Struct auto_diff::op::loss::BCEWithLogitsLoss

pub struct BCEWithLogitsLoss { /* private fields */ }

This loss combines a Sigmoid layer and the BCELoss in one single class. This version is more numerically stable than using a plain Sigmoid followed by a BCELoss as, by combining the operations into one layer, we take advantage of the log-sum-exp trick for numerical stability.

-y log (1/(1 + exp(-x))) - (1-y) log(1 - 1/(1 + exp(-x)))

Prediction comes first, label comes second.

Implementations

impl BCEWithLogitsLoss

pub fn new() -> BCEWithLogitsLoss

Trait Implementations

impl OpTrait for BCEWithLogitsLoss

fn apply(&self, input: &[Tensor], output: &[Tensor])

The first is the prediction, the second input is the label ORDER IS IMPORTANT, SECOND ARGUMENT WON’T GET GRADEINT.

fn grad(&self, input: &[Tensor], output_grad: &[Tensor], input_grad: &[Tensor])

Given the forward input value and backward output_grad, Update weight gradient. return backward input gradeint.

fn get_values(&self) -> Vec<Tensor>

access weight values

fn get_grads(&self) -> Vec<Tensor>

access gradient values

fn get_name(&self) -> String

A conventional name for the op

fn get_input_size(&self) -> usize

The number of input needs by this op.

fn get_output_size(&self) -> usize

The number of output produced by this op.

fn set_values(&self, _v: &[Tensor])