silvermecer / miprest Goto Github PK

Author : Kevin S. Brown ([email protected], github.com/thelahunginjeet), Ameya Akkalkotkar

A python package that can be used for Mixed ICA/PCA via the MIPReSt algorithm.  See
A. Akkalkotkar and K. S. Brown, "An algorithm for separation of mixed sparse and Gaussian sources",
Plos ONE, 12(4) e0175775. This package depends on the following additional packages (in addition to
standard stuff like numpy, scipy, etc.), all available and actively maintained at github.com/thelahunginjeet:

pycar
pyica

If you use:

pip install -r requirements.txt

These packages will be installed automatically when miprest is installed.

Example usage of the entire pipeline (data to decompose should be in an Nmixtures x Nsamples matrix X):

# import necessary modules
>>from miprest import miprest,plotting,stopping,pca

# initializes with default options (default directory location and options for ica)
>>mip = MIPReSt() # uses default options

# you can run the parent and decimated decompositions in either order
# this block will run 100 2-fold decimations of the input data
>>mip.parent_decomposition(X)
>>mip.decimated_decomposition(X,dfactor=2,ndec=100)

# if you haven't run BOTH decimations, this function will print an error and return (None,None)
>>R,deltaij = mip.compute_R_delta() # works on stored reproducibility data

# make the R-delta plot and inspect it to determine nsparse
>>fig = plotting.plot_R_delta(R,deltaij)
>>fig.show()

At this point, you should have inspected the R/delta plot and be able to determine
the dimension of the sparse subspace, which I will call nsparse in what follows.

# project out the sparse subspace
>>Xtilde = mip.project_sparse(X,nsparse)

# to harvest the sparse signals, use
>>Asparse,Ssparse = mip.return_sparse(nsparse)

# eigenvalue stopping rules can tell you the dimension of the Gaussian subspace
# there are many rules to choose from (see stopping.py); I'm using broken stick here.
>>stop = stopping.StoppingRule(Xtilde)
>>ng = stop.broken_stick()

# finally, you can construct a basis for the Gaussian subspace (and the 'signals')
# via principal components
>>W,S = pca.pca(Xtilde,ng) # PCA on the matrix of Gaussian components