Second-order Conditional Gradient Sliding

This is the code to reproduce all the experiments in our paper:

Second-order Conditional Gradient Sliding

Abstract

Constrained second-order convex optimization algorithms are the method of choice when a high accuracy solution to a problem is needed, due to their local quadratic convergence. These algorithms require the solution of a constrained quadratic subproblem at every iteration. We present the Second-Order Conditional Gradient Sliding (SOCGS) algorithm, which uses a projection-free algorithm to solve the constrained quadratic subproblems inexactly. When the feasible region is a polytope the algorithm converges quadratically in primal gap after a finite number of linearly convergent iterations. Once in the quadratic regime the SOCGS algorithm requires first-order and Hessian oracle calls and linear minimization oracle calls to achieve an -optimal solution. This algorithm is useful when the feasible region can only be accessed efficiently through a linear optimization oracle, and computing first-order information of the function, although possible, is costly.

Content

This repository contains the code to compare the performance of the Second-order Conditional Gradient Sliding (SOCGS) Conditional Gradients (CG), Away-step Conditional Gradients (ACG), Pairwise-step Conditional Gradients (PCG), Lazy Away-step Conditional Gradients (Lazy-ACG), Decomposition-invariant Conditional Gradients (DICG), Conditional Gradient Sliding (CGS) and the Newton Conditional Gradient (NCG) algorithms on the following instances:

1. Sparse coding over the Birkhoff polytope.
2. Inverse covariance estimation over spectrahedron.
3. Structured logistic regression over unit ball.

The code runs in Python 3.7 (tested on Windows). The files include the following:

• algorithms.py contains the implementations of the SOCGS, CG, ACG, PCG, Lazy ACG, DICG, CGS, and NCG algorithms.
• functions.py contains the objective functions being used in the numerical experiments, as well as the quadratic approximations to those functions.
• feasibleRegions.py contains the linear minimization oracles for the feasible regions used in the numerical experiments.
• auxiliaryFunctions.py contains miscelaneous functions used throughout the code.
• runExperimentsSOCGSBirkhoff.py contains code to reproduce the Sparse Coding over the Birkhoff polytope experiment.
• runExperimentsSOCGSGLasso.py contains code to reproduce the Inverse covariance estimation over spectrahedron experiment.
• runExperimentsSOCGSL1Ball.py contains code to reproduce the Structured logistic regression experiment.

Reproducing the results from the paper.

In order to reproduce the results from the paper run the following commands. Note that the code will output images and raw data, which will be stored in file directories created when running the algorithm. In the structured logistic regression example the code should automatically download the datasets used. Reproducing the results from Figures 7, 8, and 9 takes approximately 1-2 hours on a laptop with Windows 10, an Intel Core i7 2.4GHz CPU and 6GB RAM. Figure 10 and 12 take around 15-30 minutes and Figure 11 takes around 12-14 hours.

Sparse coding over the Birkhoff polytope with (Figure 7)

python runExperimentsSOCGSBirkhoff.py --max_time 600 --type_solver DICG --num_samples 10000 --dimension 80 --accuracy 1.0e-5 --accuracy_Hessian 0.1 --known_primal_gap True

Sparse coding over the Birkhoff polytope with (Figure 8)

python runExperimentsSOCGSBirkhoff.py --max_time 600 --type_solver DICG --num_samples 100000 --dimension 80 --accuracy 1.0e-5 --accuracy_Hessian 0.1 --known_primal_gap True

Inverse covariance estimation over spectrahedron with (Figure 9)

python runExperimentsSOCGSGLasso.py --max_time 600 --type_solver PCG --dimension 100 --accuracy 1.0e-5 --lambda_value 0.05 --delta_value 1.0e-5 --max_iter 100 --known_primal_gap False

Inverse covariance estimation over spectrahedron with (Figure 10)

python runExperimentsSOCGSGLasso.py --max_time 600 --type_solver PCG --dimension 50 --accuracy 1.0e-5 --lambda_value 0.05 --delta_value 1.0e-5 --max_iter 100 --known_primal_gap False

Structured logistic regression over unit ball with the gisette dataset (Figure 11)

python runExperimentsSOCGSL1Ball.py --max_time 3600 --type_solver ACG --dataset gisette --accuracy 1.0e-4 --lambda_value 0.05 --max_iter 100 --known_primal_gap False

Structured logistic regression over unit ball with the real-sim dataset (Figure 12)

python runExperimentsSOCGSL1Ball.py --max_time 120 --type_solver PCG --dataset real-sim --accuracy 1.0e-4 --lambda_value 0.05 --max_iter 100 --known_primal_gap False

Citation

Please use the following BibTeX entry to cite this software in your work:

    @inproceedings{carderera2020second,
  author    = {Carderera, Alejandro and Pokutta, Sebastian},
  title     = {Second-order Conditional Gradient Sliding},
  journal = {arXiv preprint arXiv:2002.08907},
  year      = {2020},
}

Authors

• Alejandro Carderera
• Sebastian Pokutta