1
00:00:00,330 --> 00:00:05,650
In this lesson we're going to poke around a bit more and start exploring our data.

2
00:00:05,850 --> 00:00:08,370
So far we formulated our question.

3
00:00:08,460 --> 00:00:09,570
We know what we want to do.

4
00:00:09,600 --> 00:00:15,780
We set out to classify images according to ten different categories or classes.

5
00:00:15,780 --> 00:00:19,470
We've also gathered data and very very handily.

6
00:00:19,470 --> 00:00:22,950
They've already been split into training and testing data.

7
00:00:22,950 --> 00:00:30,210
Now we're getting to this data cleaning and processing and exploration stage back in our Jupiter notebook.

8
00:00:30,440 --> 00:00:38,110
Let's insert a markdown cell here explore the data and the first thing I want to do is I want to pull

9
00:00:38,110 --> 00:00:39,590
up some of these pictures.

10
00:00:39,670 --> 00:00:50,290
So when I go back up to my inputs and I want input to things I want to see from eye Python dot display

11
00:00:50,740 --> 00:01:05,720
import display and I want to see from Charisse dot pre processing dot image import array to image I

12
00:01:05,810 --> 00:01:12,910
AMG and you know what I'm actually going to input a third thing I'm going to input that plot lib so

13
00:01:12,910 --> 00:01:20,800
I'll add import matter plot lib dot pie plot as P.L. T.

14
00:01:21,010 --> 00:01:22,500
And just below a lot.

15
00:01:22,500 --> 00:01:23,800
Percent match.

16
00:01:23,950 --> 00:01:29,820
Plot lib space inline shift enter.

17
00:01:30,140 --> 00:01:36,760
And now let me come down here and let's throw some of these images up on screen and take a closer look

18
00:01:36,760 --> 00:01:37,060
at them.

19
00:01:37,180 --> 00:01:42,970
So the reason I imported I Python display and plot lib is because I wanted to show you two different

20
00:01:42,970 --> 00:01:47,570
ways that you can display an image in Jupiter notebook.

21
00:01:47,710 --> 00:01:51,660
So the first one is with a python display.

22
00:01:51,880 --> 00:01:58,020
And what I'm going to do is actually look at the very first item in our training dataset.

23
00:01:58,120 --> 00:02:03,010
So X train underscore all square brackets zero.

24
00:02:03,150 --> 00:02:10,230
It's the very first item and shift enter will actually show us what this image looks like at the moment.

25
00:02:10,360 --> 00:02:17,680
And if we wanted to see this whole thing as an actual image then we're going to make use of this array

26
00:02:17,710 --> 00:02:22,840
to image functionality from Chris's pre processing module.

27
00:02:22,870 --> 00:02:31,120
So what we can do is store that image in a variable called pick and that will be equal to array to image

28
00:02:31,390 --> 00:02:36,090
parentheses and then we can pick one of the images from our training dataset.

29
00:02:36,160 --> 00:02:37,480
See we picked the.

30
00:02:37,810 --> 00:02:41,410
I don't know seventh image using a python display.

31
00:02:41,410 --> 00:02:49,600
We can then throw that up on the screen with display parentheses and then supply our picture as an argument.

32
00:02:49,600 --> 00:02:53,660
So this is what the seventh picture will look like now.

33
00:02:53,680 --> 00:03:00,220
I hope you can tell it this is a very tiny picture of a horse and we should be able to verify that this

34
00:03:00,220 --> 00:03:01,670
is indeed a horse.

35
00:03:01,720 --> 00:03:05,770
By looking at the image labels which are found in the y values.

36
00:03:05,890 --> 00:03:10,300
So why honest quatrain on a scroll all don't shape.

37
00:03:10,360 --> 00:03:13,650
Well show us what are y values actually look like.

38
00:03:13,860 --> 00:03:21,790
All right we've got a number higher rate of 50000 labels and it's two dimensional.

39
00:03:21,790 --> 00:03:29,740
So if I wanted to pull up the y value at index 7 then I would write y on as Katrina and a square all

40
00:03:30,160 --> 00:03:35,260
square brackets 7 and then another square brackets 0.

41
00:03:35,500 --> 00:03:40,210
And what we get is the values 7 so it doesn't say horse.

42
00:03:40,210 --> 00:03:42,220
It just says 7.

43
00:03:42,220 --> 00:03:43,690
Now why is that.

44
00:03:43,690 --> 00:03:45,700
Well let's pull up Alex's website again.

