ggt.losses.aleatoric_cov_loss

Module Contents

Functions

aleatoric_cov_loss(outputs, targets[, num_var, average])

Computes the Aleatoric Loss while including the full

ggt.losses.aleatoric_cov_loss.aleatoric_cov_loss(outputs, targets, num_var=3, average=True)

Computes the Aleatoric Loss while including the full covariance matrix of the outputs.

If you are predicting for n output variables, then the number of output neuros required for this loss is (3n + n^2)/2.

Args:

outputs: (tensor) - predicted outputs from the model targets: (tensor) - ground truth labels size_average: (bool) - if True, the losses are

averaged over all elements of the batch

Returns:

aleatoric_cov_loss: (tensor) - aleatoric loss

Formula:
loss = 0.5 * [Y - Y_hat].T * cov_mat_inv
  • [Y - Y_hat] + 0.5 * log(det(cov_mat))