prfmodel.fitters.losses.CorrelationLoss

class prfmodel.fitters.losses.CorrelationLoss(name=None, reduction='sum_over_batch_size', dtype=None)

Correlation loss.

Computes the negative Pearson correlation coefficient loss. This loss is agnostic to differences in baseline and amplitude between true values and predictions. This makes it particularly useful for GridFitter where it is the default loss.

Inherits all arguments from keras.losses.Loss.

Notes

Computes the loss by first subtracting the mean from true values and predictions and then computing their cosine similarity. Different from the usual definition of the Pearson correlation coefficient, the loss is zero when either y_true or y_pred are constant (instead of NaN).

Because the loss is invariant to baseline and amplitude, parameters that only shift or scale the prediction are not identifiable under it. See the notes of GridFitter for the consequences.

call(y_true: prfmodel.typing.Tensor, y_pred: prfmodel.typing.Tensor) → prfmodel.typing.Tensor

Compute the loss.

Parameters:

• y_true (prfmodel.typing.Tensor) – Tensor of true targets.

• y_pred (prfmodel.typing.Tensor) – Tensor of predicted targets.

Returns:

Correlation loss. Shape depends on the reduction argument in the class constructor.

Return type:

prfmodel.typing.Tensor