45
00:03:46,180 --> 00:03:51,250
So if I look at his little breakdown here I think this is really handy.

46
00:03:51,250 --> 00:03:52,520
These are in order right.

47
00:03:52,570 --> 00:04:02,550
I can see that values 7 should correspond to 0 1 2 3 4 5 6 7 corresponds to a horse.

48
00:04:02,560 --> 00:04:07,450
So what we've got here are the names for these different labels.

49
00:04:07,570 --> 00:04:14,490
Label number 0 will be our plain label number 7 will be horse label number 9 will be truck.

50
00:04:14,530 --> 00:04:17,430
Now it's no good just having this on this website.

51
00:04:17,470 --> 00:04:20,850
I think what we should do is actually add these to our notebook.

52
00:04:21,010 --> 00:04:31,300
So up here I'm going to add another markdown cell and I'm going to add some constants here so constants.

53
00:04:31,810 --> 00:04:38,770
And among those constants will be this list of names because then we can easily refer to it later on.

54
00:04:38,770 --> 00:04:43,260
So I'm just type label underscore names.

55
00:04:43,600 --> 00:04:51,200
So here will create a list that contains all the label names in order plane car bird cat Dear Dog frog.

56
00:04:51,250 --> 00:04:56,670
Horse ship and truck shift enter on that.

57
00:04:57,350 --> 00:05:04,610
We can scroll back down and what we should be able to do is hit up our label names list open some square

58
00:05:04,610 --> 00:05:12,940
brackets and we can put the label at the index number seven in here and I'll paste that right in it

59
00:05:12,950 --> 00:05:15,970
shift into Astra get horse.

60
00:05:16,120 --> 00:05:16,450
OK.

61
00:05:16,460 --> 00:05:16,930
Brilliant.

62
00:05:17,210 --> 00:05:18,460
So that makes sense to me.

63
00:05:18,710 --> 00:05:24,730
But I think we should look at another image just to double check it works and look at this image.

64
00:05:24,770 --> 00:05:30,890
Let's use matte plot lib so we can actually use matte plot lib and through the image onto a plot because

65
00:05:30,890 --> 00:05:35,510
this image is actually just an array of pixels matte plot lib knows how to handle this.

66
00:05:35,720 --> 00:05:45,860
So with Piti dot I am show parentheses x underscore a train underscore all and then say square brackets

67
00:05:46,520 --> 00:05:54,940
and then let's pick the fifth image at index for close the parentheses and then on the next line will

68
00:05:54,940 --> 00:05:58,920
at Peel T dot show and hit shift enter.

69
00:05:59,210 --> 00:06:01,890
Let's throw this image up on screen.

70
00:06:02,030 --> 00:06:07,500
So what we've got here is a chart every five pixels has got a tick mark.

71
00:06:07,760 --> 00:06:13,740
And for our convenience this is a slightly enlarged but it's definitely a car right.

72
00:06:13,790 --> 00:06:16,580
We can verify this by checking our list of label names right.

73
00:06:16,820 --> 00:06:24,500
So with Peel T dot X label we can look towards our list of labels and then insert the label on the x

74
00:06:24,500 --> 00:06:25,590
axis.

75
00:06:25,640 --> 00:06:35,020
So peel teed up X label parentheses label names square brackets and then we'll use the same technique

76
00:06:35,230 --> 00:06:35,890
as above.

77
00:06:35,890 --> 00:06:36,510
Right.

78
00:06:36,610 --> 00:06:43,180
We're going to look to our y underscore train underscore all umpiring could put some square brackets

79
00:06:43,180 --> 00:06:46,010
there when you use the same index in x4.

80
00:06:46,270 --> 00:06:51,730
And we're going to use another pair square brackets and then zero because we're quite a two dimensional

81
00:06:52,000 --> 00:06:53,850
num pi array.

82
00:06:54,130 --> 00:07:00,640
So I know it's a lot of nesting here but we're pulling out a value from our y unescorted train underscore

83
00:07:00,640 --> 00:07:08,920
all and we're using that as the index in our list of label names and we're gonna display that on our

84
00:07:08,920 --> 00:07:10,350
x axis.

85
00:07:10,690 --> 00:07:15,730
And if this works we should see the word card written here.

86
00:07:16,450 --> 00:07:20,450
If this is too small can of course enlarge the font size.

87
00:07:20,590 --> 00:07:28,370
So an extra argument here and font size 15 will make this much much more legible.

88
00:07:28,420 --> 00:07:30,250
So this is pretty cool right.

