C++ implementation of Robust Principal Components Analysis (RPCA) using the Armadillo linear algebra library. A Python wrapper for interfacing with HyperSpy is also included.
Two algorithms are included:
- Robust Orthonormal Subspace Learning (ROSL) [1]
- Online Robust PCA (OR-PCA) [2]
ROSL solves the robust PCA problem, recovering a low-rank matrix A from the corrupted observation X according to:
where E is the sparse error term.
[1] X Shu, F Porikli, N Ahuja. (2014) "Robust Orthonormal Subspace Learning: Efficient Recovery of Corrupted Low-rank Matrices". (paper)
[2] J Feng, H Xu and S Yan. (2013) "Online Robust PCA via Stochastic Optimization" (paper)
Dependencies
This library makes use of the Armadillo C++ linear algebra library, which needs to be installed first. It is recommended that you use a high-speed replacement for LAPACK and BLAS such as OpenBLAS, MKL or ACML; more information can be found in the Armadillo FAQs.
Building from source
To build the library, unpack the source and cd into the unpacked directory, then type make:
$ tar -xzf robustpca.tar.gz
$ cd rosl
$ makeThis will generate a C++ library called librosl.so, which is called by the Python module pyrosl.
For a corrupted observation matrix X, one can run the ROSL algorithm with the following required parameters:
import pyrosl
example_rosl = pyrosl.ROSL(
method='full',
rank = 5,
reg = 0.1
)
example_rosl.fit_transform(X)
A = example_rosl.model_
E = example_rosl.residuals_A simple Python script is included with examples of both ROSL and the fast ROSL+ algorithms, as well as further optional parameters. It can be run on the command line with:
$ python example.pyFurther documentation can be found in the file pyrosl.py.