1
00:00:00,650 --> 00:00:01,030
All right.

2
00:00:01,050 --> 00:00:07,920
And now we've got this fancy function create model what's happening here when we call model dot summary

3
00:00:08,070 --> 00:00:11,000
after we've created an instance of our model.

4
00:00:11,160 --> 00:00:13,840
We've got layout type Kerry's layout.

5
00:00:13,860 --> 00:00:17,910
So we go up here we've got Okay hub carriers layer that makes sense.

6
00:00:18,540 --> 00:00:25,260
We've got output shape multiple and program number five million four hundred thirty two thousand seven

7
00:00:25,260 --> 00:00:26,440
hundred thirteen.

8
00:00:26,460 --> 00:00:28,490
That's a fairly large number.

9
00:00:28,560 --> 00:00:35,980
When we come here total parameters that we have down here trainable prams and non trainable prams and

10
00:00:35,990 --> 00:00:43,750
it seems that the non trainable prams line up with this parameter number or parameter parameters sure

11
00:00:43,750 --> 00:00:44,860
for parameters.

12
00:00:45,200 --> 00:00:51,550
And we have another layer here called dense with param number one hundred twenty thousand turn of 40

13
00:00:52,340 --> 00:00:55,650
that corresponds to our dense layer here.

14
00:00:56,050 --> 00:00:56,810
Mm hmm.

15
00:00:57,290 --> 00:01:02,690
Well this is one of the amazing things in the benefits of transfer learning.

16
00:01:02,690 --> 00:01:10,910
The reason it says non trainable programs or parameters is because we're using mobile Net V2.

17
00:01:10,910 --> 00:01:19,280
And now if we come up here these are the patterns that mobile Net V2 has learned throughout these layers

18
00:01:19,280 --> 00:01:27,500
here or by repeatedly going through if we come back up to the top of this article by repeatedly going

19
00:01:27,500 --> 00:01:34,250
through our images or the images that mobile Net V2 was originally trained on better.

20
00:01:34,310 --> 00:01:36,370
That's the correct terminology there.

21
00:01:36,380 --> 00:01:38,710
So it's gone through all of these images.

22
00:01:39,050 --> 00:01:42,840
And if you're wondering what images did mobile Net V2 go through.

23
00:01:42,860 --> 00:01:55,040
Well it trained on if we read through here we go image net.

24
00:01:55,100 --> 00:01:55,790
There we go.

25
00:01:56,180 --> 00:02:04,810
So it's whites were originally obtained by training on the I L SVR RC 2012 CLSA data set for image classification.

26
00:02:04,880 --> 00:02:06,260
Image net.

27
00:02:06,260 --> 00:02:06,720
Mm hmm.

28
00:02:06,880 --> 00:02:16,840
Okay let's look up that image net they mentioned it image net is an image database organized according

29
00:02:16,840 --> 00:02:18,940
to the word knit hierarchy.

30
00:02:18,970 --> 00:02:20,870
Okay.

31
00:02:20,950 --> 00:02:24,680
Currently we have an average of over 500 images per node.

32
00:02:24,880 --> 00:02:25,620
OK.

33
00:02:25,630 --> 00:02:33,280
So essentially image net is a large database of 14 over 14 million images.

34
00:02:33,350 --> 00:02:41,250
And so that these patterns these parameters are what mobile Net has learned by training on image net.

35
00:02:42,170 --> 00:02:49,910
And so this is the beautiful thing about transfer learning is that all these patents that mobile Net

36
00:02:49,910 --> 00:02:53,050
has learned through these layers here and by the way.

37
00:02:53,060 --> 00:02:59,450
This is why they call it deep learning because essentially each one of these layers is a layer in our

38
00:02:59,450 --> 00:03:00,830
neural network.

39
00:03:00,830 --> 00:03:09,050
So if you imagine this expanded out each one of them is a layer and hence the reason they call it deep

40
00:03:09,050 --> 00:03:13,370
is because the diagrams are usually they have an input layer and then it just keeps going down and down

41
00:03:13,370 --> 00:03:14,850
and down and down and down.

42
00:03:15,140 --> 00:03:23,490
And so inherently what each of these layers actually are is a small model within themselves.

43
00:03:24,080 --> 00:03:29,470
So these are just multiple pattern finding models layered on top of each other.

44
00:03:29,570 --> 00:03:32,000
Hence the term deep learning.

45
00:03:32,000 --> 00:03:40,410
So for come back this is what our base layer is those five and a half million patterns that mobile Net

46
00:03:40,410 --> 00:03:48,600
V2 has found within the image net and now we get to utilize those patterns that it's found in our own

47
00:03:48,600 --> 00:03:49,480
problem.

48
00:03:49,500 --> 00:03:55,590
So rather than try and find them from scratch on our own which could take hours and lots of compute

49
00:03:55,590 --> 00:04:01,750
power that we don't really have we get to utilize them and train our own.

50
00:04:01,750 --> 00:04:07,990
One hundred and twenty thousand parameters so this is out our dense layer here.

51
00:04:08,100 --> 00:04:12,880
That's what we're going to be training so we're going to be passing out images to mobile Net.

52
00:04:12,880 --> 00:04:17,730
It's going to be using its baseline patents to figure out what's in our images.

53
00:04:17,860 --> 00:04:24,400
And then finally we're going to stick this layer on top and go hey we've got our own data data that

54
00:04:24,430 --> 00:04:30,790
isn't an image net but let's use what you figured out in image net for classifying which dog breed is

55
00:04:30,790 --> 00:04:32,220
in a certain photo.

56
00:04:32,650 --> 00:04:37,900
Now that we have a model we probably need to make some callbacks and you might be wondering what are

57
00:04:37,900 --> 00:04:38,680
callbacks.

58
00:04:38,680 --> 00:04:47,230
Well let's save that for the next video where we create some callbacks and then after that we'll create

59
00:04:47,230 --> 00:04:50,130
a function that's going to train our model.

60
00:04:50,590 --> 00:04:51,880
That sounds like a good idea to me.