% /solutions/conference-talks/conference-ornate-20min.fr.tex, 22/02/2006 De Sousa
\documentclass{beamer}
%\documentclass[red]{beamer}
\usepackage{pgf,pgfarrows,pgfnodes,pgfautomata,pgfheaps,pgfshade}
\usepackage{amsfonts,amsmath,amssymb,stmaryrd}
\usepackage[latin1]{inputenc}
\usepackage[T1]{fontenc}
%\usepackage[frenchb]{babel}
\usepackage{aeguill}
\usepackage{graphicx}
\usepackage{amsmath,amssymb}
\usepackage{multirow}
\usepackage{multicol}
\usepackage{colortbl}
\usepackage{graphicx}
%\usepackage{figlatex}
\usepackage{multirow}
\usepackage{times}
\usepackage{marvosym}
\usepackage{pifont}
%\usepackage[vlined,longend,algoruled,titlenumbered,french]{algorithm2e}
\usepackage[vlined,french,algoruled]{algorithm2e}
\usepackage[gen]{eurosym}

%      _         _             _                     _                     
%  ___| |_ _   _| | ___    ___| |_    ___ ___  _   _| | ___ _   _ _ __ ___ 
% / __| __| | | | |/ _ \  / _ \ __|  / __/ _ \| | | | |/ _ \ | | | '__/ __|
% \__ \ |_| |_| | |  __/ |  __/ |_  | (_| (_) | |_| | |  __/ |_| | |  \__ \
% |___/\__|\__, |_|\___|  \___|\__|  \___\___/ \__,_|_|\___|\__,_|_|  |___/
%          |___/                                                           
%

\setbeamercovered{transparent}
\usetheme{Warsaw}
% enlever [red] dans le \documentclass si utilisation de la suite :
\usecolortheme[RGB={202,31,39}]{structure} % rouge

%\setbeamertemplate{background}{\includegraphics[height=96mm,width=128mm]{fond.png}}

% definition de la couleur du fond des blocks
\definecolor{ArrierePlanBodyW}{rgb}{0.5,0.5,0.5} % si ecriture blanche
\definecolor{ArrierePlanBodyB}{rgb}{0.85,0.85,0.85} % si ecriture noire
% definition de la couleur des slides
\definecolor{ArrierePlan}{rgb}{0.9,0.9,0.9}

% selection de la couleur du fond des slides
\setbeamercolor{background canvas}{bg=ArrierePlan} 
% selection de la couleur du fond des block avec ecriture noire
%\setbeamercolor{block body}{use=structure,fg=black,bg=ArrierePlanBodyB}
% selection de la couleur du fond des block avec ecriture blanche
\setbeamercolor{block body}{use=structure,fg=white,bg=ArrierePlanBodyW}


\DeclareGraphicsExtensions{pdf,eps,jpg,png}
\graphicspath{{.}{./fig/}{img/}}

\defbeamertemplate*{footline}{section in head/foot}
{
  \leavevmode%
  \hbox{\begin{beamercolorbox}[wd=.5\paperwidth,ht=2.5ex,dp=1.125ex,leftskip=.3cm plus1fill,rightskip=.3cm]{author in head/foot}%
    \usebeamerfont{author in head/foot}\insertshortauthor
  \end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.5\paperwidth,ht=2.5ex,dp=1.125ex,leftskip=.3cm,rightskip=.3cm plus1fil]{title in head/foot}%
    \usebeamerfont{title in head/foot}\insertshorttitle \hfill \insertframenumber{} / \inserttotalframenumber\hspace*{2ex}%
  \end{beamercolorbox}}%
  \vskip0pt%
}


\setbeamercovered{dynamic}
%\logo{\includegraphics[height=1cm]{./img/logo_petit-tr.png}}
% \defbeamertemplate*{headline}{split theme}
% {%
%   \begin{beamercolorbox}{section in head/foot}
%     \insertsectionnavigationhorizontal{2cm}{\includegraphics[height=2cm]{./img/logo_petit-tr.png}}{}
%   \end{beamercolorbox}%
%  }

