Repository files navigation

sgd is an R package for largescale estimation. It features many stochastic gradient methods, built-in models,visualization tools, automated hyperparameter tuning, model checking, intervalestimation, and convergence diagnostics.

At the core of the package is the function

sgd(formula, data, model, model.control, sgd.control)

It estimates parameters for a given data set and model using stochastic gradientdescent. The optional argumentsmodel.control andsgd.control specifyattributes about the model and stochastic gradient method. Taking advantage ofthe bigmemory package, sgd also operates on data sets which are too large to fitin RAM as well as streaming data.

Example of large-scale linear regression:

library(sgd)# DimensionsN <- 1e5  # number of data pointsd <- 1e2  # number of features# Generate data.X <- matrix(rnorm(N*d), ncol=d)theta <- rep(5, d+1)eps <- rnorm(N)y <- cbind(1, X) %*% theta + epsdat <- data.frame(y=y, x=X)sgd.theta <- sgd(y ~ ., data=dat, model="lm")

Any loss function may be specified. For convenience the following arebuilt-in:

• Linear models
• Generalized linear models
• Method of moments
• Generalized method of moments
• Cox proportional hazards model
• M-estimation

The following stochastic gradient methods exist:

• (Standard) stochastic gradient descent
• Implicit stochastic gradient descent
• Averaged stochastic gradient descent
• Averaged implicit stochastic gradient descent
• Classical momentum
• Nesterov's accelerated gradient

Check out the vignette invignettes/ or examples indemo/.In R, the equivalent commands arevignette(package="sgd") anddemo(package="sgd").

To install the latest version from CRAN:

install.packages("sgd")

To install the latest development version from Github:

# install.packages("devtools")devtools::install_github("airoldilab/sgd")

sgd is written byDustin Tran,Junhyung Lyle Kim andPanos Toulis. Please feel free to contribute bysubmitting any issues or requests—or by solving any current issues!

We thank all other members of the Airoldi Lab (led by Prof. Edo Airoldi) for their feedback and contributions.

@article{tran2015stochastic,  author = {Tran, Dustin and Toulis, Panos and Airoldi, Edoardo M},  title = {Stochastic gradient descent methods for estimation with large data sets},  journal = {arXiv preprint arXiv:1509.06459},  year = {2015}}