In the previous lesson, we looked at how
you would go from the binary classifiers we'd been looking at
throughout the course, to update it from
multi-class classifier. In this video, we'll look
at the Rock Paper Scissors workbook and explore how it perform multi-class
classification. The first step is
to get the data. There are two zip files, one for the training data,
and one for the test set. Once we have the data, we'll unzip it into
the subdirectories here. Let's now take a look at some
of the files within that. You can see there are 840 of each class and
a few filenames. Let's now plot a few of the files so we can see what
the data looks like. As we can see, we have
a few different hands with different skin colors, and both male and female hands. Let's now build the model. Note that while the images
are 300 by 300, we are setting up
the image generators to give us 150 by 150 variance. It will resize them
on the fly and augment the ones in
the training directory. We'll also print out the model
if you want to inspect it. Now, the training begins. I'll speed up the video so
that you can see the progress. Keep an eye on the accuracy
and the validation accuracy. I'm only training for 25 epochs, based on the chart you
saw in the last lesson. But by the time we
reach the 10th epoch, we're already doing quite well. By the time we finish, the training data
is above 98% and the validation data is
at 95 percent accuracy. This is highly specialized data
that's optimized for this lesson and not a
great real-world scenario for Rock, Paper, and Scissors. I'll discuss why in a moment but let's first
plot the accuracy. We can see the training improving and
trending towards one. The validation zig-zags a bit, but it's always between 0.9 and one after the first few epochs. Now, let's explore testing it with some images that
it hasn't previously seen.