\defbeamertemplate*{headline}{page title}{
  \leavevmode%
  \begin{beamercolorbox}[wd=.08\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.45cm]{./img/logo_fresnel.pdf}\vfill}%
  \end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.655\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.45cm]{./img/UPC.jpg}\vfill}%
  \end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.1\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfill\hspace*{1mm}\includegraphics[height=1.25cm]{./img/lifc.pdf}\vfill}%
  \end{beamercolorbox}%
   \begin{beamercolorbox}[wd=.17\paperwidth,ht=1.45cm]{section in head/foot}%
      \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.45cm]{./img/ufc2.pdf}\vfill}%
  \end{beamercolorbox}%
}

\defbeamertemplate*{headline}{section in head/foot}
{
  \leavevmode%
  \begin{beamercolorbox}[wd=.1\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.25cm]{./img/lifc.pdf}\vfill}%
  \end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.39\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfil\insertsectionnavigation{.4\paperwidth}\vfil}%
  \end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.43\paperwidth,ht=1.45cm]{subsection in head/foot}%
   \vbox to 1.45cm{\vfil\insertsubsectionnavigation{.4\paperwidth}\vfil}%
  \end{beamercolorbox}%
  %\begin{beamercolorbox}[wd=.17\paperwidth,ht=1.45cm]{section in head/foot}%
  %    \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.45cm]{./img/ufc2.pdf}\vfill}%
  %\end{beamercolorbox}%
  %\begin{beamercolorbox}[wd=.235\paperwidth,ht=1.45cm]{section in head/foot}%
  %    \vbox to 1.45cm{\vfill\hspace*{-0.1mm}\includegraphics[height=1.45cm]{./img/UPC.jpg}\vfill}%
  %\end{beamercolorbox}%
  \begin{beamercolorbox}[wd=.1\paperwidth,ht=1.45cm]{section in head/foot}%
    \vbox to 1.45cm{\vfill\includegraphics[height=1.45cm]{./img/logo_fresnel.pdf}\vfill}%
  \end{beamercolorbox}%

}




% CUSTOM JMN
%%%%%%%%%%%%%
% EXERCICES %
%%%%%%%%%%%%%
\newcounter{exo}
\setcounter{exo}{0}
\newcounter{labelExo}

\newcommand{\Exos}[1]{\stepcounter{exo}{\vspace{0.5cm}
     {\textbf{Exercice \arabic{exo}~:\ #1}}}}

\newcommand{\Solution}{\vspace{0.5cm}
     {\textbf{\emph{Solution de l'exercice \arabic{exo}}}}}

%%%%%%%%%%%%%%%
% DEFINITIONS %
%%%%%%%%%%%%%%%
\newcounter{def}
\setcounter{def}{0}
\newcounter{labelDef}

\renewcommand{\Definition}[1]{\stepcounter{def}{%\vspace{0.5cm}
   {\textbf{Définition \arabic{def}~:\ #1 }}}}


\def\N{{ I\!\!N}}
\def\R{{ I\!\!R}}
\def\Z{{ Z\!\!\! Z}}

\def\Q{I\kern-0.60em Q}
\def\C{I\kern-0.60em C}
\def\F{I\kern-0.20em F}
\def\K{I\kern-0.20em K}

% fleche d'affectation pour les algo
\def\plv{$ \leftarrow$\hspace{1ex}}

\titlegraphic{%
  \centerline{%
    %% \parbox{.165\textwidth}{%
    %%  \includegraphics[width=0.165\textwidth]{logo-2lignes-moyen-tr.png}} 
    %%\hfill
    %%\parbox{.22\textwidth}{%
    %%  \includegraphics[width=0.11\textwidth]{logo_tutelle4.jpg}} 
    %\hfill 
    %\parbox{.22\textwidth}{
    %  \includegraphics[width=0.11\textwidth]{logo_petit-tr.png}}
    \hfill
    \parbox{.5\textwidth}{%
      \includegraphics[width=0.5\textwidth]{CITLogo2.jpg}} 
    \hfill
  }
}


\title[Snake GPU] % (facultatif, à utiliser uniquement si le titre de l'article est trop long)
{\scshape GPU-accelerated snake}

\subtitle {GPU implementation of a region-based segmentation algorithm (snake) for large images}

\author[G. Perrot] % (facultatif, à utiliser seulement avec plusieurs auteurs)
{Gilles Perrot$¹$, Stéphane Domas$¹$, Raphaël Couturier$¹$, Nicolas Bertaux$²$.}
% - Composez les noms dans l'ordre dans lequel ils apparaîtrons dans l'article
% - Utilisez la commande \inst{?} uniquement si les auteurs ont des affiliations
%   différentes.

