Variance Reduced Optimizers in PyTorch

This repo contains implementation of SVRG, SARAH (SpiderBoost), SCSG and Geom-SARAH algorithms based on PyTorch. It was used to produce experiments for the paper Adaptivity of Stochastic Gradient Methods for Nonconvex Optimization (Horvath et al., 2020).

To replicate our experiments, first, recreate conda environment from environment.yml. Run scripts are available in runs/ directory.

If you find this useful, please consider citing:

@article{horvath2020adaptivity,
  title={Adaptivity of stochastic gradient methods for nonconvex optimization},
  author={Horv{\'a}th, Samuel and Lei, Lihua and Richt{\'a}rik, Peter and Jordan, Michael I},
  journal={arXiv preprint arXiv:2002.05359},
  year={2020}
}

References:

• SVRG: Accelerating Stochastic Gradient Descent using Predictive Variance Reduction (Johnson and Zhang, 2013)
• SCSG: Less than a single pass: Stochastically controlled stochastic gradient method. (Lei and Jordan, 2017)
• SARAH: SARAH: A novel method for machine learning problems using stochastic recursive gradient (Nguyen et al., 2017)
• SpiderBoost: SpiderBoost and Momentum: Faster Stochastic Variance Reduction Algorithms (Wang et al., 2019)
• Geom-SARAH: Adaptivity of Stochastic Gradient Methods for Nonconvex Optimization (Horvath et al., 2020)