Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with
or
.
Download ZIP
Browse files

journal: model training

  • Loading branch information...
commit d00c5f6fd0633bcbea8261237a5c493df105f336 1 parent cebf0c6
@spyked spyked authored
Showing with 20 additions and 1 deletion.
  1. +20 −1 journal/journal.tex
View
21 journal/journal.tex
@@ -266,13 +266,32 @@ \subsection*{Letter recognition}
The generated model, however, it's not very good. We will debug it for the
later stages.
-\section*{Final version}
+\section{Final version}
\label{sec:final}
\subsection*{Model training}
+Model training is done by using scripts provided by libSVM along with a tool
+named \texttt{gen2svm} that we previously developed to convert the images into
+a format readable by the SVM library. libSVM uses the \texttt{svm-train}
+executable to train a model. However, first the data provided needs to be
+scaled using \texttt{svm-scale}. This stage can already be executed in an
+automated fashion by a Python script called \texttt{easy.py}, that also
+displays a Gnuplot graph of the model while it's being constructed.
+
+\texttt{easy.py} builds an initial model that it then improves upon by doing
+n-fold cross-validation using a default number of sets $n = 5$. It then
+evaluates and outputs the accuracy of the obtained model on the entire training
+set.
+
+The script uses a Gaussian kernel to learn the classification model.
+
\subsection*{Model testing}
+After obtaining a model, we need to test it on a separate data set. We do this
+using mostly the same infrastructure, as described in Section
+\ref{sec:architecture}.
+
\subsection*{Finding a suitable model}
\subsection*{Results}
Please sign in to comment.
Something went wrong with that request. Please try again.