Skip to contents

General negative log likelihood loss function

loss_negative_log_likelihood.Rd

Bear in mind, that the number of the model outputs must reflect the number of distribution parameters. For example, if you use normal distribution (tfprobability::tfd_normal()), which is described with two parameters (mean and standard deviation), the model should return two values per each timestep. In othr words, it produces a distribution as a forecast rather than a point estimate. When the model is trained, we have two options to generate the final forecast:

  • use the expected value of the distribution (e.g. mean for normal distribution)

  • sample a value from the distribution Additionally, having the distribution we can compute prediction intervals. Remeber also about the constraints imposed on the parameter values, e.g. standard deviation must be positive.

Usage

loss_negative_log_likelihood(...)

Arguments

distribution

A probability distribution function from tfprobability package. Default: tfprobability::tfd_normal()

References

  1. Cross-Entropy, Negative Log-Likelihood, and All That Jazz

  2. D. Salinas, V. Flunkert, J. Gasthaus, T. Januschowski, DeepAR: Probabilistic forecasting with autoregressive recurrent networks, International Journal of Forecasting(2019)

Examples

y_pred <- array(runif(60), c(2, 10, 2))
y_true <- array(runif(20), c(2, 10, 1))

loss_negative_log_likelihood(
    distribution = tfprobability::tfd_normal,
    reduction = 'auto'
 )(y_true, y_pred)
#> tf.Tensor(14.322809, shape=(), dtype=float32)
loss_negative_log_likelihood(reduction = 'sum')(y_true, y_pred)
#> tf.Tensor(286.45618, shape=(), dtype=float32)