tf.nn.sigmoid_cross_entropy_with_logits

TensorFlow 1 version View source on GitHub

Computes sigmoid cross entropy given logits.

Measures the probability error in discrete classification tasks in which each class is independent and not mutually exclusive. For instance, one could perform multilabel classification where a picture can contain both an elephant and a dog at the same time.

For brevity, let x = logits, z = labels. The logistic loss is

  z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
= z * -log(1 / (1 + exp(-x))) + (1 - z) * -log(exp(-x) / (1 + exp(-x)))
= z * log(1 + exp(-x)) + (1 - z) * (-log(exp(-x)) + log(1 + exp(-x)))
= z * log(1 + exp(-x)) + (1 - z) * (x + log(1 + exp(-x))
= (1 - z) * x + log(1 + exp(-x))
= x - x * z + log(1 + exp(-x))

For x < 0, to avoid overflow in exp(-x), we reformulate the above

  x - x * z + log(1 + exp(-x))
= log(exp(x)) - x * z + log(1 + exp(-x))
= - x * z + log(1 + exp(x))

Hence, to ensure stability and avoid overflow, the implementation uses this equivalent formulation

max(x, 0) - x * z + log(1 + exp(-abs(x)))

logits and labels must have the same type and shape.

labels A Tensor of the same type and shape as logits.
logits A Tensor of type float32 or float64.
name A name for the operation (optional).

A Tensor of the same shape as logits with the componentwise logistic losses.

ValueError If logits and labels do not have the same shape.