`omnibias.torch`

The PyTorch backend. Install with pip install omnibias-torch.

import omnibias.torch as ot

Operator-typed layers

Symbol	Description
`OperatorMultiBiasUnit` / `OMBU`	the trainable scalar-operator primitive
`GrowableOperatorMultiBiasUnit` / `GrowableOMBU`	OMBU with a learnable `K`
`GrowStrategy`	curriculum/annealing strategy for the growable unit
`OperatorBlock`	typed wrapper selecting `K`/stencil from an `op` tag
`cmbLinear`	`nn.Linear` with an inline `OperatorBlock`
`cmbConv1d`, `cmbConv2d`	`Conv1d`/`Conv2d` with an inline `OperatorBlock`
`AnalyticGaussianConv1d`, `AnalyticGaussianConv2d`, `analytic_gaussian_taps`	analytic Gaussian convolutions

Symbol	Description
`get_activation(name)`	fetch an `ActivationSpec`
`list_activations()`	list registered names
`register_activation(...)`	register a custom activation
`is_registered(name)`	membership check
`ActivationSpec`	the spec type (re-exported)

Directional (jets) and multivariate (multivariate jets).

directional: mlp_jet, layer_jet, affine_jet, compose_jet, tower_to_jet, jet_to_tower, plus the L'Hôpital helpers lhopital_ratio, limit_of_ratio, removable_value.
multivariate: mlp_jet_mv, layer_jet_mv, affine_jet_mv, compose_jet_mv, identity_jet, jet_multiply, jet_partials, jet_gradient, jet_hessian.

information / divergences: kl_divergence, js_divergence, renyi_divergence, f_divergence, hellinger_distance, total_variation_distance, chi_squared_divergence, cross_entropy, entropy, renyi_entropy, tsallis_entropy, mutual_information, fisher_information, wasserstein1, wasserstein2_gaussian, wassersteinp, sliced_wasserstein, sinkhorn_distance, …
GLM / exponential family: glm_mean, glm_variance, exponential_family_cumulants, fit_natural_parameter, moment_match.
moments: gaussian_moment_propagation, delta_method_gaussian, delta_method_moments.
probability: cdf, soft_histogram, ks_statistic, binned_calibration_error, empirical_band_mass, model_band_mass.

omnibias.torch.architectures ships PINNHeat, CmbNet, CvxLasso, and CvxLogistic.