Beginning Anomaly Detection Using

LPSRUW

WRUFK

LPSRUW

WRUFKQQ

DV

QQ

LPSRUW

WRUFKYLVLRQ

LPSRUW

WRUFKYLVLRQWUDQVIRUPV

DV

WUDQVIRUPV

LPSRUW

WRUFKRSWLP

DV

RSWLP

LPSRUW

WRUFKQQIXQFWLRQDO

DV

)

LPSRUW

QXPS\

DV

QS

GHYLFH WRUFKGHYLFH FXGD 

LI

WRUFKFXGDLVBDYDLODEOH 

HOVH

FSX

Figure B-3.  Importing the basic modules needed to create your network

Figure B-4.  The code in Figure 

B-3

 in a Jupyter cell

appendix B   intro to pytorch

---

366

FODVV

&11 QQ0RGXOH 

GHI

BBLQLWBB VHOI 

VXSHU

&11VHOI BBLQLWBB

VHOIFRQY QQ&RQYG 

VHOIFRQY QQ&RQYG 

VHOIGHQVH QQ/LQHDU   

VHOIGHQVH QQ/LQHDU QXPBFODVVHV

GHI

IRUZDUG VHOI[ 

[ )UHOX VHOIFRQY [

[ )UHOX VHOIFRQY [

[ )PD[BSRROG [

[ )GURSRXW [

[ [YLHZ   

[ )UHOX VHOIGHQVH [

[ )GURSRXW [

[ VHOIGHQVH [

UHWXUQ

)ORJBVRIWPD[ [GLP 

Figure B-5.  Defining the model

appendix B   intro to pytorch

---

367

With that out of the way, you can define both the training and testing functions (see 

Figure 

B-7

 and Figure 

B-8

 for the training function, and Figure 

B-9

 and Figure 

B-10

 for 

the testing function).

Figure B-6.  The code in Figure 

B-5

 in a Jupyter cell

appendix B   intro to pytorch

---

368

GHI

WUDLQ PRGHOGHYLFHWUDLQBORDGHUFULWHULRQRSWLPL]HUHSRFK

VDYHBGLU PRGHOFNSW 

WRWDOBVWHS OHQ WUDLQBORDGHU

IRU

L LPDJHVODEHOV 

LQ HQXPHUDWH

WUDLQBORDGHU 

LPDJHV LPDJHVWR GHYLFH

ODEHOV  ODEHOVWR GHYLFH

)RUZDUGSDVV

RXWSXWV PRGHO LPDJHV

ORVV FULWHULRQ RXWSXWVODEHOV

%DFNZDUGDQGRSWLPL]H

RSWLPL]HU]HURBJUDG

ORVVEDFNZDUG

RSWLPL]HUVWHS

LI

L   

SULQW

(SRFK>^`^`@6WHS>^`^`@/RVV

^I` IRUPDW HSRFKQXPBHSRFKVLWRWDOBVWHSORVVLWHP

WRUFKVDYH PRGHOVWDWHBGLFW  S\WRUFKBPQLVWBFQQFNSW

Figure B-7.  The training algorithm. The for loop takes each pair of image and 

labels and passes them into the GPU as a tensor. They then go into the model, and 

the gradients are calculated. The information about the epoch and loss are then 

output

appendix B   intro to pytorch

---

369

The training function takes in the following parameters:

• 

model: An instance of a model class. In this case, it’s an instance of 

the CNN class defined above.

• 

device: This basically tells PyTorch what device (if the GPU is an 

option, which GPU to run on, and if not, the CPU is the device) to run 

on. In this case, you define the device right after the imports.

• 

train_loader: The loader for the training data set. In this case, you 

use a data_loader because that’s how the MNIST data is formatted 

when importing from torchvision. This data loader contains the 

training samples for the MNIST data set.

• 

criterion: The loss function to use. Define this before calling the train 

function.

• 

optimizer: The optimization function to use. Define this before 

calling the train function.

• 

epoch: What epoch is running. In this case, you call the training 

function in a for loop while passing in the iteration as the epoch.

The testing function is shown in Figure 

B-9

 and Figure 

B-10

.

Figure B-8.  The code in Figure 

B-7

 in a Jupyter cell

appendix B   intro to pytorch

---

370