\institute[] % (facultatif mais généralement nécessaire)
{
 $¹$University of Franche-Comté - Distributed Numerical Algorithmics group (AND),\\
 $²$Fresnel Institute - Physics and Image Computing group (PhyTI) 
}
% - Utilisez la commande \inst uniquement s'il y a plusieurs affectations
% - Faîtes quelque chose de simple, personne ne s'intéresse à votre adresse.

\date[] % (facultatif)
{\small 31 august - 1 september}

%\subject{}
% Inséré uniquement dans la page d'information du fichier PDF. Peut être
% supprimé.



% Si vous avez un fichier nommé "université-logo-nomfichier.xxx", où xxx
% est un format graphique accepté par latex ou pdflatex (comme par exemple .png),
% alors vous pouvez insérer votre logo ainsi :

%\pgfdeclareimage[height=1cm]{logo-lifc}{img/logo-lifc.jpg}
%\logo{\pgfuseimage{logo-lifc}}



% À supprimer si vous ne voulez pas que la table des matières apparaisse
% au début de chaque sous-section :
% 
% \AtBeginSubsection[] {
%   \begin{frame}<beamer>{Lignes directrices}
%     \tableofcontents[currentsection,currentsubsection]
%   \end{frame}
% }


% Si vous souhaitez recouvrir vos transparents un à un,
% utilisez la commande suivante (pour plus d'info, voir la page 74 du manuel
% d'utilisation de Beamer (version 3.06) par Till Tantau) :

%\beamerdefaultoverlayspecification{<+->}
\colorlet{textvert}{green}

\begin{document}

\setbeamertemplate{headline}[page title]
\setbeamertemplate{footline}[default]

\begin{frame}[default]
  \titlepage
\end{frame}


\begin{frame}{Table of contents}
  \tableofcontents
  % Vous pouvez, si vous le souhaitez ajouter l'option [pausesections]
\end{frame}

\setbeamertemplate{headline}[section in head/foot]
\setbeamertemplate{footline}[section in head/foot]

\section{Context}
\begin{frame}{Image segmentation}
\begin{block}{Definition, goal}
\begin{itemize}
  \item Dividing an image in two homogeneous regions.
    \item Reducing the amount of data needed to code information. 
      \item Helping the human perception in certain cases.
\end{itemize}
\end{block}
 \begin{block}{Image characteristics}
 \begin{itemize}
   \item 16 bit-coded gray levels,
   \item From 10~Mpixels to more than 100~Mpixels,
   \item Very noisy.
 \end{itemize}
 \end{block}
\end{frame}

% \begin{frame}{Images of our interest}
% \begin{block}{Origins}
% \begin{itemize}
%   \item Synthetic Aperture RADAR (S.A.R.),
%     \item Ultrasonic (medical imaging), 
%       \item Photographic (IR, nightshots).
% \end{itemize}
% \end{block}
% \begin{block}{Characteristics}
% \begin{itemize}
%   \item 16 bit-coded gray levels,
%   \item From 10~Mpixels to more than 100~Mpixels,
%   \item Very noisy.
% \end{itemize}
% \end{block}
% \end{frame}

