NICE_GP_HIK_Core/tutorial/tutorial.tex

\documentclass[a4paper]{scrreprt}

\usepackage[utf8x]{inputenc}
\usepackage[OT1]{fontenc}
\usepackage[ngerman]{babel}
\selectlanguage{ngerman}
\usepackage{graphicx}
\usepackage{geometry}
\geometry{top=25mm,bottom=25mm,left=30mm,right=25mm}
\usepackage{setspace}
\onehalfspacing
\usepackage{color}

\usepackage{amsmath}
\usepackage{amssymb}

\usepackage[pdfpagelabels]{hyperref}
\hypersetup{colorlinks=false,pdfborder= 0 0 0}

\usepackage{booktabs}

\include{notations}

% define everything for a nice header
\usepackage{scrpage2}
\pagestyle{scrheadings}
% header also on first page of new chapter
\renewcommand*{\chapterpagestyle}{scrheadings} 
% font of header
\renewcommand{\headfont}{\normalfont}
% header
\ihead{\includegraphics[width=0.1\linewidth]{img/hanfried-en-blue}}
 \chead{}
\ohead{\includegraphics[width=0.3\linewidth]{img/fsuText-en}}
\setlength{\headheight}{21mm} % height of header

\title{Tutorial for GP-HIK-CORE}
\author{Alexander Freytag, Erik Rodner \\ \url{firstname.lastname@uni-jena.de}}
\setlength{\itemsep}{-0.5em}

\newcommand{\confSection}[1]{\textit{Section: } #1}
\newcommand{\variable}[1]{\textit{Variable: } #1}
\newcommand{\variableType}[1]{\textit{Type: } #1}
\newcommand{\default}[1]{\textit{Default: } #1}
\newcommand{\infos}[4]{
	\begin{itemize}
	  \setlength{\itemsep}{-0.5em}
	  \item \confSection{#1}
	  \item \variable{#2}
	  \item \variableType{#3}
	  \item \default{#4}
	\end{itemize}
     }

\definecolor{darkred}{rgb}{0.5,0,0}

\begin{document}

\maketitle

\setlength{\parindent}{0pt}
\renewcommand{\chaptername}{chapter}

\chapter{Preface}
\label{chap:Preface}
This document shall help to quickly use the modul gp-hik-core\footnote{\url{https://github.com/cvjena/gp-hik-core}} of our library NICE\footnote{\url{https://github.com/cvjena/nice-core}}.
It is structured as follows: 

\paragraph{Chapter~\ref{chap:SetUp}: How to set-up the library} We briefly explain things to be regarded while setting up our library.

\paragraph{Chapter~\ref{chap:ParamConfig}: Parameters and Configurations}
In \chaptername~\ref{chap:ParamConfig} we give an overview of the majority of built-in parameters, explain their influence briefly and present default settings.

\paragraph{Chapter~\ref{chap:DemoProgs}: Demo-Programs and how to use them}
In \chaptername~\ref{chap:DemoProgs} we introduce some demo programs we created to show how to switch between some parameter settings.


%--------------- setup --------------------
\chapter{Setting up the library}
\label{chap:SetUp}


\textbf{Step 1: Obtain our main library}\newline
\texttt{git clone https://github.com/cvjena/nice-core.git}
\vspace{2em}

\textbf{Step 2: configure everything}\newline
\texttt{cd nice-core/} \newline
\texttt{source setenv.sh}
\vspace{2em}

\textbf{Step 3: Obtain the gp-hik-core module}\newline
\texttt{git clone https://github.com/cvjena/gp-hik-core.git}
\vspace{2em}

\textbf{Step 4: Build everything}\newline
\texttt{cd cd gp-hik-core/} \newline
\texttt{make}
\vspace{2em}

\textbf{Step 5: Verify that everything works properly}\newline
\texttt{make check}



%--------------- parameter configs -------------------- 
\chapter{Parameter Configurations}
\label{chap:ParamConfig}

\section{Classification in general}
  \paragraph{Specify the interative linear solver}
    \infos{GPHIKClassifier}{ils\_method}{string}{CG}
    \begin{tabular}{ll}
      \textbf{Setting} & \textbf{Explanation} \\
	CG     &  the conjugate gradients method\\
	CGL    &  the conjugate gradients method using Lanczos process \\
	SYMMLQ &  the symmetric LQ (SYMMLQ) method using Lanczos process\\
	MINRES &  the minimum residual method using Lanczos process \\
    \end{tabular}

  \paragraph{Specify the optimization method}
    \infos{GPHIKClassifier}{optimization\_method}{string}{greedy}
    \begin{tabular}{ll}
      \textbf{Setting} & \textbf{Explanation} \\
	greedy            &  greedy 1D search in a pre-defined range\\
	downhillsimplex   &  DHS in multiple dimensions \\
	none              &  no optimization at all\\
    \end{tabular}

\section{Classification with Quantization}
  \paragraph{(De-)Activate the Quantization}
    \infos{GPHIKClassifier}{use\_quantization}{bool}{false}
  \paragraph{Specify precision of quantization}
    \infos{GPHIKClassifier}{num\_bins}{integer}{$100$}

\section{Uncertainty Prediction}
\paragraph{Activate or de-active the computation of predictive uncertainties}
  \infos{GPHIKClassifier}{uncertaintyPredictionForClassification}{bool}{false}

