Why there is a - in front of the regression_loss?

[Wang-Yu-Qing]

in zero_inflated_lognormal.py line 76:

      regression_loss = -tf.keras.backend.mean(
      positive * tfd.LogNormal(loc=loc, scale=scale).log_prob(safe_labels),
      axis=-1)

     return classification_loss + regression_loss

In the paper, the Loss equals CrossEntropyLoss + LogNormalLoss, so why there is a minus in front of the LogNormalLoss?

[minh-nguyenhoang]

It's for negative log-likelihood (which you should see .log_prob()). In the normal regression setting, like MSE assume a normal distribution with known variance, the log-likelihood is -(x-mu)^2 + constant, which translate to the MSE loss as - -(x-mu)^2. Subtitute normal with laplace distribution, you have MAE loss.