omnibias.torch

0.5.0rc1Release candidatePython >=3.10Apache-2.0

PyTorch activation derivatives, trainable operator layers, network jets, and direct differential-operator contractions.

import omnibias.torch

Choose this backend for torch.Tensor workflows and torch.nn.Module training loops. The forward spatial-derivative paths can participate in ordinary parameter backpropagation, so a PDE residual can train the weights that produced it. Shape and architecture restrictions belong to the particular jet or contraction API.

Use activation fast paths for elementwise activation derivatives, directional jets for derivatives along an input direction, multivariate jets for mixed partials, and direct contractions when you only need an operator such as a Laplacian. These compute different objects and have different costs. A direct Laplacian avoids materializing a dense Hessian on supported paths; it is not a promise of dimension-independent work.

The current default is a release candidate. Start in float64 when validating high-order numerical identities, retain the model's intended device and dtype, and assert both output shape and finite parameter gradients. The high-order residual and deep-Laplacian learning paths provide complete executable examples.

Module contract & mathematical conventions

PyTorch activation derivatives and Taylor jets for physics-informed models.

Shared polynomial coefficients define the derivative tower. Operator blocks, mixed jets, and PINN architectures reuse it without nested coordinate autodiff. Architecture and optimization helpers are available in their own modules.

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.

Module source ↗Source revision 9a2d73df0573Running reference examples →

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
`ActivationSpec`	`omnibias.core.ActivationSpec`
`AdaptivePackBank`	`omnibias.torch.AdaptivePackBank`
`AnalyticGaussianConv1d`	`omnibias.torch.AnalyticGaussianConv1d`
`AnalyticGaussianConv2d`	`omnibias.torch.AnalyticGaussianConv2d`
`BankSpec`	`omnibias.core.BankSpec`
`BiasScan`	`omnibias.torch.BiasScan`
`BirkhoffOMBU`	`omnibias.torch.BirkhoffOMBU`
`BlockSpec`	`omnibias.core.BlockSpec`
`DEQConfig`	`omnibias.core.DEQConfig`
`DEQNotContractive`	`omnibias.core.DEQNotContractive`
`DEQResult`	`omnibias.torch.DEQResult`
`DEQSolverUnknown`	`omnibias.core.DEQSolverUnknown`
`EquivariantScan`	`omnibias.torch.EquivariantScan`
`GradientSecant`	`omnibias.core.GradientSecant`
`GrowStrategy`	`omnibias.torch.GrowStrategy`
`GrowableOMBU`	`omnibias.torch.GrowableOMBU`
`GrowableOperatorMultiBiasUnit`	`omnibias.torch.GrowableOMBU`
`JetLineSearchConfig`	`omnibias.core.JetLineSearchConfig`
`LearnablePReLU`	`omnibias.torch.LearnablePReLU`
`LineSearchResult`	`omnibias.core.LineSearchResult`
`MultiPackUnit`	`omnibias.torch.BirkhoffOMBU`
`OMBU`	`omnibias.torch.OMBU`
`OperatorBlock`	`omnibias.torch.OperatorBlock`
`OperatorMultiBiasUnit`	`omnibias.torch.OMBU`
`TemperedActivation`	`omnibias.torch.TemperedActivation`
`WeightLossJetSpec`	`omnibias.core.WeightLossJetSpec`
`affine_jet`	`omnibias.torch.affine_jet`
`affine_jet_mv`	`omnibias.torch.affine_jet_mv`
`analytic_gaussian_taps`	`omnibias.torch.analytic_gaussian_taps`
`antiderivative_jet`	`omnibias.torch.antiderivative_jet`
`arrangement_w_block`	`omnibias.core.arrangement_w_block`
`block_direction`	`omnibias.torch.block_direction`
`block_exact_search`	`omnibias.torch.block_exact_search`
`block_exact_sweep`	`omnibias.torch.block_exact_sweep`
`cmbConv1d`	`omnibias.torch.cmbConv1d`
`cmbConv2d`	`omnibias.torch.cmbConv2d`
`cmbLinear`	`omnibias.torch.cmbLinear`
`compose_jet`	`omnibias.torch.compose_jet`
`compose_jet_mv`	`omnibias.torch.compose_jet_mv`
`compose_jet_riccati`	`omnibias.torch.compose_jet_riccati`
`deq_du_dW`	`omnibias.torch.deq_du_dW`
`deq_solve`	`omnibias.torch.deq_solve`
`deq_vjp`	`omnibias.torch.deq_vjp`
`derivative_jet`	`omnibias.torch.derivative_jet`
`get_activation`	`omnibias.torch.get_activation`
`grow_ombu`	`omnibias.torch.grow_ombu`
`identity_jet`	`omnibias.torch.identity_jet`
`is_registered`	`omnibias.torch.is_registered`
`jet_attention`	`omnibias.torch.jet_attention`
`jet_exp`	`omnibias.torch.jet_exp`
`jet_gradient`	`omnibias.torch.jet_gradient`
`jet_hessian`	`omnibias.torch.jet_hessian`
`jet_line_search`	`omnibias.torch.jet_line_search`
`jet_line_search_on_ray`	`omnibias.torch.jet_line_search_on_ray`
`jet_multiply`	`omnibias.torch.jet_multiply`
`jet_partials`	`omnibias.torch.jet_partials`
`jet_reciprocal`	`omnibias.torch.jet_reciprocal`
`jet_softmax`	`omnibias.torch.jet_softmax`
`jet_to_tower`	`omnibias.torch.jet_to_tower`
`last_linear_block`	`omnibias.core.last_linear_block`
`layer_jet`	`omnibias.torch.layer_jet`
`layer_jet_mv`	`omnibias.torch.layer_jet_mv`
`lhopital_ratio`	`omnibias.torch.lhopital_ratio`
`limit_of_ratio`	`omnibias.torch.limit_of_ratio`
`list_activations`	`omnibias.torch.list_activations`
`mlp_jet`	`omnibias.torch.mlp_jet`
`mlp_jet_mv`	`omnibias.torch.mlp_jet_mv`
`multipack_response`	`omnibias.torch.multipack_response`
`ombu_bias_block`	`omnibias.core.ombu_bias_block`
`one_layer_loss`	`omnibias.torch.one_layer_loss`
`one_layer_loss_grad`	`omnibias.torch.one_layer_loss_grad`
`one_layer_loss_hessian`	`omnibias.torch.one_layer_loss_hessian`
`one_layer_loss_jet`	`omnibias.torch.one_layer_loss_jet`
`one_layer_newton_direction`	`omnibias.torch.one_layer_newton_direction`
`register_activation`	`omnibias.torch.register_activation`
`removable_value`	`omnibias.torch.removable_value`
`scan_response`	`omnibias.torch.scan_response`
`soft_argmax_offset`	`omnibias.torch.soft_argmax_offset`
`steerable_basis`	`omnibias.torch.steerable_basis`
`tower_to_jet`	`omnibias.torch.tower_to_jet`

← Package overview Runnable learning examples →

omnibias.torch

Imports & aliases​

Imports & aliases