Fully First-Order Constrained Bilevel Optimization

This repo includes the implementation of the fully first-order gradient descent algorithm for constrained bilevel optimization.

Dependencies

• CvxpyLayer: https://github.com/cvxgrp/cvxpylayers
• Pytorch
• Numpy

Linear inequality constraints

• Non-fully first order method (implemented using CvxpyLayer): python3 main.py --solver=cvxpylayer --seed=0 --ydim=20 --n-constraints=5

• Fully first-order method: python3 main.py --solver=ffo --seed=0 --ydim=20 --n-constraints=5

Bilinear inequality constraints

The same analysis and approach work for the bilinear case. The bilinear case is implemented in main_bilinear.py. You can also run the non-fully first order and fully first-order methods for the bilinear case using the following commands:

• Non-fully first order method (implemented using CvxpyLayer): python3 main_bilinear.py --solver=cvxpylayer --seed=0 --ydim=20 --n-constraints=5

• Fully first-order method: python3 main_bilinear.py --solver=ffo --seed=0 --ydim=20 --n-constraints=5

You can change the optimizer inside main.py