\section{Snake algorithm}

% \begin{frame}{Active contour hypothesis}
% \begin{itemize}
%   \item A background region $B$ and a target region $T$, containing  $n_T$ and $n_B$ pixels, compose a $L x H$ pixel image.
%   \item Probability density functions (PDF) $p^B$ and $p^T$ are of the same family (Gaussian in this example).
%   % inutile ds les hypotheses, on ne dit pas qu'on ls connait  
%   %\item Vectors of parameters $\Theta^B$ and $\Theta^T$ of the PDFs $p^B$ and $p^T$ are unknown.
%   \item The contour $\varGamma$ which surrounds $T$ is chosen polygonal.
%   \item [\ding{253}] Implementation of a region-based algorithm. 
% \end{itemize}
% \end{frame}

\begin{frame}{Algorithm  basics : criterion}
\begin{columns}
\begin{column}[l]{5cm}
  \resizebox{5cm}{5cm}{\input{img/fig_target.pdf_t}}
\end{column}
\begin{column}[l]{7.5cm}
  \begin{itemize}
  \item The goal is to find the most likely contour $\varGamma$ (number and positions of nodes).
  \item The criterion used is a \emph{Generalized Likelihood}.\\In the Gaussian case, it is given by 
    $$GL = \dfrac{1}{2}\left[ n_B.log\left({\widehat{\sigma_B}}^2\right) + n_T.log\left({\widehat{\sigma_T}}^2\right)\right]$$
    where $\widehat{\sigma_\Omega}$ is the estimation of the deviation $\sigma$ for the region $\Omega$.
  \end{itemize}
\end{column}
\end{columns}
\end{frame}

\begin{frame}{Algorithm  basics : criterion}
\begin{columns}
\begin{column}[l]{5cm}
  \resizebox{5cm}{5cm}{\input{img/fig_target.pdf_t}}
\end{column}
\begin{column}[l]{7.5cm}
  \begin{itemize}
  \item Parameters estimation is done by 1-D sums on along the contour.
  \item Every pixel coordinates are needed.
  \end{itemize}
\end{column}
\end{columns}
\end{frame}

% \begin{frame}{Algorithm basics : parameters estimation}
%   \begin{itemize}
%   \item In the Gaussian case,  $p_{\Omega}$ has two parameters, average and deviation, which are estimated by maximum likelihood.\\ 
%     If $z$ is the gray level of the pixel of coordinates $(i,j)$:
%     $$
%     \left(
%       \begin{array}{l}
%         \widehat{\mu_{\Omega}} = \frac{1}{n_{\Omega}} \displaystyle\sum_{(i,j)\in \Omega} z(i,j) \\
%         \widehat{\sigma^2_{\Omega}} = \frac{1}{n_{\Omega}} \displaystyle\sum_{(i,j)\in \Omega} \left(z(i,j) - \widehat{\mu_{\Omega}}\right)^2 \\
%       \end{array}
%     \right.
%     $$
%   \item These estimations have to be computed for each test state of the contour $\varGamma$: time-consuming.
%   \end{itemize}
% \end{frame}

% \begin{frame}{Algorithm basics : parameters estimation}
%   \begin{itemize}
% %peut-être illustrer ici
%   \item Based on the Green-Ostogradsky theorem, Chesnaud has shown how to replace those 2-dimensions sums inside the contour 
%     by 1-dimension sums along the contour.
%     \item This optimization implies: 
%       \begin{itemize}
%       \item the precomputation of a few matrices (called cumulated images) containing the potential \emph{contributions} of each pixel of the image,
%         \item the use of constant lookup tables of weighting coefficients to determine the \emph{contributions} of each segment of pixels.
%       \end{itemize}
%   \end{itemize}
% \end{frame}

% \begin{frame}{CPU sequential implementation : outlines}
%   \begin{enumerate}
%   \item Precomputation of the cumulated images.
%     \item Choice of the initial contour $\varGamma_0$: rectangle near the borders.
%       \item Changes made to the contour in order to find the most likely one:
%         \begin{enumerate}[i]
%         \item Move every node as long as it gives a better criterion. 
%           \item Add nodes in the middle of each big enough segment.
%         \end{enumerate}
%   \end{enumerate}
% \end{frame}

