Physarum-Differentiable-LP-Layer

This repository contains the reference code for our paper Physarum Powered Differentiable Linear Programming Layers and Applications (AAAI-2021)

Requirements

• Python 3
• PyTorch 1.1+

Use our physarum solver to solve a given LP differentiably:

from physarum_solver import physarum_solve

# construct the LP problem in standard form with A, b, c, which can be from either constant or output of some deep network
# then solve the LP problem by calling the solver
x_sol = physarum_solve(A, b, c, step_size=0.5, max_iter=10)
# x_sol is the solution to this LP and can be used in desired further processing

Use our physarum solver in video object segmentation (DMM_Net):

cd ./DMM_Net

Follow the instructions in ./DMM_Net to prepare the data and do the training/testing.

We replaced the original solver from DMM_Net which is customized for matching problem with our physarum solver which works for general LPs, including the matching problem as a special case.

The replacement happens in DMM_Net/dmm/modules/submodules/relax_match.py

To test the performance of our solver on randomly constructed LP problems and compare with the solver from DMM_Net:

cd DMM_Net/dmm/modules/submodules/
python relax_match.py

Use our physarum solver in few-shot learning (MetaOptNet):

cd ./MetaOptNet

Follow the instructions in ./MetaOptNet to prepare the data and do the training/testing.

We replaced the original L-2 SVM solved using Optnet with our L-1 SVM solved using our physarum solver.

The L-1 SVM and our physarum solver are used in ./MetaOptNet/models/classification_heads_pairwise_physarum.py