examodels

Experimental

This package is experimental and is not yet stable. It is under active development: the API may change without notice, releases are unversioned in practice, and the compiler interface in particular is still moving. It has not been through the round of real use that would justify calling it otherwise.

Use it for research and experiments, expect to pin a commit if you depend on it, and please report what breaks.

Python interface to ExaModels.jl — SIMD-parallel algebraic modelling and automatic differentiation for nonlinear programs, on CPU threads or GPUs.

import examodels as exa

N = 10
core = exa.Core()
x = exa.add_var(core, N, start=[-1.2 if i % 2 == 0 else 1.0 for i in range(N)])

exa.add_obj(core, lambda i: 100 * (x[i-1]**2 - x[i])**2 + (x[i-1] - 1)**2,
            over=range(1, N))

exa.add_con(core, lambda i: 3 * x[i+1]**3 + 2 * x[i+2] - 5
            + exa.sin(x[i+1] - x[i+2]) * exa.sin(x[i+1] + x[i+2])
            + 4 * x[i+1] - x[i] * exa.exp(x[i] - x[i+1]) - 3,
            over=range(0, N - 2))

sol = exa.Model(core).solve()
print(sol.status, sol.objective, sol[x])

You do not need Julia installed: the backend runtime is downloaded into the environment on first use. And for a script that is re-run rather than kept alive, the model cache compiles the model ahead of time once, after which every run is free of compilation overhead entirely.

Contents

• Installation
  • Solvers
  • GPUs
  • The compiler
  • Development
• Mathematical abstraction
• Tutorial
  • Getting started
  • Index sets
  • Adding terms to constraints
  • Parameters
  • The first call, and every call after it
  • GPU
  • AC optimal power flow
  • Oracle constraints
  • Two-stage stochastic models
• Recipes and compiled libraries
  • Compiling
  • Using the result
  • Giving the data to the library instead
• The model cache — no compilation overhead after the first run
  • Write it as a recipe
  • What the key covers
  • Fixed models cache too
  • What it needs, and how it degrades
  • Where entries live
• CuPy interchange
  • Ownership
  • Coexisting in one process
• API
  • Building a model
  • The built model
  • Handles
  • Index sets
  • Recipes
  • The compiler and the model cache
  • Solvers and backends
  • Oracles, two-stage models and wrappers
  • CuPy interchange
  • Elementwise functions

Relation to the Julia manual

The pages here mirror the ExaModels.jl manual. Three of its pages have no counterpart, for reasons rather than by omission:

• JuMP interface — JuMP is a Julia modelling language; the corresponding Python entry points are the ones documented here.

• Developing solvers / Upgrading — both concern writing Julia against the backend.

• Quadrotor and Distillation examples — not yet ported; the constructs they use (multi-dimensional blocks, subexpressions, product index sets) are all supported and covered in the tutorial.