Differentiable Proximal Algorithm Modeling for Large-Scale Optimization
> ∇-Prox is a domain-specific language (DSL) and compiler that transforms optimization problems into differentiable proximal solvers. Departing from handwriting these solvers and differentiating via autograd, ∇-Prox requires only a few lines of code to define a solver that can be *specialized based on user requirements w.r.t memory constraints or training budget* by optimized algorithm unrolling, deep equilibrium learning, and deep reinforcement learning. ∇-Prox makes it easier to prototype different learning-based bi-level optimization problems for a diverse range of applications. We compare our framework against existing methods with naive implementations. ∇-Prox is significantly more compact in terms of lines of code and compares favorably in memory consumption in applications across domains.