Skip to main content

omnibias.jax

The JAX backend. Bit-identical to omnibias.torch because both import the same core coefficients. Install with pip install omnibias-jax.

import omnibias.jax as oj

Closed-form neural-field kernels

The closed-form operators for the one-layer field f(x) = b + Σ_h c_h · σ(W_h·x + b_h) on ℝᴰ.

SymbolReturns
neural_field_valuef(x)
neural_field_laplacianΔf(x)
neural_field_hessian∇²f(x)
neural_field_value_and_laplacian(f, Δf)
neural_field_value_grad_laplacian(f, ∇f, Δf)
neural_field_value_grad_hessian(f, ∇f, ∇²f)

These are the kernels behind the first Laplacian tutorial.

Born–Oppenheimer derivatives

Used for nuclear-Hessian / molecular work.

coulomb_potential · make_local_energy · make_bo_force · make_bo_hessian · vibrational_frequencies

Activation registry

get_activation · list_activations · register_activation · is_registered · JaxActivationSpec (alias of ActivationSpec pinned to jax.Array).

Jets

  • directional: mlp_jet, layer_jet, affine_jet, compose_jet, tower_to_jet, jet_to_tower, 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.

Statistical surfaces

Mirrors the torch backend: divergences (kl_divergence, js_divergence, renyi_divergence, wasserstein1, wasserstein2_gaussian, …), GLM helpers (glm_mean, glm_variance, exponential_family_cumulants, fit_natural_parameter, moment_match), moments (gaussian_moment_propagation, delta_method_*), and probability (cdf, soft_histogram, ks_statistic, binned_calibration_error, empirical_band_mass, model_band_mass).

Higher-order field operators

The iterated neural_field_polylaplacian (Δᵏ) used in the high-order PDEs guide lives alongside these kernels in the JAX field surface.

See also