Problem.objective

t3toolbox.backend.optimizers.Problem.objective(x_cores, sample=None, data=None)

def objective(self, x_cores: Tangent, sample=None, data=None):

½‖S(x)-data‖² (+ ρ(x) if regularized) on the full data (sample=None) or an explicit minibatch; no frame sweep (cheap – for the line search / the full-batch stop signal). The reg term reads x_cores directly, so on the manifold it assumes x_cores is left-orthogonal (true for a retraction output – the line-search points; dev/regularization_design.md §4a).

Parameters:

x_cores (Tangent)