Solver for Large-Scale Rank-One Semidefinite Relaxations

Overview

STRIDE: spectrahedral proximal gradient descent along vertices

A Solver for Large-Scale Rank-One Semidefinite Relaxations

About

STRIDE is designed for solving high-order semidefinite programming (SDP) relaxations of nonconvex polynomial optimization problems (POPs) that admit rank-one optimal solutions. STRIDE is the first algorithmic framework that blends fast local search on the nonconvex POP with global descent on the convex SDP. Specifically, STRIDE follows a globally convergent trajectory driven by a proximal gradient method (PGM) for solving the SDP, while simultaneously probing long, but safeguarded, rank-one "strides", generated by fast nonlinear programming algorithms on the POP, to seek rapid descent.

If you find STRIDE helpful or use it in your projects, please cite:

@article{Yang21arxiv-stride,
  title={STRIDE along Spectrahedral Vertices for Solving Large-Scale Rank-One Semidefinite Relaxations},
  author={Yang, Heng and Liang, Ling and Toh, Kim-Chuan and Carlone, Luca},
  journal={arXiv preprint arXiv:2105.14033},
  year={2021}
}

Dependencies

In order to run the example code example_quasar.m, please download the following two packages and provide paths to them in example_quasar.m:

  • SDPNAL+: STRIDE uses the ADMM+ subroutine in SDPNAL+ to warmstart.
  • Manopt: in example_quasar.m, STRIDE uses Manopt to perform local search to generate rank-one strides.

Example

We provide a starting example about how to use STRIDE to solve the QUASAR semidefinite relaxation in the script example_quasar.m, you can simply run the script in Matlab.

We also provide an example about using MOSEK to solve the same QUASAR problems, you can run the script example_quasar_mosek.m in Matlab (for which please download MOSEK).

Surprise: you should see STRIDE being 50 times faster on data/quasar_100_1.mat (100 measurements, 20 seconds vs. 1000 seconds) and 30 times faster on data/quasar_50_1.mat (50 measurements, 2 seconds vs. 60 seconds). Note that MOSEK cannot solve larger problems than data/quasar_100_1.mat, but STRIDE has successfully solved problems with up to 1000 measurements (in which case the SDP has millions of constraints, see our paper). However, the goal of STRIDE is not to replace MOSEK -for generic SDP problems that have small to medium size, MOSEK is still the go-to solver- but to provide a solution for large-scale SDPs arising from rank-one semidefinite relaxations that are far beyond the reach of MOSEK.

For more examples of using STRIDE for machine perception applications, please navigate to the repo CertifiablyRobustPerception.

How to use STRIDE

The function signature for STRIDE is

[out,Xopt,yopt,Sopt] = PGDSDP(blk,At,b,C,X0,options)

where PGDSDP stands for projected gradient descent in solving a generic SDP problem (which is the backbone of STRIDE). We now describe the detailed input and out of STRIDE.

Input

  • blk,At,b,C: standard SDP data in SDPT3 format. A standard SDP problem can be fully described by blk,At,b,C, where blk describes the sizes of the positive semidefinite constraints (i.e., blocks, we do not support other conic constraints such as second-order cone and nonnegative orthant), At,b describes the linear constraints, and C describes the linear cost function. blk,At,C should be Matlab cell arrays, while b should be a Matlab array. Please refer to the SDPT3 user guide for details. We provide two example problem data for the QUASAR SDP in the subfolder data. If you are interested in how to generate standard SDP problem data from semidefinite relaxations of polynomial optimization problems, please navigate to the repo CertifiablyRobustPerception.

  • X0: a primal initial guess for the SDP problem. Set X0 = [] if no initial guess is available. A good way of providing an initial primal guess is to use fmincon in Matlab to solve the original polynomial optimization problem (if the POP admits a manifold structure, Manopt should be preferred), obtain a local optimizer, and lift the local optimizer to a rank-one feasible point of the SDP. Please read our paper for more details.

