qrpca(R package)

qrpca behaves similarly prcomp. But employs a QR-based PCA instead of applying singular value decomposition on the original matrix. The code uses torch under the hood for matrix operations and supports GPU acceleration.

Installation

Source installation from GitHub:

install.packages('remotes')
remotes::install_github("RafaelSdeSouza/qrpca")
library(qrpca)

Usage

An example of using the package to run a PCA:

library(qrpca)
set.seed(42)
  N <- 1e4
  M <- 1e3
  X <- matrix(rnorm(N*M), M)
  system.time(prcomp(X))
  system.time(qrpca(X))
  system.time(qrpca(X,cuda = TRUE))

For an astronomical example, we use a datacube from MaNGA for galaxy. It comprises a tensor [74,74,4563], of two spatial dimensions and one spectral dimension. The following code reads the cube and flattens the spectra into a matrix of dimension [5476,4563]

require(qrpca);require(reticulate)
require(FITSio);require(ggplot2);
require(dplyr);require(reshape2)
cube <- "manga-7443-12703-LOGCUBE.fits"
df <- readFITS(cube)
n_row <- dim(df$imDat)[1]
n_col <- dim(df$imDat)[2]
n_wave <- dim(df$imDat)[3]
data.2D  <- array_reshape(df$imDat,
c(n_row*n_col,n_wave),order = c("F"))
pca  <- qrpca(data.2D)
# Function to extract the k-th eigenmap
eigenmap <- function(pcobj, k = 1){
 x <- as.matrix(pcobj$x)
 out <- matrix(x[,k],nrow=n_row,ncol=n_col)
 out
}

map <- eigenmap(pca) %>% melt()

ggplot(map,aes(x=Var1,y=Var2,z=value)) +
 geom_raster(aes(fill=value)) +
 scale_fill_viridis_c(option="C") +
 theme(legend.position = "none") 

Dependencies

torch

References

• Sharma, Alok and Paliwal, Kuldip K. and Imoto, Seiya and Miyano, Satoru 2013, International Journal of Machine Learning and Cybernetics, 4, 6, doi: 10.1007/s13042-012-0131-7