dimensional input Tensor rescaling them so that the elements of the n-dimensional output Tensor lie in the range [0,1] and sum to 1 while keeping the zero elements to zero.

The softmax zero on the set $\mathbf{x} = (x_1, ..., x_n)$ is given by :

$$\text{softmax zero}(x_i) = \begin{cases} 0 & \qquad x_i = 0 \\ \frac{e^{x_i}}{ \sum_{x \in {S}} e^{x}} & \qquad \text{otherwise} \end{cases}$$

where $S$ in a subset of $\mathbf{x}$ given by

$$\ S = \{ (x_1, \ldots, x_k) \mid 1 \leq k \leq n, x_j \neq 0 \text{ for } 1 \leq j \leq k \}$$

Args

• tensor(@Tensor<T>) - The input tensor.

• axis(usize) - The axis along which to compute the softmax zero.

Returns

A Tensor of fixed point numbers with the same shape than the input Tensor.

Type Constraints

Constrain input and output types to fixed point tensors.

Examples

use core::array::{ArrayTrait, SpanTrait};

use orion::operators::tensor::{TensorTrait, Tensor, FP8x23Tensor};
use orion::operators::nn::{NNTrait, FP8x23NN};
use orion::numbers::{FP8x23, FixedTrait};

use core::debug::PrintTrait;

fn softmax_zero_example() -> Tensor<FP8x23> {
    let tensor = TensorTrait::<FP8x23>::new(
        shape: array![2, 2].span(),
        data: array![
            FixedTrait::new(0, false),
            FixedTrait::new(8388608, false),
            FixedTrait::new(16777216, false),
            FixedTrait::new(25165824, false),
        ]
            .span(),
    );

    return NNTrait::softmax_zero(@tensor, 1);
}
>>> [[0,0x800000],[2256043,6132564]]
    // The fixed point representation of
    // [[0, 1],[0.2689, 0.7311]]