Final presentation .pdf and .tex upload

Projekty/WojewodaReckoSperkowski/Finalna_tex/bibfile.bib

@@ -0,0 +1,8 @@
+@InProceedings{Chelaramani_2021_WACV,
+    author    = {Chelaramani, Sahil and Gupta, Manish and Agarwal, Vipul and Gupta, Prashant and Habash, Ranya},
+    title     = {Multi-Task Knowledge Distillation for Eye Disease Prediction},
+    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
+    month     = {January},
+    year      = {2021},
+    pages     = {3983-3993}
+}

Projekty/WojewodaReckoSperkowski/Finalna_tex/main.tex

@@ -0,0 +1,321 @@
+\documentclass [xcolor=svgnames, t] {beamer} 
+\usepackage[utf8]{inputenc}
+\usepackage{booktabs, comment} 
+\usepackage[absolute, overlay]{textpos} 
+\usepackage{pgfpages}
+\usepackage[font=footnotesize]{caption}
+\useoutertheme{infolines} 
+\usepackage{advdate}
+\definecolor{gold}{RGB}{254, 206, 0}
+
+\setbeamercolor{title in head/foot}{bg=gray, fg=black}
+\setbeamercolor{author in head/foot}{bg=myuniversity}
+\setbeamertemplate{page number in head/foot}{}
+\usepackage{csquotes}
+\usepackage{multicol}
+\usepackage{amsmath}
+\usepackage[makeroom]{cancel}
+\usepackage{bibentry}
+\usepackage{multirow}
+\usepackage{graphicx}
+\usepackage{textpos}
+
+\usepackage{tikz}
+
+\usetheme{Madrid}
+\definecolor{myuniversity}{RGB}{0, 85, 150}
+\usecolortheme[named=myuniversity]{structure}
+\usepackage{tikz}
+\nobibliography*
+
+
+\title[MTKD Presentation]{Multi-task Knowledge Distillation for Eye Disease Prediction}
+\subtitle{Our implementation of Knowledge Distillation on Multi-task Learning}
+\titlegraphic{\includegraphics[height=2cm]{facultyLOGO.png}}
+\author[Warsztaty Badawcze 2022L]{
+	Mateusz Sperkowski, Szymon Rećko, Malwina Wojewoda}
+\date{\today}
+
+\addtobeamertemplate{navigation symbols}{}{%
+    \usebeamerfont{footline}%
+    \usebeamercolor[fg]{footline}%
+    \hspace{1em}%
+    \insertframenumber/\inserttotalframenumber
+}
+
+\begin{document}
+
+{
+\usebackgroundtemplate{\includegraphics[height=\paperheight,width=\paperwidth]{black.png}}
+\setbeamertemplate{navigation symbols}{}
+\begin{frame}[plain]
+\end{frame}
+}
+
+\begin{frame}
+\maketitle
+\end{frame}
+
+
+\logo{\includegraphics[scale=0.13]{logo.png}~%
+}
+
+\section{Introduction}
+\begin{frame}{Paper Reproduction}
+\centering
+\textbf{\small “Non-reproducible single occurrences are of no significance to
+science.”} 
+{\footnotesize Karl Popper (1959), “The logic of scientific discovery”, p. 66}\\
+\vspace{7mm}
+\includegraphics[width=\linewidth]{eye.png}
+\vspace{5mm}
+{\small \vspace{-2em} \bibentry{Chelaramani_2021_WACV}}
+\end{frame}
+
+
+\section{Table of Contents}
+\begin{frame}
+\frametitle{Table of Contents}
+\tableofcontents
+\end{frame}
+
+
+\section{The Implementation}
+\subsection{Used methods of Transfer Learning}
+\begin{frame}{MTKD}
+    \centering
+\begin{figure}
+\centering
+\begin{minipage}{.5\textwidth}
+  \centering
+  \vspace{10mm}
+  \begin{figure}[H]
+  \includegraphics[width=.9\linewidth]{MTL scheme.png}
+  \caption{Multi-task Learning}
+  \end{figure}
+\end{minipage}%
+\begin{minipage}{.5\textwidth}
+  \centering
+  \vspace{10mm}
+  \begin{figure}[H]
+  \includegraphics[width=\linewidth]{KD scheme.png}
+  \caption{Knowledge distilation}
+  \end{figure}
+\end{minipage}
+\end{figure}
+\end{frame}
+
+\begin{frame}{Pipeline Architecture}
+\vspace{10mm}
+    \centering
+    \begin{figure}[H]
+    \includegraphics[width=1\textwidth]{pipe.png}
+    \caption{Training phases of the MTKD Pipeline}
+    \end{figure}
+\end{frame}
+
+
+\subsection{What worked and what didn't}
+\begin{frame}{What worked and what didn't}
+\begin{itemize}
+    \item We managed to implement the Pipeline from the paper with reasonable changes. 
