lxysl / pbSGD

Powered Stochastic Gradient Descent Methods for AcceleratedNonconvex Optimization

pbSGD

Powered Stochastic Gradient Descent Methods for Accelerated Nonconvex Optimization

Introduction

We propose a novel technique for improving the stochastic gradient descent (SGD) mthod to train deep networks, which we term pbSGD. The proposed pbSGD method simply raises the stochastic gradient to a certain power $\gamma$ elementwise during iterations and introduces only one additional parameter, namely, the power exponent $\gamma$.

This nonlinear transform function improving the convergence of SGD is orthogonal and complementary to other techniques for accelerating gradient-based optimization methods such as learning rate schedules.

Installation

pbSGD requires Python >= 3.6.0 and torch >= 0.4.0. The torch version is currently provided in this repository.

We're supporting pbSGD installation via pip. You can just run

pip install pbSGD

in your terminal and install pbSGD into your python environment. On the other hand, as pbSGD is modified based on SGD implementation, you can also copy the file into your own project and simply import it.

Examples

We provide demos including experiments on CIFAR-10 and CIFAR-100 to reproduce the results in our work.

Contributions

The CIFAR-10 and CIFAR-10 demo code are borrowed from pytorch-cifar and pytorch-cifar-models respectively.