Scaling issues with large trees and large datasets on windows #77
Description
In some simulation studies, stochtree (R package) is dramatically slower in Windows than in MacOS or Linux [all are running on performant hardware, but of course not completely apples-to-apples]. Features that appear to trigger this massive performance differential:
- Data generating processes that encourage deep trees (because of deep interactions between features)
- Large sample sizes that support growing deep trees
- A large cutpoint grid in the grow-from-root algorithm (the
cutpoint_grid_sizeparameter in thestochtree::bartandstochtree::bcffunction signatures)
To view this performance gap, run the following code on both Windows and MacOS / Linux
# Load libraries library(stochtree) library(rnn) # Random seed random_seed <- 1234 set.seed(random_seed) # Fixed parameters sample_size <- 500000 alpha <- 1.0 beta <- 0.1 ntree <- 50 num_iter <- 10 num_gfr <- 10 num_burnin <- 0 num_mcmc <- 10 cutpoint_grid_size <- 10 min_samples_leaf <- 1 nu <- 3 lambda <- NULL q <- 0.9 sigma2_init <- NULL sample_tau <- F sample_sigma <- T # Initial DGP setup n0 <- 500 p <- 10 n <- n0*(2^p) k <- 2 p1 <- 20 noise <- 0.1 # Full factorial covariate reference frame xtemp <- as.data.frame(as.factor(rep(0:(2^p-1),n0))) xtemp1 <- rep(0:(2^p-1),n0) x <- t(sapply(xtemp1,function(j) as.numeric(int2bin(j,p)))) X_superset <- x*abs(rnorm(length(x))) - (1-x)*abs(rnorm(length(x))) # Generate outcome M <- model.matrix(~.-1,data = xtemp) M <- cbind(rep(1,n),M) beta.true <- -10*abs(rnorm(ncol(M))) beta.true[1] <- 0.5 non_zero_betas <- c(1,sample(1:ncol(M), p1-1)) beta.true[-non_zero_betas] <- 0 Y <- M %*% beta.true + rnorm(n, 0, noise) y_superset <- as.numeric(Y>0) # Downsample to desired n subset_inds <- order(sample(1:nrow(X_superset), sample_size, replace = F)) X <- X_superset[subset_inds,] y <- y_superset[subset_inds] system.time({ bart_obj <- stochtree::bart( X_train = X, y_train = y, alpha = alpha, beta = beta, min_samples_leaf = min_samples_leaf, nu = nu, lambda = lambda, q = q, sigma2_init = sigma2_init, num_trees = ntree, num_gfr = num_gfr, num_burnin = num_burnin, num_mcmc = num_mcmc, cutpoint_grid_size = cutpoint_grid_size, sample_tau = sample_tau, sample_sigma = sample_sigma, random_seed = random_seed ) })