Beginning Anomaly Detection Using

IURP

VNOHDUQPHWULFV

LPSRUW

URFBDXFBVFRUH

GHI

WHVW PRGHOGHYLFHWHVWBORDGHU 

SUHGV >@

\BWUXH >@

7HVWWKHPRGHO

PRGHOHYDO 6HWPRGHOWRHYDOXDWLRQPRGH

ZLWK

WRUFKQRBJUDG 

FRUUHFW 

WRWDO  

IRU

LPDJHVODEHOV

LQ

WHVWBORDGHU

LPDJHV LPDJHVWR GHYLFH

ODEHOV ODEHOVWR GHYLFH

RXWSXWV PRGHO LPDJHV

BSUHGLFWHG WRUFKPD[ RXWSXWVGDWD

WRWDO ODEHOVVL]H 

FRUUHFW  SUHGLFWHG  ODEHOV VXP LWHP 

GHWDFKHGBSUHG SUHGLFWHGGHWDFK FSX QXPS\

GHWDFKHGBODEHO ODEHOVGHWDFK FSX QXPS\

IRU

I

LQ UDQJH

OHQ GHWDFKHGBSUHG 

SUHGVDSSHQG GHWDFKHGBSUHG>I@

\BWUXHDSSHQG GHWDFKHGBODEHO>I@

SULQW

7HVW$FFXUDF\RIWKHPRGHORQWKHWHVWLPDJHV

^` IRUPDW FRUUHFWWRWDO

SUHGV QSH\H QXPBFODVVHV >SUHGV@

\BWUXH QSH\H QXPBFODVVHV >\BWUXH@

DXF URFBDXFBVFRUH SUHGV\BWUXH

SULQW

$8&^`IRUPDW DXF

Figure B-9.  The code for the testing algorithm. Once again, the for loop takes 

the image and label pairs and passes them through the model to get a prediction. 

Then, once every pair has a prediction, the AUC score is calculated

appendix B   intro to pytorch

---

371

Notice that you use the AUC score as part of the testing metric. You don’t have to do 

this, but it might be a better indicator of the model’s performance than plain accuracy, so 

it was included in this example.

The parameters the model takes in are

•