Tags give the ability to mark specific points in history as being important
-
v0.15.0
c1cff02a · ·v0.15.0 — Loss framework redesign + init-audit infrastructure Headline changes ---------------- Loss framework: - Level-2 pointwise/reduce contract. Subclasses override pointwise() returning a per-element score map; default calculate_score composes reduce(pointwise(x, y), mask=y_mask). Masking + all-masked→0 guarantee handled centrally. - UncertaintyLoss: heteroscedastic-NLL data term for any pointwise base (family="laplace"|"gauss", logvar_clamp for numerical safety). - Regularizer: generic single-tensor penalty (identity/abs/square/ exp/entropy/huber). Companion to UncertaintyLoss for the +log_var term; also usable for L1 sparsity, L2 magnitude, etc. - CorrectedLoss: modern wrapper analogue of cLossCombiner using torch.Tensor.unfold. 2.46x faster on GPU (see benchmarks/). Init-audit infrastructure (@audit_subclasses): - New class decorator applied to Loss and Dataset. - Catches at class-definition time: missing super().__init__() calls, framework kwargs declared but not forwarded, and framework kwargs not accepted at all. Breaking changes: - base_loss / base → base_metric (UncertaintyLoss, CorrectedLoss). - sigma → log_var (UncertaintyLoss). - Model.__init__ no longer accepts io_spec kwarg. - L1RegisteredUncertaintyLoss removed. - Loss/Dataset subclasses without **kwargs forwarding fail at class definition with a clear message. Deprecated (scheduled for v0.16.0 removal): - cLossCombiner → use CorrectedLoss per base metric. - Observer / Observable / EventBus → use Hook system. Test count: 1381 passing.