PACbayesianNMF 0.1.3

Implementing NMF with a PAC-Bayesian approach relying upon block gradient descent

################################# README.txt############################################### # Name of the project : PACBayesianNMF

# Purpose : Implementation of a PAC-Bayesian approach to Non-Negative Matrix

# Factorization, usein block gradient descent.

# Authors : Astha Gupta <[e-mail]>

# Benjamin Guedj <[e-mail]>

# Reference : Source: https://arxiv.org/abs/1601.01345

# P. Alquier and B. Guedj (2016). “A Sharp Oracle Inequality

# for Bayesian Non-Negative Matrix Factorization”, arXiv preprint.

# Please refer to the paper above for a full description on the

# implemented algorithm.

#License : GPLv3

---

# List of important files and functions

# #################################################################################

# “.pacbayesiannmfblockGradientDescent.py”

# Contains class called blockGradientDescent

# -> setDataMatrix(self,dataMatrix): function to set dataMatrix

# -> setNoOfPatterns(self,K): function to set no of patterns to find

# -> setConditionOnAllSteps(self,concondition_on_step = 1e-2,

# condition_on_inside_step_U = 1e-3,

# condition_on_inside_step_V = 1e-3): Set exit conditions for block

# gradient descent

# -> setConditionOnOutsideStep(self,concondition_on_step = 1e-2):

# Set value for the most outside step, minimize for UV

# -> setConditionOnInsideStepU(self,condition_on_inside_step_U = 1e-3):

# Set value for the inner loop that minimizes for U

# -> setConditionOnInsideStepV(self,condition_on_inside_step_V = 1e-3):

# Set value for the inner loop that minimized for V

#

# -> def applyBlockGradientDescent(self,b = 1e6,lmbd = (float(1)/4)*100,

# pas = 1e-3,printflag = 0):

# This is the main function that applies blockGradientDescent

# b is used to inforce sparcity

# lmbd is lambda from the algorithm

# pas is constant used in algorithm for calculating new U and V

# ##################################################################################

---

# Usage: Import package, create a 2d dataMatrix with each row as a datapoint having values

# between 0 to 1. Create an object of class with parameters as dataMatrix and an

# integer that specifies no of patterns to be detected. Set other parameters using

# set*() methods. Finally to apply Block Gradient Descent use

# applyBlockGradientDescent() method

#

# from pacbayesiannmf import *

# z = blockGradientDescent(dataMatrix,2)

# z.setConditionOnAllSteps(1e-4,1e-6,1e-6)

# (U,V,crit,out)= z.applyBlockGradientDescent(printflag = 1)

#

# Most important output is V, which contains the signal. Each column of V

# crit gives as an array with distance between actual datamatrix and estimated UV with each step

# out is a list of values of different variables with each step, helps in debugging

---

################################## END#####################################################