  • options: a Matlab structure that provides more information. There are many available parameters in options, but there are two parameters that are required:

    • options.rrFunName: a string that provides the name of the Matlab function that implements a local search scheme. For example, in the provided example example_quasar.m, we use options.rrFunName = 'local_search_quasar' to tell STRIDE that the function local_search_quasar.m implements the local search scheme.

    • options.SDPNALpath: a string that provides the path to the software package SDPNAL+. STRIDE uses the admmplus subroutine in SDPNAL+ to warmstart. The other optional parameters are described in more details below.

Output

  • Xopt,yopt,Sopt: an (approximate) optimal solution to the SDP. In many cases, STRIDE can solve the SDP to very high accuracy (even better than MOSEK). The printout of STRIDE will show the KKT residuals at Xopt,yopt,Sopt.
  • out: a Matlab structure that contains other information such as run history and runtime.

Available parameters

We now list all the available but optional parameters in options:

  • options.S0: a dual initial guess. Typically it is difficult to have a good guess on the dual variables. If not provided, STRIDE uses ADMM+ to generate dual initial guess. However, in some cases, one can exploit problem structure to provide clever dual initializations, please checkout our paper for details.

  • options.tolADMM: accuracy tolerance for using ADMM+. We note that this is perhaps the most important parameter to tune for a fast performance. Setting options.tolADMM very low (e.g., 1e-12) will ask ADMM+ to provide a very accurate warmstart (in the price of more ADMM+ iterations and runtime) so that the main STRIDE algorithm will converge very fast. Setting options.tolADMM very high (e.g., 1e-4) will not require an accurate warmstart from ADMM+ (so very few ADMM+ iterations and less runtime), but it may take many STRIDE main PGD iterations. We recommend tuning this parameter for each specific problem. For the QUASAR examples in this repo, options.tolADMM = 1e-4 works very well.

  • options.maxiterADMM: maximum ADMM+ iterations, default 1e4.

  • options.tolPGD: accuracy tolerance for STRIDE, in terms of maximum relative KKT residual, default 1e-6.

  • options.pgdStepSize: step size for projected gradient descent. We recommend setting options.pgdStepSize = 10.

  • options.maxiterPGD: maximum outer iterations of STRIDE (in performing projected gradient descent), default 10.

  • options.lbfgsmemory: memory of L-BFGS, default 10.

  • options.maxiterLBFGS: maximum iterations of L-BFGS, default 1000.

  • options.lbfgseps: boolean value to decide if using inexactness in L-BFGS (what we call modified L-BFGS), default options.lbfgseps = true. In practice we found this does not have significant effect on the convergence speed.

  • options.rrOpt: a array that contains the indices of the eigenvectors to be rounded in local search, default options.rrOpt = 1:3 and STRIDE generates rounded hypotheses from the leading 3 eigenvectors.

  • options.rrPar: a Matlab structure that contains all user-defined information needed to perform local search. For a template about how to implement a local search scheme, please see below.

Implement your local search scheme

The function signature for a local search scheme is

[Xhat,fhat,info] = local_search_func(Xbar,C,rrPar,rrOpt,roundonly)

where local_search_func is the string that needs to be passed to STRIDE's function call by using options.rrFunName = 'local_search_func', so that STRIDE can evaluate the local_search_func.m function to generate rank-one hypotheses.

We now explain the input and output of local_search_func.

Input

  • Xbar: a primal SDP iterate, generated by STRIDE's projected gradient descent backbone. Xbar has the same format as X0 and Xopt and is a cell array of positive semidefinite matrices (block structure defined by blk).

  • C: linear cost function, same as the C in standard SDP data.

  • rrPar: a Matlab structure that contains any data that are necessary for performing local search using Xbar. For example, rrPar can contain suitable data from the original POP. This rrPar is provide by using options.rrPar when calling STRIDE.

  • rrOpt: a array that contains the indices of the eigenvectors to be rounded in local search. This rrOpt is provided by using options.rrOpt when calling STRIDE.

