Changelog • RmlxStats

RmlxStats 0.4.0

mlxs_lm_fit() now uses Rmlx’s specialized code for QR on the GPU when the design has more than 10 million elements, with qr_method available for explicit selection.

mlxs_glm() now moves to float64 on the cpu where necessary to compute more accurate estimates.
New mlxs_glm_control() function.
mlxs_lm() and mlxs_glm() now reject rank-deficient x. A bug which meant we calculated qr(x) twice has now been fixed.
mlxs_lm(), mlxs_lm_fit(), and mlxs_glm_control() now expose rank_tol to tune rank-deficiency detection. Set rank_tol = FALSE to skip rank checks entirely.
Bugfix: mlxs_lm() now drops unused factor levels.
New bread(), estfun() and hatvalues() methods for mlxs_lm to allow for sandwich-style robust standard errors.
More mlxs_lm methods now return base R objects by default, controllable via the output argument.
confint.mlxs_lm() and confint.mlxs_glm() can now return bootstrap confidence intervals. So can the respective summary() methods.
Speedups for some augment(), predict() and summary() methods.

Added mlxs_prcomp(), a prcomp()-style PCA interface with exact and randomized truncated MLX-backed decomposition paths. Benchmarks show this greatly outperforms base R prcomp() and other specialised packages for fast PCA.
Reworked mlxs_glmnet(). It can now outperform glmnet::glmnet() for large problems (roughly n x p > 5,000,000).
Added mlxs_cv_glmnet() as a cross-validation wrapper for the MLX-backed elastic-net path fits, analogous to glmnet::cv.glmnet().
Export mlxs_lm_fit() so advanced users can call the MLX-backed QR solver directly.