Trait collenchyma_nn::Softmax

pub trait Softmax<F>: NN<F> {
    fn softmax(
        &self, 
        x: &mut SharedTensor<F>, 
        result: &mut SharedTensor<F>
    ) -> Result<(), Error>;
    fn softmax_plain(
        &self, 
        x: &SharedTensor<F>, 
        result: &mut SharedTensor<F>
    ) -> Result<(), Error>;
    fn softmax_grad(
        &self, 
        x: &mut SharedTensor<F>, 
        x_diff: &mut SharedTensor<F>, 
        result_diff: &mut SharedTensor<F>
    ) -> Result<(), Error>;
    fn softmax_grad_plain(
        &self, 
        x: &SharedTensor<F>, 
        x_diff: &SharedTensor<F>, 
        result_diff: &mut SharedTensor<F>
    ) -> Result<(), Error>;
}

Provides the functionality for a Backend to support Softmax operations.

Required Methods

fn softmax(
    &self,
    x: &mut SharedTensor<F>,
    result: &mut SharedTensor<F>
) -> Result<(), Error>

Computes a Softmax over the input Tensor x with complete memory management.

Saves the result to result.

For a no-memory managed version see softmax_plain.

fn softmax_plain(
    &self,
    x: &SharedTensor<F>,
    result: &mut SharedTensor<F>
) -> Result<(), Error>

Computes the softmax over the input Tensor x without any memory management.

Saves the result to result.

Attention:
For a correct computation result, you need to manage the memory allocation and synchronization yourself.
For a memory managed version see softmax.

fn softmax_grad(
    &self,
    x: &mut SharedTensor<F>,
    x_diff: &mut SharedTensor<F>,
    result_diff: &mut SharedTensor<F>
) -> Result<(), Error>

Computes the gradient of a Softmax over the input Tensor x with complete memory management.

Saves the result to result_diff.

For a no-memory managed version see softmax_grad_plain.

fn softmax_grad_plain(
    &self,
    x: &SharedTensor<F>,
    x_diff: &SharedTensor<F>,
    result_diff: &mut SharedTensor<F>
) -> Result<(), Error>

Computes the gradient of a softmax over the input Tensor x without any memory management.

Saves the result to result_diff.

Attention:
For a correct computation result, you need to manage the memory allocation and synchronization yourself.
For a memory managed version see softmax_grad.

Implementors

• impl Softmax<f32> for Backend<Native>
• impl Softmax<f64> for Backend<Native>
• impl Softmax<f32> for Backend<Cuda>
• impl Softmax<f64> for Backend<Cuda>