  • roundonly: a boolean value that decides if STRIDE should just perform rounding (without local search). If roundonly = true, then the user should specify a routine that generates a rounded feasible POP point from Xbar. If roundonly = false, then the user should specify a routine that not only generates a rounded POP iterate, but also perform local search starting from the rounded POP iterate, using suitable nonlinear programming techniques.

Output

  • Xhat: a rank-one SDP iterate, generated by rounding, local search and lifting from Xbar.

  • fhat: value of the SDP objective function attained by Xhat, by using the cost matrix C.

  • info (optional output): a structure that contains the following information:

    • info.nlpsuccess: a boolean value that indicates whether the local search has been successful (for example, if the nonlinear programming solver has failed, then info.nlpsuccess = false).
    • info.minidx: the index of the eigenvector, from which the local search solution is best. For example, if rrOpt = 1:3, and the local solution obtained from rounding the second eigenvector attained the lowest cost, then info.minidx = 2.
    • info.pobjs: the objective values of all local search solutions.
    • info.diffpobj: which is simply info.diffpobj = info.pobjs(1) - fhat.

Although the local_search_func may sound complicated to implement, it is quite natural, because it is simply how one would implement a local optimization method for the POP. Please see utils/local_search_quasar.m for how we implemented a local search scheme for the QUASAR SDP relaxation. Note that one of the major contributions of STRIDE is to use the original POP to attain fast convergence, so please spend time on implementing this local search function for your problem.

Acknowledgements

STRIDE is implemented by Heng Yang (MIT) and Ling Liang (NUS). We would like to thank the feedback and resources from Prof. Kim-Chuan Toh (NUS), and Prof. Luca Carlone (MIT).

You might also like...
Open source implementation of AceNAS: Learning to Rank Ace Neural Architectures with Weak Supervision of Weight Sharing

AceNAS This repo is the experiment code of AceNAS, and is not considered as an official release. We are working on integrating AceNAS as a built-in st

 COD-Rank-Localize-and-Segment (CVPR2021)
COD-Rank-Localize-and-Segment (CVPR2021)

COD-Rank-Localize-and-Segment (CVPR2021) Simultaneously Localize, Segment and Rank the Camouflaged Objects Full camouflage fixation training dataset i

Gradient-free global optimization algorithm for multidimensional functions based on the low rank tensor train format

ttopt Description Gradient-free global optimization algorithm for multidimensional functions based on the low rank tensor train (TT) format and maximu

Pytorch based library to rank predicted bounding boxes using text/image user's prompts.
Pytorch based library to rank predicted bounding boxes using text/image user's prompts.

pytorch_clip_bbox: Implementation of the CLIP guided bbox ranking for Object Detection. Pytorch based library to rank predicted bounding boxes using t

ViViT: Curvature access through the generalized Gauss-Newton's low-rank structure

ViViT is a collection of numerical tricks to efficiently access curvature from the generalized Gauss-Newton (GGN) matrix based on its low-rank structure. Provided functionality includes computing

This is the solution for 2nd rank in Kaggle competition: Feedback Prize - Evaluating Student Writing.

Feedback Prize - Evaluating Student Writing This is the solution for 2nd rank in Kaggle competition: Feedback Prize - Evaluating Student Writing. The

(Personalized) Page-Rank computation using PyTorch

torch-ppr This package allows calculating page-rank and personalized page-rank via power iteration with PyTorch, which also supports calculation on GP

Rank 3 : Source code for OPPO 6G Data Generation Challenge
Rank 3 : Source code for OPPO 6G Data Generation Challenge

OPPO 6G Data Generation with an E2E Framework Homepage of OPPO 6G Data Generation Challenge Datasets H1_32T4R.mat H2_32T4R.mat Please put the original

Open-AI's DALL-E for large scale training in mesh-tensorflow.

DALL-E in Mesh-Tensorflow [WIP] Open-AI's DALL-E in Mesh-Tensorflow. If this is similarly efficient to GPT-Neo, this repo should be able to train mode

Comments
  • Problem with running the example.

    Problem with running the example.

    Thanks for sharing this great solver!