 \begin{frame}{Snake algorithm in action}
   \begin{columns}
     \begin{column}[T]{7cm}
       \begin{overprint}      
         \onslide<1>  \includegraphics[width=7cm]{./img/sequences/cochon_it0.png}
         \onslide<2>  \includegraphics[width=7cm]{./img/sequences/cochon_it1.png} 
         \onslide<3>  \includegraphics[width=7cm]{./img/sequences/cochon_it21.png} 
         \onslide<4>  \includegraphics[width=7cm]{./img/sequences/cochon_it22.png} 
         \onslide<5>  \includegraphics[width=7cm]{./img/sequences/cochon_it4.png}  
       \end{overprint}
     \end{column}
     \begin{column}[T]{5cm}
       \only<1> {
         \begin{itemize}
         \item 150~Mpixels image.
         \item Initial contour: 4 nodes.
         \end{itemize}
       }
       \only<2> {
         \begin{itemize}
         \item End of first iteration: no more move can be of interest.
         \end{itemize}
       }
       \only<3> {
         \begin{itemize}
         \item Nodes added in the middle of segments. 
         \end{itemize}
       }
       \only<4> {
         \begin{itemize}
         \item End of second iteration. 
         \end{itemize}
       }
       \only<5> {
         \begin{itemize}
         \item End of fifth iteration\\(36 nodes).
	  \item Fast segmentation.
	  \item Efficient with noise.
         \end{itemize}
       }
     \end{column}
   \end{columns}
 \end{frame}


\section{GPU Implementation}
\begin{frame}{GPU design : key points}
  \begin{itemize}
  \item The parallelism of a modern GPU lays on a SIMT paradigm (Single Instruction Multiple Threads): the same instruction 
    is processed by a great number of threads at a time (up to $2^{16}$).
    \item Threads are compounded in independants blocks with no possible synchronization between blocks.
      \item Threads in a block share a small amount of shared memory (16-48~KBytes).
        \item There are restrictive conditions to be fullfilled in order to make efficent accesses to global and shared memory.
          \item Data transfers between CPU and GPU are slow.
  \end{itemize}
\end{frame}

% \begin{frame}{GPU implementation: precomputations}
%   \begin{itemize}
%   \item One of the cumulated images is not to be computed anymore: values are evaluated on the fly.
%   \item An inclusive parallel prefixsum is performed on each row of the image for each matrix to be processed ($z$,$z^2$).
%   \item[\ding{253}] Speedup is around x7 for images larger than 100~MPixels. Comparison is done with the SSE/CPU implementation of the PhyTI group.
%     \item[\ding{253}] Higher speedups (x15) are obtained with specific versions for constant image sizes.
%   \end{itemize}
% \end{frame}


\begin{frame}{GPU implementation}
When modifying the shape of the contour, we have to compute:
\begin{itemize}
\item Parameters of the corresponding regions.
\item The value of the criterion.  
\item The parallelization needs reside essentially in the parameters estimation.\\Two possible parallelism levels:
  \begin{itemize}
  \item One contour per thread.
  \item One pixel per thread.
  \item[\ding{253}] The \emph{one pixel per thread} rule is far more efficient, due to memory access constraints.
  \end{itemize}
\end{itemize}
\end{frame}


\begin{frame}{GPU implementation: parallelization}
  \begin{columns}
    \begin{column}[T]{6cm}
      \resizebox{6cm}{6cm}{\input{./img/topologie.pdf_t}}
    \end{column}
    \begin{column}[T]{6.5cm}
      \begin{itemize}
      \item Every 16 segments for every node are processed in parallel.
      \item Fits GPU specific parallelism: each segment pixel is processed by a thread.
      \item Criterion values are obtained after several reduction stages.
      \item All nodes are possibly moved in one step.
      \end{itemize}
    \end{column}
  \end{columns}
\end{frame}

% \begin{frame}{GPU implementation: data structure}
% The main idea is to organize, in a single array, every pixels of every segments to be processed.\\
% Thus, for a given state of the contour ($N$ nodes), we:
%   \begin{enumerate}
%   \item Find the largest segment to be processed. It gives: 
%     \begin{itemize}
%     \item the block size $bs$ of the computing grid, 
%     \item the number of blocks needed for each segment ($N_{TB}$).
%     \end{itemize} 
%   \item Compute in parallel,  the coordinates of every pixels of the $16.N$ segments to be considered,
%   \item Make some parallel reductions to finally obtain parameters estimation.
%   \end{enumerate}
% \end{frame}

