Function dfdx::losses::binary_cross_entropy_with_logits_loss

source · [−]

pub fn binary_cross_entropy_with_logits_loss<T: Reduce<AllAxes>>(
    logits: T,
    target_probs: T::NoTape
) -> T::Reduced

Expand description

Binary Cross Entropy With Logits in numerically stable way.

Computes target_probs * log(sigmoid(logits)) + (1 - target_probs) * log(1 - sigmoid(logits)) as (1 - target_probs) * logits + log(1 + exp(-logits)).

Inputs

logits - unnormalized inputs. NOT output of sigmoid
target_probs - target values between 0 and 1.

Example

let logits = Tensor1D::new([-1.0, -0.5]);
let target_probs = Tensor1D::new([1.0, 0.25]);
let loss = binary_cross_entropy_with_logits_loss(logits.traced(), target_probs);

Numerically Stable Derivation

See https://www.tensorflow.org/api_docs/python/tf/nn/sigmoid_cross_entropy_with_logits for more information on this.