    When I run the example file, I got the following error: image

    As some of the codes are in p files, I do not how to debug the program. I wonder if there is a way to solve this problem?

    The laptop I am using has: Matlab 2021b manopt 5.0 sdpnal+ downloaded from the Google drive link.

    I tried both Linux and Windows systems.

    Thank you!!

    opened by SangliTeng 0
  • About inequality constraints

    About inequality constraints

    Thank you for your impressive work.

    When I test POP with the inequality constraint g, I get an error that appears to be a dimension mismatch of the internal variable. This error occurs even when uncommenting the inequality constraint g in the example BQP program. It would be nice to provide an example of a POP problem that can handle inequality constraints. Thanks a lot.

    opened by WGBmushi 1
  • .p files?

    .p files?

    Hi, thanks for the amazing work on specialized SDP solvers

    I noticed most of the implementation is in .p files, is there a way to access the contents of those for reading and modifying the code?

    opened by Linusnie 0
  • Can it be used with CVX toolbox?

    Can it be used with CVX toolbox?

    I have some SDP problems to solve and I have implemented them in MATLAB using CVX. But the sedumi solver is not good enough, I want to try it with your solver, but I don't know how to start. Do you have any suggestions? Thank you in advance.

    opened by 19982084685 1
Owner
null
Sudoku solver - A sudoku solver with python

sudoku_solver A sudoku solver What is Sudoku? Sudoku (Japanese: 数独, romanized: s

Sikai Lu 0 May 22, 2022
Simple Linear 2nd ODE Solver GUI - A 2nd constant coefficient linear ODE solver with simple GUI using euler's method

Simple_Linear_2nd_ODE_Solver_GUI Description It is a 2nd constant coefficient li

:) 4 Feb 5, 2022
This implements one of result networks from Large-scale evolution of image classifiers

Exotic structured image classifier This implements one of result networks from Large-scale evolution of image classifiers by Esteban Real, et. al. Req

null 54 Nov 25, 2022
[ICLR 2021] Rank the Episodes: A Simple Approach for Exploration in Procedurally-Generated Environments.

[ICLR 2021] RAPID: A Simple Approach for Exploration in Reinforcement Learning This is the Tensorflow implementation of ICLR 2021 paper Rank the Episo

Daochen Zha 48 Nov 21, 2022
An efficient and effective learning to rank algorithm by mining information across ranking candidates. This repository contains the tensorflow implementation of SERank model. The code is developed based on TF-Ranking.

SERank An efficient and effective learning to rank algorithm by mining information across ranking candidates. This repository contains the tensorflow

Zhihu 44 Oct 20, 2022
Source Code for DialogBERT: Discourse-Aware Response Generation via Learning to Recover and Rank Utterances (https://arxiv.org/pdf/2012.01775.pdf)

DialogBERT This is a PyTorch implementation of the DialogBERT model described in DialogBERT: Neural Response Generation via Hierarchical BERT with Dis

Xiaodong Gu 67 Jan 6, 2023
Rank 1st in the public leaderboard of ScanRefer (2021-03-18)

InstanceRefer InstanceRefer: Cooperative Holistic Understanding for Visual Grounding on Point Clouds through Instance Multi-level Contextual Referring

null 63 Dec 7, 2022
Official PyTorch Implementation of Rank & Sort Loss [ICCV2021]

Rank & Sort Loss for Object Detection and Instance Segmentation The official implementation of Rank & Sort Loss. Our implementation is based on mmdete

Kemal Oksuz 229 Dec 20, 2022
TensorFlow Ranking is a library for Learning-to-Rank (LTR) techniques on the TensorFlow platform

TensorFlow Ranking is a library for Learning-to-Rank (LTR) techniques on the TensorFlow platform

null 2.6k Jan 4, 2023
This is the pytorch implementation for the paper: Generalizable Mixed-Precision Quantization via Attribution Rank Preservation, which is accepted to ICCV2021.

GMPQ: Generalizable Mixed-Precision Quantization via Attribution Rank Preservation This is the pytorch implementation for the paper: Generalizable Mix

null 18 Sep 2, 2022