+    \item We found labeled retinal fundus images similar to those used by the authors, with 2 similar tasks.
+
+    \item We didn't manage to implement two other tasks.
+    \item We didn't manage to thoroughly experiment on many hyperparameters and pretrained models.
+    \item We had a problem with getting good results on model trained using KD.
+    \item Crucial elements that were not described in the article:
+    \begin{itemize}
+
+        \item The shared layers from  pretrained ResNet-50.
+        \item Weights in the sum of losses for Model 2.
+        \item Precise method of ensembling the teachers.
+
+    \end{itemize}
+\end{itemize}
+
+\end{frame}
+
+\subsection{Our Results}
+\begin{frame}{Results}
+    \centering
+
+\begin{table}[]
+\begin{figure}[H]
+\centering
+\resizebox{\textwidth}{!}{%
+\begin{tabular}{|c|cccc|cccc|}
+\hline
+\multirow{2}{*}{} &
+  \multicolumn{4}{c|}{Task 1} &
+  \multicolumn{4}{c|}{Task 2} \\ \cline{2-9} 
+ &
+  \multicolumn{2}{c|}{Balanced Accuracy} &
+  \multicolumn{2}{c|}{F1} &
+  \multicolumn{2}{c|}{Balanced Accuracy} &
+  \multicolumn{2}{c|}{F1} \\ \hline
+Model &
+  \multicolumn{1}{l|}{Train} &
+  \multicolumn{1}{l|}{Test} &
+  \multicolumn{1}{l|}{Train} &
+  \multicolumn{1}{l|}{Test} &
+  \multicolumn{1}{l|}{Train} &
+  \multicolumn{1}{l|}{Test} &
+  \multicolumn{1}{l|}{Train} &
+  \multicolumn{1}{l|}{Test} \\ \hline
+M1{[}1{]} &
+  \multicolumn{1}{c|}{0.911} &
+  \multicolumn{1}{c|}{0.822} &
+  \multicolumn{1}{c|}{0.864} &
+  0.715 &
+  \multicolumn{1}{c|}{-} &
+  \multicolumn{1}{c|}{-} &
+  \multicolumn{1}{c|}{-} &
+  - \\ \hline
+M1{[}2{]} &
+  \multicolumn{1}{c|}{-} &
+  \multicolumn{1}{c|}{-} &
+  \multicolumn{1}{c|}{-} &
+  - &
+  \multicolumn{1}{c|}{0.889} &
+  \multicolumn{1}{c|}{0.811} &
+  \multicolumn{1}{c|}{0.888} &
+  0.789 \\ \hline
+M1{[}1,2{]} &
+  \multicolumn{1}{c|}{0.843} &
+  \multicolumn{1}{c|}{0.758} &
+  \multicolumn{1}{c|}{0.799} &
+  0.755 &
+  \multicolumn{1}{c|}{0.874} &
+  \multicolumn{1}{c|}{0.854} &
+  \multicolumn{1}{c|}{0.832} &
+  0.772 \\ \hline
+Ensemble &
+  \multicolumn{1}{c|}{0.946} &
+  \multicolumn{1}{c|}{0.861} &
+  \multicolumn{1}{c|}{0.920} &
+  0.791 &
+  \multicolumn{1}{c|}{0.870} &
+  \multicolumn{1}{c|}{0.837} &
+  \multicolumn{1}{c|}{0.828} &
+  0.743 \\ \hline
+M2 &
+  \multicolumn{1}{c|}{0.857} &
+  \multicolumn{1}{c|}{0.2} &
+  \multicolumn{1}{c|}{0.803} &
+  0.291 &
+  \multicolumn{1}{c|}{0.883} &
+  \multicolumn{1}{c|}{0.456} &
+  \multicolumn{1}{c|}{0.821} &
+  0.322 \\ \hline
+M3 &
+  \multicolumn{1}{c|}{0.2} &
+  \multicolumn{1}{c|}{0.197} &
+  \multicolumn{1}{c|}{0.042} &
+  0.116 &
+  \multicolumn{1}{c|}{0.385} &
+  \multicolumn{1}{c|}{0.327} &
+  \multicolumn{1}{c|}{0.324} &
+  0.291 \\ \hline
+\end{tabular}%
+}
+\end{table}
+\caption{Results of the trained models}
