% EXAM CANDIDATE #Y3898772
% NETWORK BUILD INSTRUCTIONS
%
\documentclass{article}

\usepackage[en-GB]{datetime2}
\usepackage[a4paper]{geometry}
\usepackage{fancyvrb, xcolor, hyperref}

\newcommand\examnumber{Examination Candidate \#Y3898772}
\title{CNN Build Instructions}
\author{\examnumber}

\colorlet{ttcol}{darkgray}
\renewcommand\texttt[1]{\textcolor{ttcol}{\ttfamily#1}}
\fvset{formatcom={\small\color{ttcol}}}

\begin{document}
\maketitle
This document describes a suitable method of bootstrapping, building, and
testing the \emph{Flowers} convolutional neural network, as submitted by
\examnumber\ in partial fulfilment of the IMLO module assignment. The
instructions given here assume a Linux shell on a CUDA-equipped machine, such as
\texttt{csgpu13.cs.york.ac.uk}.
\begin{enumerate}
    \item \textbf{Building the Python virtual environment.}
        \begin{enumerate}
            \item Using the system Python interpreter, navigate to the
                \texttt{classifier/} project folder and execute \texttt{python
                -m venv .venv}. Activate the virtual environment with
                \texttt{source .venv/bin/activate}.
            \item Install all third-party library dependencies with \texttt{pip
                install -r requirements.txt}. This will download and install (or
                recover from the system cache) PyTorch with Torchvision for
                tensor and neural network support, and SciPy for the parsing of
                MATLAB files distributed with \emph{Flowers}.
        \end{enumerate}
    \item \textbf{Training and evaluating the model.}
        Still with the virtual environment activated, execute \texttt{python
        ModelDriver.py}; this will train the model for 200 epochs\footnotemark,
        periodically checkpointing the best network parameters to
        \texttt{data/flowers-net.pt}, and then evaluate the trained model.
        %
        \footnotetext{The number of training epochs $E$ can be modified through
        the sole parameter to \texttt{\_ModelDriver.train\_model}, although in
        the interests of avoiding excessive file-system transactions on
        early-epoch (continually improving) models, parameters representing
        maximally accurate networks will only be committed to persistent storage
        after $\lfloor E/3 \rfloor$ epochs.}
        %
        The output of training over the first epoch will appear as follows:
        \VerbatimInput{training-log.txt}

        The model will then self-evaluate, having re-loaded the network
        parameters of the iteration which incurred the maximal validation
        accuracy: \VerbatimInput{testing-log.txt}
\end{enumerate}
\end{document}