Skip to contents

Changelog

Source: NEWS.md

emeans 0.1.0

Initial CRAN release: an R port of the e-means clustering in faiss (contrib/e_means, commit 2b648f0605f2ef3d14c4705350f8a0b214463878).

  • emeans() fits e-means, a mini-batch alternative to k-means, and returns an "emeans" object with the same cluster, centers, totss, withinss, tot.withinss, betweenss and size as a kmeans() fit, from a closing sweep that labels every training row by its nearest final center. It supports Euclidean and spherical clustering, a max_points_per_centroid subsample, user-supplied starting centers and fractional nepochs. Randomness comes from R’s RNG, so set.seed() makes a fit reproducible.
  • print(), summary(), predict() and fitted() methods for "emeans" objects.
  • Experimental em_begin(), em_accumulate(), em_update(), em_end_epoch(), em_finalize() and em_centers() run training one batch at a time, for data that does not fit in memory.
  • C++ CPU kernels (Rcpp, RcppArmadillo, optional OpenMP) in double precision, with a direct kernel and a blocked BLAS kernel. options(emeans.kernel =) selects between them and options(emeans.threads =) sets the default thread count.
  • Optional GPU path through the {torch} package (float32): device = "mps" for Apple GPUs, "torch:cpu" and an experimental "cuda" that has not yet been tested on NVIDIA hardware.
  • tidy(), glance() and augment() methods for {generics}, following broom’s columns for kmeans(), and an "emeans" engine for tidyclust::k_means().
  • Results were cross-checked pass by pass against the upstream Python implementation with replayed random draws; see tools/crosscheck/ in the GitHub repository.
  • Four vignettes: getting started, large data, tidymodels and GPU.