omnibias-convex

Optimize inside explicit constraints. A barrier path approaches the constrained quadratic optimum.

Static poster · Narrow-screen animation · How this visual is computed

Linear or quadratic objectives and explicit constraints enter; primal candidates, dual information, implicit gradients and checked bounds leave. This is an optimization engine and certificate backend, not a generic neural-network layer factory.

The animation uses computed outputs to explain this package. Frame transitions are illustrative unless a training step is explicitly identified; it is not a performance comparison.

API reference · Source · Tests · Talk to Derivon

The mathematical connection

The log-barrier continuation follows an interior path toward a constrained solution. It shares the idea of progressively removing a smooth approximation with temperature collapse, but its parameter and feasibility semantics are those of the barrier problem. Bias collapse is not needed for its closed-form quadratic/barrier derivatives. Finite termination tolerances must be checked.

Run this README

The examples use omnibias-convex[torch] on Python >=3.10. Their installed-wheel profile selects runtime features, not an editable workspace. Install the prepared prerelease from PyPI:

python -m pip install --pre "omnibias-convex[torch]==0.1.0a2"

For local development before publication, build and test the coordinated wheelhouse using the release guide. The package's wheel profile executes the examples below outside the source checkout.

Existing published consumers may need historical primitive versions; see the compatibility policy.

Why this package exists

Many scientific and decision workflows need a constrained solve inside a larger model. Convex separates problem data, a numerical candidate and its optimality checks. This makes it possible to inspect feasibility and use a differentiable layer where the chosen solve and its regularity assumptions support it.

What you can build

• Linear and quadratic objectives with explicit inequality constraints.
• Log-barrier solver configuration and implicit differentiation layers.
• Dual lower bounds, optimality enclosures and reusable warm starts.

Use convex as a continuous optimization primitive or a bound backend for discrete applications. Start with a small independently solvable problem, check residuals and slacks, then integrate the layer into the surrounding training objective. Named application frontends belong outside this package.

A working example

import torch
from omnibias.convex.torch import solve_qp

# Minimize 0.5*x**2 - 0.5*x over -1 <= x <= 1; optimum is x=0.5.
solution = solve_qp([[1.]], [-0.5], [[1.], [-1.]], [1., 1.], x0=[0.])
assert solution.converged
assert torch.all(solution.slack > 0)
assert abs(float(solution.x[0]) - 0.5) < 1e-3
print(solution.gap)

Choose the right contract

A finite barrier iterate is not automatically an exact optimizer. Inspect converged, gap and slack rather than treating a returned vector as certified. Implicit gradients need appropriate regularity, especially when active constraints change. The solver and certificate functions expose different claims.

Explore and validate

The API guide contains the generated module/export inventory. Use it to find the focused implementation rather than guessing a symbol from another package. The capability map connects the primitives to larger scientific workflows.

From the main repository, run the package’s regression suite:

uv run pytest packages/omnibias-convex/tests -q

License

AGPL-3.0-or-later or commercial. See LICENSE and the licensing policy. Comply with the AGPL terms or obtain a signed commercial grant; commercial use alone does not require payment. Commercial terms describe the alternative. Previously distributed Apache editions, where applicable, retain their original grants.