This issue serves as a reminder and a tracker for the discussion about convenience functionality for predictions in Liesel. Currently, we do not really offer much in that regard.
I propose that we could let certain lsl.Group child-classes know about how they want to be used for prediction, for example:
class PSpline(lsl.Group):
def predict(self, samples, new_data) -> Array:
coef_samples = samples[self.coef.name]
basis_matrix = self.basis_matrix.eval_basis_funs(new_data)
smooth = np.tensordot(
basis_matrix, coef_samples, axes=([1], [-1])
)
return np.moveaxis(smooth, 0, -1) # makes axes fit to samples
What's more, @wiep mentioned that it may be nice to utilize the model graph for predictions.
This issue serves as a reminder and a tracker for the discussion about convenience functionality for predictions in Liesel. Currently, we do not really offer much in that regard.
I propose that we could let certain
lsl.Groupchild-classes know about how they want to be used for prediction, for example:What's more, @wiep mentioned that it may be nice to utilize the model graph for predictions.