orcanet.lib.losses

OrcaNet custom loss functions.

Module Contents

Functions

lkl_normal_tfp(y_true, y_pred)

Normal distribution using tfp. See lkl_normal.

lkl_normal(y_true, y_pred)

Negative normal log-likelihood function for n regression output neurons

Attributes

EPS

orcanet.lib.losses.EPS[source]
orcanet.lib.losses.lkl_normal_tfp(y_true, y_pred)[source]

Normal distribution using tfp. See lkl_normal.

orcanet.lib.losses.lkl_normal(y_true, y_pred)[source]

Negative normal log-likelihood function for n regression output neurons with clipping for increased stability.

For stability in the case of outliers, the loss l_i is capped at a maximum of 10 * |pred_i - true_i| for each sample.

Parameters
y_truetf.Tensor

Shape (bs, 2, n) or (bs, 2). y_true[:, 0] is the label of shape (bs, n) (true), and y_true[:, 1] is not used (necessary as tf 2.1 requires y_true and y_pred to have same shape).

y_predtf.Tensor

Shape (bs, 2, n) or (bs, 2). The output of the network. y_pred[:, 0] is mu, and y_pred[:, 1] is sigma.