89
00:07:30,250 --> 00:07:39,160
We've got our images that are stored as Num higher arrays in our data set and we've got two ways of

90
00:07:39,280 --> 00:07:42,800
displaying them in our Jupiter notebook.

91
00:07:42,910 --> 00:07:46,430
The first way is using a python display.

92
00:07:46,990 --> 00:07:51,590
And the second way is actually throwing them up onto a mat plot lib chart.

93
00:07:51,610 --> 00:07:57,490
Now I quite like Matt put lib actually one of the reasons is that's very easy to display these images

94
00:07:57,790 --> 00:08:02,550
side by side and just take a look at a few more from this dataset.

95
00:08:02,560 --> 00:08:06,190
So I might ask you to solve this as a challenge.

96
00:08:06,190 --> 00:08:13,910
Can you write a for loop to display the first 10 images from the X underscore train underscore all the

97
00:08:13,910 --> 00:08:15,310
data set.

98
00:08:15,340 --> 00:08:22,420
Remember this is a number primary and I'd like you to display all these images in a row and I'd like

99
00:08:22,420 --> 00:08:24,750
you to show the label below.

100
00:08:24,760 --> 00:08:31,450
Each image and if you can format the images in such a way that we're not seeing any of these tick marks

101
00:08:31,450 --> 00:08:34,310
here like we are in the cell above.

102
00:08:34,390 --> 00:08:37,900
I'll give you a few seconds to pause the video and give this a go

103
00:08:41,620 --> 00:08:42,340
ready.

104
00:08:42,340 --> 00:08:44,310
Here's the solution.

105
00:08:44,380 --> 00:08:49,510
The first thing I'm going to do is an asset the size of my chart or of my figure.

106
00:08:49,870 --> 00:09:00,810
So PDT don't figure parentheses fixed size equals parentheses 15 by 5.

107
00:09:00,820 --> 00:09:05,900
Well size my chart nicely and then I'm going to write this for loop.

108
00:09:05,950 --> 00:09:09,850
For i in range 10.

109
00:09:09,880 --> 00:09:16,490
This for loop will go through the numbers 0 through 9 and I can show you this if I print I here.

110
00:09:17,500 --> 00:09:24,270
But inside my loop what I want to do is actually create some subplots using my plot lib.

111
00:09:24,640 --> 00:09:35,140
This is what the challenge was about so I'll say BLT dot subplot and then my subplot will require three

112
00:09:35,140 --> 00:09:36,530
inputs right.

113
00:09:36,730 --> 00:09:46,180
One how many rows 10 how many columns and then because we're in a for loop we're going to have to figure

114
00:09:46,180 --> 00:09:49,650
out which subplot we want to create at each iteration.

115
00:09:49,660 --> 00:09:56,350
So in this case it's going to be I plus one because the first plot should be number one and the last

116
00:09:56,350 --> 00:10:00,440
subplot should be number 10 but I starts from zero.

117
00:10:00,460 --> 00:10:01,840
It goes up until 9.

118
00:10:01,960 --> 00:10:13,280
So this is why if I plus one and then we'll use PDT dot I am show parentheses x underscore train underscore

119
00:10:13,360 --> 00:10:16,440
all square brackets.

120
00:10:16,440 --> 00:10:27,100
I let's see what this gives us what we see here is 10 images in a row in a figure that assigns 15 by

121
00:10:27,160 --> 00:10:29,660
fives and number seven here.

122
00:10:29,680 --> 00:10:36,330
So of course our horse and the image at index number four is our call to remove those tick marks.

123
00:10:36,340 --> 00:10:47,830
We can do that with peeled T dot y ticks parentheses square brackets and peel T dot X takes parentheses

124
00:10:48,040 --> 00:10:51,320
square brackets so square brackets are empty.

125
00:10:51,580 --> 00:10:55,240
So this means that my tick marks will be blank.

126
00:10:55,240 --> 00:10:59,860
Now all I have to do is insert those labels below the images.

127
00:10:59,860 --> 00:11:05,830
So I want to do that with peeled T dot X label and then want to use the very same technique I've got

128
00:11:05,830 --> 00:11:06,510
above.

129
00:11:06,700 --> 00:11:14,450
I want to take my list of label names and in the square brackets I'm going to look towards my dataset

130
00:11:14,860 --> 00:11:23,500
of training labels y on a squat train underscore all square brackets and again I for the index and square

131
00:11:23,500 --> 00:11:27,460
brackets zero for that second dimension.

132
00:11:27,460 --> 00:11:30,120
Let me hit shift enter on this.

