Function dfdx::tensor_ops::bce_with_logits

pub fn bce_with_logits<S: Shape, E: Dtype, D: BinaryKernel<BCEKernelOp, E>, LTape, RTape>(
    logits: Tensor<S, E, D, LTape>,
    probs: Tensor<S, E, D, RTape>
) -> Tensor<S, E, D, LTape>where
    LTape: Tape<E, D> + Merge<RTape>,
    RTape: Tape<E, D>,

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.

Numerically Stable Derivation

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