+\end{figure}
+\end{frame}
+
+\subsection{Diagrams}
+\begin{frame}{MTKD}
+    \centering
+\begin{figure}
+\centering
+\begin{minipage}{.5\textwidth}
+  \centering
+  \vspace{10mm}
+  \begin{figure}[H]
+  \includegraphics[width=0.9\linewidth]{M2_loss.png}
+  \caption{Loss on M2}
+  \end{figure}
+\end{minipage}%
+\begin{minipage}{.5\textwidth}
+  \centering
+  \vspace{10mm}
+  \begin{figure}[H]
+  \includegraphics[width=1.1\linewidth]{M3_loss.jpg}
+  \caption{Loss on M3}
+  \end{figure}
+\end{minipage}
+\end{figure}
+\end{frame}
+
+
+
+
+
+\section{Pros$\And$Cons of the Architecture}
+\begin{frame}{Pros$\And$Cons of the Architecture}
+    \begin{multicols}{2}
+    Pros :
+    \begin{itemize}
+    \vspace{14}
+        \item  Possibility of using a small labelled dataset together with a larger unlabeled one.
+
+
+        \\
+        \item Potentially less costs connected to labeling a dataset.
+        \\
+
+        \item Teacher ensemble method is more effective than just a single teacher for KD.
+    \end{itemize}
+    \columnbreak
+    Cons :
+    \begin{itemize}
+        \item High computing requirements for all task combinations. 
+        \item Large number of hyperparameter choices.
+        \item High requirements for in memory storage of the weights.
+    \end{itemize}
+    \end{multicols}
+
+    \centering
+\end{frame}
+
+\section{Possible future expansions of project}
+\begin{frame}{Possible future expansions of project}
+\vspace{3mm}
+Possible paths for improvements of the pipeline model.
+\vspace{2mm}
+   \begin{itemize}
+        \item Expand model to the unused tasks.
+       \item Generalize code for any dataset/model architectures.
+       \item Test the architecture on other hyperparameters.
+   \end{itemize}
+\end{frame}
+
+\begin{frame}{What we've learned}
+\begin{itemize}
+    \item Structure \& implementation of MTL and KD architectures
+    \item Joining multiple loss functions
+    \item Code and data sharing are crucial for reproduction
+    \item Hyperparameters tuning
+\end{itemize}
+\end{frame}
+
+
+\begin{frame}{Summary}
+    \centering
+    \begin{itemize}
+    \item Reproduction is a crucial part of the scientific method and every scientist should strive to ensure that their paper is reproducible. In case of Computer Science, sharing the code is of upmost importance.
+    \item We successfully analized and implemented a Multi-task Knowledge distillation pipeline.
+    \item It's worth to consider joining multiple transfer learning methods for a single application.
+    \end{itemize}
+    \vspace{7mm}
+    Feel free to send us questions (eg. MS Teams).\\ Thanks for listening!
+\end{frame}
+
+
+\begin{frame} [allowframebreaks]\frametitle{References}
+\bibliographystyle{apalike}
+\bibliography{bibfile}
+\end{frame}
+
+\end{document}