Hands-On Deep Learning for Images with TensorFlow

[ 28 ]
Turning images into tensors
In the previous section, we learned a bit about what a tensor is. Now, we're going to use
that knowledge to prepare image data as tensors for machine learning. First, we'll ask a
question: why are we working with data in floating points? Then, we will learn the
difference between samples and the data points at the end of them. Finally, we will
normalize the data for use in machine learning.
So, why a floating point? Well, the real reason is that machine learning is fundamentally a
math optimization problem, and when we're working with floating points, the computer is
trying to optimize a series of mathematical relationships to find learned functions that can
then predict outputs. So, preparing our data for machine learning does involve
reformatting normal binary data, such as an image, into a series of floating point numbers,
which isn't how we'd normally deal with images in terms of image processing, but it's
what's required in order to get machine learning algorithms to engage.
Now, let's talk about samples. By convention, samples are always the first dimension in
your multidimensional array of data. Here, we have multiple samples because machine
learning fundamentally works by looking at a wide array of different data points across a
wide array of different samples and then learning a function to predict outcomes based on
that.
So, each image in our 
train_images
multidimensional array is one of the samples we're
going to be looking at. But as you can see in the Grey scale image (arrays of array) screenshot,
the samples we have right now are definitely not in floating point; these are still in 8-bit
integers.
So, we have to come up with a principled method to transform our images from 8 bit into
floating point.
Now, we're going to start looking at what it really takes to prepare data for machine
learning by looking at normalization. What this really means is that you take your data (in
this case, it's numbers on the range of 
0
to 
255
) and then divide it by another number so
that you squash down the range from 
0
to 
1
:

---

Image Data
Chapter 2
[ 29 ]
Normalization output
This is needed for numerical stability in machine learning algorithms. They simply do
better, converge faster, and become more accurate when your data is normalized on the
range of 
0
to 
1
.
And that's it! We've seen how to deal with input data. Two things to remember: we're going
to be turning everything into floating points, and it's best if we normalize the data on the
range of 
0
to 
1
.

---

Image Data
Chapter 2