1
00:00:00,000 --> 00:00:02,280
Did you find a solution? Well, of

2
00:00:02,280 --> 00:00:03,960
course yours might be
different from mine but

3
00:00:03,960 --> 00:00:06,030
let me show you what
I did in the case

4
00:00:06,030 --> 00:00:08,505
of the cat my model
thought was a dog.

5
00:00:08,505 --> 00:00:11,220
So let's go back to the notebook,
and we'll run the code.

6
00:00:11,220 --> 00:00:13,920
I'll upload this image to
see how it classifies.

7
00:00:13,920 --> 00:00:15,375
It's a crop of the cats,

8
00:00:15,375 --> 00:00:16,605
and lo and behold,

9
00:00:16,605 --> 00:00:18,360
it classifies as a cat.

10
00:00:18,360 --> 00:00:21,525
Let's open it, and compare
it to the original image,

11
00:00:21,525 --> 00:00:23,955
and we'll see that
just by cropping I was

12
00:00:23,955 --> 00:00:26,550
able to get it to change
its classification.

13
00:00:26,550 --> 00:00:28,020
There must have been something in

14
00:00:28,020 --> 00:00:31,340
the uncropped image that
matched features with a dog.

15
00:00:31,340 --> 00:00:32,330
Now I thought that was

16
00:00:32,330 --> 00:00:34,490
a very interesting
experiment, didn't you?

17
00:00:34,490 --> 00:00:36,530
Now what do you
think the impact of

18
00:00:36,530 --> 00:00:38,770
cropping might've
had on training?

19
00:00:38,770 --> 00:00:41,240
Would that have trained
the model to show that this

20
00:00:41,240 --> 00:00:43,715
was a cat better than
an uncropped image.

21
00:00:43,715 --> 00:00:45,860
That's food for thought,
and something to explore in

22
00:00:45,860 --> 00:00:49,770
the next lesson but first
let's go back to the workbook.