Code webpage: Stochastic Variance Reduction Methods for Saddle Point Problems
This page contains information about the software implementation of Stochastic Variance Reduction Methods (e.g. SAG, SAGA, SVRG) for Saddle Point Problems, used in the paper:
Stochastic Variance Reduction Methods for Saddle Point Problems.
P.Balamurugan, Francis Bach.
In Advances in Neural Information Processing Systems, 2016. (To appear)
Preprint available at hal arxiv.
You can download the current version of the code from the link NIPS_sagsaddle_code.zip
Please note: This code is available free only for non-commercial purposes.
How To Use
Unzip the archive NIPS_sagsaddle_code.zip under some path /USER_PATH (this depends on your machine).
Under the path /USER_PATH, a directory named NIPS_sagsaddle_code is
created. The full path of this directory is /USER_PATH/NIPS_sagsaddle_code
Open a Matlab shell window.
Change directory using the following command on the Matlab shell prompt
Type the following command on the Matlab prompt
The function simulations_sagsaddle_nips(ftype, gtype) accepts the following arguments:
ftype: Regularizer options for function f(x)
- 2 for L1-Norm
- 5 for cluster norm
gtype: Loss options for function g(y)
- 1 for Squared Hinge-loss
- 6 for AUC Loss
Please note that a toy data set is used in the code. For other data, please make appropriate changes to the code.