\paragraph{Specify how to compute predictive uncertainties}
  \infos{GPHIKClassifier}{varianceApproximation}{string}{approximate\_fine}

  \begin{tabular}{ll}
    \textbf{Setting} & \textbf{Explanation} \\
      approximate\_rough &  use the RAPU method (perhaps with quantization, see our ACCV'12 paper)\\
      approximate\_fine &  use the FAPU method \\
      exact &  use the PUP method\\
      none &  deactivate the computation \\
  \end{tabular}


\section{Various}
% 
  \paragraph{Generalizations of HIK}
    \infos{GPHIKClassifier}{transform}{string}{absexp}

    \begin{tabular}{ll}
      \textbf{Setting} & \textbf{Explanation} \\
	absexp &  absolute value and exponential operation -- pow(fabs(x), exponent)\\
	exp &  exponential operation -- exp(fabs(x), exponent)\\
	MKL &  weights for Multiple Kernel Learning approach\\
	WeightedDim &  weights for each dimension\\
    \end{tabular}

  \paragraph{Useful Output}
    \infos{GPHIKClassifier}{verbose}{bool}{false}

  \paragraph{Computation Time Output}
    \infos{GPHIKClassifier}{verboseTime}{bool}{false}

  \paragraph{Useful Debug Output}
    \infos{GPHIKClassifier}{debug}{bool}{false}

\section{GP related settings}
  \paragraph{Set the GP noise for model regularization}
    \infos{GPHIKClassifier}{noise}{double}{$0.01$}
  \paragraph{(De-)Activate balanced learning}
    \infos{GPHIKClassifier}{learn\_balanced}{bool}{false}
  \paragraph{(De-)Activate aumatic determination of noise}
    \infos{GPHIKClassifier}{optimize\_noise}{bool}{false}


% \chapter{Class Structure}


\chapter{Demo Programs}
\label{chap:DemoProgs}

\section{Toy Example for Classification}

A simple toy example with synthetic data can be run with the program  \texttt{toyExample}. You may call the program with a default configuration via 

  \texttt{../BUILD\_x86\_64/gp-hik-core/progs/toyExample -config ./configs/toyExample.conf}.

\paragraph{Usage and Accuracies}
This program uses synthetic data of $3$ classes with $49$ dimensions. For training, $20$ examples per class are used, whereas we provide $50$ examples per class for testing. 

Without any changes, you should obtain an accuracy of $99.33\%$. If you activate the quantization approach (either in the config file or via 
\texttt{-GPHIKClassifier:use\_quantization true}) the accuracy drops to $89.33\%$ since the sampled features cover small ranges in the input space. However, 
when increasing the number of quantization steps to $1,000$, the resulting accuracy is again $99.33\%$.

\paragraph{Quantization: Runtimes and Memory}
When switching between with and without quantization, you should notice the differences in resulting runtimes. On our computer (single core, $3.4$GHz), 
we obtain the following results:

\begin{center}
 \begin{tabular}{ccccc}
    \textbf{Quantization} & \textbf{Training [s]} & \textbf{Testing [s]} & \textbf{Memory [kB]} & \textbf{Accuracy [\%]}\\
     \hline
    no                    & $5.06$                & $0.19$               & $26,228$             &  $99.33$\\
    yes, $100$            & $5.09$                & $0.01$               & $26,488$             &  $89.33$ \\
    yes, $1,000$          & $5.13$                & $0.01$               & $27,540$             &  $99.33$ \\
  \end{tabular}
\end{center}

\section{Evaluation of Fast Min Kernel}
\begin{itemize}
 \item Name of Program
    \begin{itemize}
      \item[ ] \texttt{completeEvaluationFastMinkernel}
    \end{itemize}
 \item Input:
    \begin{itemize}
      \item[\texttt{-n}] number of examples
      \item[\texttt{-d}] number of dimensions
      \item[\texttt{-v}] additional output
    \end{itemize}
 \item Usage: three main parts
    \begin{enumerate}
      \item initialization (FMK vs. computation of $\kernelMatrix$)
      \item kernel multiplication $\kernelMatrix \cdot \alpha $
      \item kernel sum $\kernelVector^T \cdot \alpha $
    \end{enumerate}
\end{itemize}

\chapter{Closing words}
This library was built to provide a fast and memory efficient possibility for bayesian inference in large-scale scenarios. 
The algorithms are published in the following papers.
\vspace{1em}

Large-scale classification (training, optimization, testing):
\begin{itemize}
  \item \textbf{Large-Scale Gaussian Process Classification with Flexible Adaptive Histogram Kernels} by \textit{Erik Rodner and Alexander Freytag and Paul Bodesheim and Joachim Denzler} (ECCV 2012. 85--98) and
\end{itemize}
Rapid uncertainty computation, incremental and active learning:
\begin{itemize}
  \item \textbf{Rapid Uncertainty Computation with Gaussian Processes and Histogram Intersection Kernels} by \textit{Alexander Freytag and Erik Rodner and Paul Bodesheim and Joachim Denzler} (ACCV 2012. ), which was awarded with the \textcolor{darkred}{\bf ``Best paper honorable mention''}.
\end{itemize}


In case of any problems or suggestions for improvement, don't hesitate to contact us.

\vspace{10em}
\begin{center}
  \includegraphics[width=0.5\linewidth]{img/logoV2blue}
\end{center}


\end{document}