Skip to content

cyber-meow/Double_stepsize_extragradient

Repository files navigation

Double_stepsize_extragradient

For reproducing experimental results of the paper Explore Aggressively, Update Conservatively: Stochastic Extragradient Methods with Variable Stepsize Scaling.

Requirements

Python3, NumPy, SciPy, Autograd

Usage

To reproduce the results for

Bilinear games

python(3) main_quadratic_quardric.py\
  --algo=[algo] --nb_iterations=100000\
  --init_stepsize_gamma=[gamma1] --init_stepsize_eta=[eta1] --offset=[offset]\
  --save_dir=[log_dir]

Strongly convex-concave problem (Bilinear+Quadratic+Quadric)

python(3) main_quadratic_quardric.py\
  --algo=[algo] --nb_iterations=100000\
  --q2a_coef=1 --q4a_coef=1 --q2b_coef=1 --q4b_coef=1\
  --init_stepsize_gamma=[gamma1] --init_stepsize_eta=[eta1] --offset=[offset]\
  --save_dir=[log_dir]

Linear quadratic Gassian GAN (Covariance learning)

python(3) main_covariance_learning.py\
  --algo=[algo] --nb_iterations=1000000\
  --init_stepsize_gamma=[gamma1] --init_stepsize_eta=[eta1] --offset=[offset]\
  --save_dir=[log_dir]

In the above, algo is either EG or OG. The choice of gamma1, eta1 and offset are as specified in the paper. After the execution of the script, several .npy files containing the necessary information to plot the figures are generated in log_dir. In more details, the evolution of the relevant convergence measure for each single run is recorded.

About

Explore Aggressively, Update Conservatively: Stochastic Extragradient Methods with Variable Stepsize Scaling https://arxiv.org/abs/2003.10162

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages