Beginning Anomaly Detection Using

LPSRUW

PDWSORWOLES\SORW

DV

SOW



PDWSORWOLELQOLQH

SOWLPVKRZ [BWUDLQ>@FPDS JUD\

SOWVKRZ

Figure 3-17.  Importing matplotlib.pyplot to see what these training images 

look like

Figure 3-18.  The output of running the code in figure 

3-17

Chapter 3   IntroduCtIon to deep LearnIng

---

88

ILJ SOWILJXUH ILJVL]H 

L 

IRU

I

LQ UDQJH

\BWUDLQVKDSH>@ 

LI

\BWUDLQ>I@  

DQG

L 

SOWVXESORW L

SOWLPVKRZ [BWUDLQ>I@FPDS JUD\

SOW[WLFNV >@

SOW\WLFNV >@

L L

SOWVKRZ

Figure 3-19.  Code to generate a plot that shows some example images for a 

specific class

Chapter 3   IntroduCtIon to deep LearnIng

---

89

Now, extend the shape by a dimension. Right now, the dimensions of the training 

and testing sets are as shown in Figure 

3-21

 and Figure 

3-22

.

Figure 3-20.  The output of running the code in Figure 

3-19

. Notice the amount of 

variation, as well as anomalous data that you would barely consider as numbers

SULQW

[BWUDLQ^`?Q[BWHVW^`?QIRUPDW 

[BWUDLQVKDSH[BWHVWVKDSH

Figure 3-21.  Code to output the shapes of the training and testing data sets

Figure 3-22.  The output of running the code in Figure 

3-21

Chapter 3   IntroduCtIon to deep LearnIng

---

90

For the purposes of training your model, you want to extend this shape to (60000, 28, 

28, 1) and (10000, 28, 28, 1).

A property of images is that there are three dimensions for color images and two for 

grey scale images. Grey scale images are simply row x column since they don’t have color 

channels. Color images, on the other hand, can be formatted as row x column x channel 

or channel x row x column. For color images, the variable channel is 3 because you want 

to know the pixel values for red, green, and blue (RGB).

In this case, it’s grey scale, so you don’t have to worry about the channel variable, but 

the following code will account for both cases if you end up using a data set with color 

such as the CIFAR-10 data set. CIFAR-10 is extremely similar to MNIST, but this time you 

are classifying the 32x32 images based on labels such as cars, birds, ships, etc. and they 

are in color. Run the code in Figure 

3-23

.

Now convert the values to float32 and divide by 255. Right now, the values are all 

integer values that range from 0 to 255, but you want to convert those values to float and 

make them 0 to 1. This is a process called