1
00:00:00,750 --> 00:00:06,480
Hi and welcome to Chapter 15 point tree where we're about to build a flow of classify and we're going

2
00:00:06,480 --> 00:00:08,580
to use transfer learning to do this.

3
00:00:08,580 --> 00:00:10,280
So let's take a look at how we actually.

4
00:00:10,370 --> 00:00:13,230
What is Flora ossify our flower data set.

5
00:00:13,230 --> 00:00:14,920
I should say so.

6
00:00:15,060 --> 00:00:20,150
It comes from the Oxford University's visual geometry group as called 17.

7
00:00:20,310 --> 00:00:26,270
And that's because there are 17 categories of flowers and their images in each class say the sets and

8
00:00:26,270 --> 00:00:27,200
not that much.

9
00:00:28,110 --> 00:00:33,900
So this is some sample images from the flow Josephite the flowers 17 there is that this is the web page

10
00:00:33,900 --> 00:00:34,940
from Oxford University.

11
00:00:34,950 --> 00:00:40,170
And this is the link you can go to if you want to download it from day itself or you can use that link

12
00:00:40,230 --> 00:00:44,580
I have on the left here on the demi site panel.

13
00:00:44,640 --> 00:00:49,440
Please use that link to actually download it because I've already preprocess the data into a format

14
00:00:49,470 --> 00:00:54,330
that is easily imported into Karris if you downloaded from Oxford University site you're going to have

15
00:00:54,330 --> 00:00:55,760
to do a pre-processing itself.

16
00:00:55,770 --> 00:01:00,630
And I don't think if you're a beginner you're not going to find that fun at all although it's a good

17
00:01:00,630 --> 00:01:02,740
exercise to do sometimes.

18
00:01:03,600 --> 00:01:08,790
So anyway our approach to this problem is that we're going to actually use a pre-trained Fiji A16 model

19
00:01:09,540 --> 00:01:14,490
with all of its way it's frozen except the top layer and we're only going to train the top ahead of

20
00:01:14,490 --> 00:01:17,490
the model with a final output of 17 classes.

21
00:01:17,490 --> 00:01:21,370
So let's go back to our I and notebook and get this done.

22
00:01:21,710 --> 00:01:22,100
OK.

23
00:01:22,140 --> 00:01:24,750
So welcome back to our virtual machine.

24
00:01:24,780 --> 00:01:28,820
I hope you downloaded the flowers dataset and extracted it to this folder here.

25
00:01:29,040 --> 00:01:34,170
That's this folder called transfer linning and financing and Plaisted right here so we can quickly just

26
00:01:34,170 --> 00:01:37,910
inspect it taking a look at some of those pictures.

27
00:01:38,330 --> 00:01:42,120
Let's put it on toenail view and it looks quite nice.

28
00:01:42,120 --> 00:01:46,280
So as you can see we don't have that many images in this data set.

29
00:01:46,380 --> 00:01:51,380
So let's see what kind of accuracy we can get without transfer learning on the Viji model.

30
00:01:51,390 --> 00:01:53,380
So let's go to it here.

31
00:01:53,790 --> 00:02:02,170
So no let me just close some of these windows open and let's quickly go back to this one here so you

32
00:02:02,170 --> 00:02:03,350
can actually see how I do it.

33
00:02:03,360 --> 00:02:05,080
It's 15.

34
00:02:05,080 --> 00:02:07,090
And we go to making a flower classifier.

35
00:02:07,210 --> 00:02:08,440
That's this file here.

36
00:02:08,830 --> 00:02:10,260
So now that we're in the file.

37
00:02:10,300 --> 00:02:11,800
Let's take a look at what's going on.

38
00:02:11,800 --> 00:02:15,770
So we import the BTG model that's easily done here.

39
00:02:16,120 --> 00:02:23,470
Viji was designed to work open 24 or 224 by 224 pixel image input's Isiah's.

40
00:02:23,500 --> 00:02:26,450
So let's keep the standard size and go forward.

41
00:02:26,530 --> 00:02:32,200
So let's load the model with out his weights or with the weights of image's nuts without the top layer.

42
00:02:32,410 --> 00:02:34,360
I should say so we do that.

43
00:02:34,420 --> 00:02:36,960
And let's just print out the layers in this model.

44
00:02:37,060 --> 00:02:37,560
OK.

45
00:02:37,930 --> 00:02:44,740
So as you can see default actually is loaded here and by default all the layers are trainable.

46
00:02:44,740 --> 00:02:52,370
True that means the default in of when you load EGD all the weights are trainable.

47
00:02:52,630 --> 00:02:55,090
So we now have to set this true to false.

48
00:02:55,090 --> 00:02:56,490
So that's what we do here.

49
00:02:56,860 --> 00:03:03,010
So we loaded with our top head with Image net weights and we set all the treatable as we said this flag

50
00:03:03,090 --> 00:03:04,210
to false.

51
00:03:04,270 --> 00:03:08,030
So let's do this quickly and that's done there.

52
00:03:08,520 --> 00:03:13,450
And now let's create the function where we add a fully connected head.

53
00:03:13,510 --> 00:03:17,960
This is where we delay as we add now back to the top of our Viji that network.

54
00:03:18,190 --> 00:03:24,340
Notice this is different to the layers we added in the mobile network and that's because PDG has a different

55
00:03:24,340 --> 00:03:26,000
design to mobile and that.

56
00:03:26,020 --> 00:03:30,190
So you're going to have to look at the final design BTG and replace easily as here.

57
00:03:30,340 --> 00:03:35,700
And this here this densely a number of densely as dense units here.

58
00:03:36,190 --> 00:03:38,440
By default we are going to use 256.

