crested.tl.losses.PoissonLoss

class crested.tl.losses.PoissonLoss(log_transform=True, eps=1e-07, reduction='sum_over_batch_size', name='PoissonLoss')

Custom Poisson loss for count data with optional log(x + 1) transformation.

This loss function computes the Poisson loss, optionally applying log(x + 1) transformations to predictions and/or targets to ensure non-negativity.

Parameters:

• log_transform (bool (default: True)) – If True, applies log(x + 1) transformation to both predictions and targets.

• eps (float (default: 1e-07)) – Small value to avoid log(0).

• reduction (str (default: 'sum_over_batch_size')) – Type of reduction to apply to the loss. Default: “sum_over_batch_size”.

Attributes table

dtype

Methods table

call(y_true, y_pred)	Compute the Poisson loss.
from_config(config)
get_config()	Return the configuration of the loss function.

Attributes

PoissonLoss.dtype

Methods

PoissonLoss.call(y_true, y_pred)

Compute the Poisson loss.

Parameters:

• y_true – True target values (counts or log(x + 1)-transformed counts).

• y_pred – Predicted values (counts or log(x + 1)-transformed counts).

Returns:

The Poisson loss value for each sample.

classmethod PoissonLoss.from_config(config)

PoissonLoss.get_config()

Return the configuration of the loss function.