tf.nn.log_poisson_loss

Computes log Poisson loss given log_input.

Gives the log-likelihood loss between the prediction and the target under the assumption that the target has a Poisson distribution. Caveat: By default, this is not the exact loss, but the loss minus a constant term [log(z!)]. That has no effect for optimization, but does not play well with relative loss comparisons. To compute an approximation of the log factorial term, specify compute_full_loss=True to enable Stirling's Approximation.

For brevity, let c = log(x) = log_input, z = targets. The log Poisson loss is

  -log(exp(-x) * (x^z) / z!)
= -log(exp(-x) * (x^z)) + log(z!)
~ -log(exp(-x)) - log(x^z) [+ z * log(z) - z + 0.5 * log(2 * pi * z)]
    [ Note the second term is the Stirling's Approximation for log(z!).
      It is invariant to x and does not affect optimization, though
      important for correct relative loss comparisons. It is only
      computed when compute_full