omnibias.curvature.torch
Scoped parameter-curvature formulas, Fisher and Gauss–Newton constructions, natural steps, and sharpness utilities.
import omnibias.curvature.torchModule contract & mathematical conventions
Torch, multi-layer exact-curvature sharpness.
The arbitrary-depth, PyTorch counterpart of omnibias.curvature.sharpness
(one-layer, JAX). Curvature functionals are computed matrix-free from exact
Hessian-vector products, so they scale to deep networks and drop into any
training loop -- an MLP, a JetMLP PINN, or the omnibias.score.flow CNF velocity
field.
Import as::
from omnibias.curvature.torch import sharpness_aware_loss, sam_objective
This subpackage requires the optional torch extra
(pip install "omnibias-curvature[torch]"); it is intentionally *not*
imported by the JAX-only top-level omnibias.curvature.
Shared formulas do not imply universal floating-point identity. Verify dtype, shape, derivative-order support, and the stated numerical assumptions for the path you use.
Imports & aliases
These names share a definition with another public module. Follow a name for its full contract, parameters, and examples.
| Import from this module | Canonical reference |
|---|---|
CollapseResult | omnibias.curvature.torch.CollapseResult |
ExactSAM | omnibias.curvature.torch.ExactSAM |
HessianOperator | omnibias.curvature.torch.HessianOperator |
curvature_sharpness | omnibias.curvature.torch.curvature_sharpness |
dense_hessian | omnibias.curvature.torch.dense_hessian |
hessian_eigenvalue_extremes | omnibias.curvature.torch.hessian_eigenvalue_extremes |
hessian_frobenius_sq | omnibias.curvature.torch.hessian_frobenius_sq |
hessian_top_eigenvalue | omnibias.curvature.torch.hessian_top_eigenvalue |
hessian_trace | omnibias.curvature.torch.hessian_trace |
hutchinson_frobenius_sq | omnibias.curvature.torch.hutchinson_frobenius_sq |
hutchinson_trace | omnibias.curvature.torch.hutchinson_trace |
hvp | omnibias.curvature.torch.hvp |
min_norm_solve | omnibias.curvature.torch.min_norm_solve |
numerical_rank | omnibias.curvature.torch.numerical_rank |
rank_collapse | omnibias.curvature.torch.rank_collapse |
regularization_path | omnibias.curvature.torch.regularization_path |
regularized_solve | omnibias.curvature.torch.regularized_solve |
sam_objective | omnibias.curvature.torch.sam_objective |
sam_sharpness_gap | omnibias.curvature.torch.sam_sharpness_gap |
sharpness_aware_loss | omnibias.curvature.torch.sharpness_aware_loss |
top_eigenvalue | omnibias.curvature.torch.top_eigenvalue |