% \begin{frame}{GPU implementation: data structure}
% \begin{center}
%   \begin{overprint}
%     \onslide<1> \resizebox{8cm}{1.43cm}{\input{img/contribs_segments_seg.pdf_t}}
%     \onslide<2> \resizebox{8cm}{3.78cm}{\input{img/contribs_segments_16seg.pdf_t}}
%     \onslide<3> \resizebox{8cm}{6cm}{\input{img/contribs_segments_even.pdf_t}}
%     \onslide<4> \resizebox{8cm}{6cm}{\input{img/contribs_segments.pdf_t}}
%   \end{overprint}
% \end{center}
% \end{frame}

\begin{frame}{GPU implementation: first results}
  \begin{itemize}
  \item Global speedup around x7-x8 for image sizes from 15 to 150~Mpixels.
  \item First iterations have higher speedups: 
    \begin{itemize}
    \item several large segments,
    \item few inactive threads in the grid.
    \end{itemize}
%   \item Last iterations are sometimes slower than on CPU: 
%     \begin{itemize}
%     \item a lot of small segments,
%     \item more inactive threads in the grid.
%     \end{itemize}
  \end{itemize}
\end{frame}

\begin{frame}{GPU implementation: smart init (reasons)}
  \begin{itemize}
  \item The target shape is often far from initial contour,
  \item It causes the very first iteration to be much more time-consuming than the other ones.
  %\item Horizontal segments contributions are null.
  %\item Vertical segments contributions computations can be fast, through a specific process.
  \item[\ding{253}] It's fast on GPU to find a rectangle near the target. But it needs to overrule the 'one thread/one pixel' principle.
  \end{itemize}
\end{frame}

\begin{frame}{GPU implementation: smart init (process)}
\begin{columns}
\begin{column}[l]{7cm}
  \resizebox{7cm}{6.75cm}{\input{img/smart_init1.pdf_t}}
\end{column}
\begin{column}[l]{5.5cm}
  \begin{itemize}
  \item Realize a periodic sampling of a few hundreds of J-coordinates.
  \item Evaluate in parallel every possible rectangle of diagonal $(0 , j_L)-(H , j_H)$.
  \item Select the one with the best GL criterion.
  \item $j_L$ and $j_H$ are now considered as constants.
  \end{itemize}
\end{column}
\end{columns}
\end{frame}

\begin{frame}{GPU implementation: smart init (process)}
\begin{columns}
\begin{column}[l]{7cm}
  \resizebox{7cm}{6.75cm}{\input{img/smart_init2.pdf_t}}
\end{column}
\begin{column}[l]{5.5cm}
  \begin{itemize}
  \item Given $j_L$ and $j_H$.
  \item Realize a periodic sampling of a few hundreds of I-coordinates.
  \item Evaluate in parallel every possible rectangle of diagonal $(i_L , j_L)-(i_H , j_H)$.
  \item Select the one with the best GL criterion.
  \end{itemize}
\end{column}
\end{columns}
\end{frame}

\begin{frame}{GPU implementation: improvement}
  \begin{itemize}
  \item Global speedup around x10 for image sizes from 15 to 150~Mpixels and a small enough target (as in the example)
  \item Less than 0.6 second for the 150~Mpixels image of this example.
  \end{itemize}
\end{frame}

\section{Conclusion}
\begin{frame}{Conclusion, future works}
  \begin{itemize}
  \item Interesting speedups
  \item Original algorithm is not GPU-friendly
  \item Future works:
    \begin{itemize}
    \item Finding a more suited structure to describe the contour.
    \item Switching to a statistical model independant from a PDF: the potts model.
    \item Benefit from recent features of CUDA v4 (overlapping, multiple kernels).
    \item Extend to a multiple targets algorithm, based on this single target elementary piece of code.
    \end{itemize}
  \end{itemize}
\end{frame}


%\begin{frame}[allowframebreaks]{References}
%\bibliographystyle{plain}
%\bibliography{biblio}
%\nocite{*}
%\end{frame}


\end{document}