59
00:03:38,440 --> 00:03:47,550
However this function allows us to specify it in here we can add 128 and it would be 128 units here.

60
00:03:47,890 --> 00:03:50,480
So let's leave the default right.

61
00:03:50,500 --> 00:03:57,220
And then you said drop out who said these things we input a number of classes which is 17 from the flow

62
00:03:57,220 --> 00:04:01,450
was data set 17 17 Sivam should make sense you know.

63
00:04:01,780 --> 00:04:04,730
And we just concatenated models here.

64
00:04:05,110 --> 00:04:08,800
Well the parts of the model to get the full model and then printed out.

65
00:04:08,800 --> 00:04:13,690
So let's take a look at and we see there 14 million parameters.

66
00:04:13,880 --> 00:04:18,150
It's less than between 19 and 16 sorry BTD 19.

67
00:04:18,440 --> 00:04:23,180
And with treatable parameters only 135 tells him that's quite good.

68
00:04:23,720 --> 00:04:25,060
So let me just run this.

69
00:04:25,130 --> 00:04:33,150
So we have fresh and no we just do it data generators here to deflower validation and Floetry unfold

70
00:04:33,250 --> 00:04:35,290
as we said our size.

71
00:04:35,320 --> 00:04:38,210
We can go actually just keep it at 16.

72
00:04:38,490 --> 00:04:38,910
All right.

73
00:04:38,950 --> 00:04:43,140
And keep going here.

74
00:04:43,260 --> 00:04:49,500
So now we declare all callbacks right here and we just create we create a callback array which we pass

75
00:04:49,500 --> 00:04:51,740
in here and let's run this now.

76
00:04:51,850 --> 00:04:55,430
So I to leave you to run this over and run this already.

77
00:04:55,450 --> 00:04:56,800
And it takes quite some time.

78
00:04:57,040 --> 00:05:01,540
But what I want you to observe is look at the validation accuracy in 25 books.

79
00:05:01,540 --> 00:05:06,230
The highest we get was actually 95 percent which is quite good.

80
00:05:06,820 --> 00:05:11,500
So you keep going see did it ever pass 95 tree at one time.

81
00:05:11,560 --> 00:05:12,990
So this is quite good.

82
00:05:13,240 --> 00:05:19,370
So we've got 95 percent accuracy using transfer linning using Viji 16 in translating.

83
00:05:19,630 --> 00:05:22,710
So let's keep going let's see what else we can do.

84
00:05:22,750 --> 00:05:24,080
OK.

85
00:05:24,430 --> 00:05:26,020
So this section here.

86
00:05:26,020 --> 00:05:27,620
Can we speed this up.

87
00:05:27,730 --> 00:05:31,060
So let's try resizing the images to 64 by 64.

88
00:05:31,200 --> 00:05:34,820
You remember it was assigned to a can 224 224.

89
00:05:34,910 --> 00:05:37,660
Now let's do this to 64.

90
00:05:37,930 --> 00:05:44,100
So let's use this comment to setting the input size.

91
00:05:44,100 --> 00:05:49,660
Now to.

92
00:05:49,780 --> 00:05:55,670
All right and do the standard thing where we load with image that way it's we don't include the top

93
00:05:55,780 --> 00:06:01,810
specified in U shape and we make the last train with three syllables.

94
00:06:02,190 --> 00:06:04,040
So that's good.

95
00:06:04,050 --> 00:06:07,050
And now let's move on to this.

96
00:06:07,460 --> 00:06:13,330
Let us actually start treating the small so as we can see this model has a different input sites.

97
00:06:14,180 --> 00:06:16,010
And let's see what we get.

98
00:06:16,010 --> 00:06:18,940
So I've trained this before so you don't have to do it.

99
00:06:18,950 --> 00:06:26,180
So what I want you to see though is that what what's happened here previously before actually did not

100
00:06:26,180 --> 00:06:30,130
used the callbacks or that's it in view but I should have thought it and I.

101
00:06:30,410 --> 00:06:32,490
But what I've done now is a discipline we do.

102
00:06:32,540 --> 00:06:41,660
So we see some callbacks feedback from stopping so we see it's not increasing monitoring patients is

103
00:06:41,660 --> 00:06:42,310
good.

104
00:06:42,320 --> 00:06:45,740
So at the end Epopt 12 is what we use.

105
00:06:45,770 --> 00:06:49,530
So let's go back to Iraq 12 pastorate ago.

106
00:06:49,920 --> 00:06:53,210
That's this one 82 percent.

107
00:06:53,230 --> 00:06:58,340
So 82 percent was our best loess validation loss and our best accuracy.

108
00:06:58,340 --> 00:07:06,500
So you can see by resizing the images a 64 by 64 which is a substantial decrease in size 2 to 24 by

109
00:07:06,500 --> 00:07:09,580
224 we got it into possessory.

110
00:07:09,860 --> 00:07:10,860
How much was it again.

111
00:07:11,520 --> 00:07:11,850
Sorry.

112
00:07:11,950 --> 00:07:13,930
82 percent accuracy.

113
00:07:14,060 --> 00:07:20,570
So that's not too bad to be fair actually sorry 86 percent accuracy we got that was fifteen point five

114
00:07:20,570 --> 00:07:22,150
six five two.

115
00:07:22,370 --> 00:07:22,730
Right.

116
00:07:22,730 --> 00:07:24,540
So that is actually this one.

117
00:07:25,010 --> 00:07:26,140
So yep.

118
00:07:26,150 --> 00:07:27,620
So this is good.

119
00:07:27,710 --> 00:07:29,960
It's not great but is pretty good.