133
00:11:30,640 --> 00:11:31,810
And there we go.

134
00:11:31,810 --> 00:11:36,910
There are our tiny little labels underneath our tiny little images.

135
00:11:37,030 --> 00:11:39,530
And because that's very hard to read.

136
00:11:39,760 --> 00:11:44,470
I'll increase the font size once again to say maybe 14.

137
00:11:44,500 --> 00:11:47,050
Try that.

138
00:11:47,050 --> 00:11:47,500
There we go.

139
00:11:47,500 --> 00:11:49,450
That's a lot more legible.

140
00:11:49,450 --> 00:11:50,570
Brilliant.

141
00:11:50,590 --> 00:11:58,910
So what we see here is that our list of a label names matches up nicely with the numbers in the y underscore

142
00:11:58,910 --> 00:12:03,850
train underscore all of an umpire Ray seems to be indeed a bird.

143
00:12:03,910 --> 00:12:08,190
And this seems to be a ship and we've got a cat and a deer.

144
00:12:08,590 --> 00:12:13,220
I think this sort of evil looking red eyed frog is a definitely a keeper.

145
00:12:13,850 --> 00:12:20,800
But yeah what we're seeing here are indeed the first 10 images in our training dataset.

146
00:12:20,950 --> 00:12:28,770
So the next thing to understand about this data set is how these images are formatted so x underscore

147
00:12:28,780 --> 00:12:32,450
train underscore all square brackets zero.

148
00:12:32,560 --> 00:12:43,720
That frog is formatted like so but if we look at the shape of it we see that it is 32 by 32 by 3.

149
00:12:43,750 --> 00:12:56,040
Now this is interesting because 32 and 32 are the width and the height and three is the number of channels.

150
00:12:56,050 --> 00:12:59,770
So when I say channels I'm talking about colors.

151
00:12:59,770 --> 00:13:01,180
We've got a red channel.

152
00:13:01,210 --> 00:13:04,220
We've got a green channel and we've got a blue channel.

153
00:13:04,570 --> 00:13:08,440
So this is gonna be in our GP format.

154
00:13:08,440 --> 00:13:13,390
If we look at our number higher rate as a whole so X and it's Katrina on the square all that shape we

155
00:13:13,390 --> 00:13:16,330
get to see that this thing has actually four dimensions.

156
00:13:16,330 --> 00:13:16,960
Right.

157
00:13:16,990 --> 00:13:26,800
So we've got 50000 entries and each of these entries is 32 pixels by 32 pixels by three.

158
00:13:26,800 --> 00:13:28,150
Let's make this more explicit.

159
00:13:28,150 --> 00:13:28,450
Right.

160
00:13:28,480 --> 00:13:32,660
So let's see number of images comma X come out.

161
00:13:32,710 --> 00:13:33,620
Why comma.

162
00:13:33,670 --> 00:13:39,550
C is equal to X and a quatrain underscore all that shape.

163
00:13:39,550 --> 00:13:46,630
If we do this then we can add a nice little print statement here below with an F string so print open

164
00:13:46,630 --> 00:13:56,780
parentheses f single quotes and then we can write images is equal to curly braces.

165
00:13:56,970 --> 00:14:01,230
Number of images no underscore images.

166
00:14:01,310 --> 00:14:09,450
And then we can insert some formatting so backslash T for a tab and then a pipe maybe and then we'll

167
00:14:09,450 --> 00:14:15,280
see width is equal to curly braces.

168
00:14:15,390 --> 00:14:24,860
X backslash t pipe height is equal to curly braces.

169
00:14:24,980 --> 00:14:33,510
Y backslash t channels is equal to curly braces.

170
00:14:33,730 --> 00:14:40,480
See me at that pipe symbol here and hit shift center.

171
00:14:40,540 --> 00:14:43,600
I think this is a really nice thing to have in a notebook.

172
00:14:43,600 --> 00:14:49,720
If you're going to take a break and come back to it later on to be reminded of the structure of the

173
00:14:49,720 --> 00:14:55,000
data all in all we've got 50000 training images.

174
00:14:55,000 --> 00:14:58,750
We've got a width in pixels of 32.

175
00:14:58,750 --> 00:15:00,850
So this is our x value.

176
00:15:00,850 --> 00:15:04,240
We've got a height in pixels of 32.

177
00:15:04,240 --> 00:15:09,310
This is our y value and we've got three color channels per pixel.

178
00:15:09,340 --> 00:15:16,550
So we've got three values per pixel now considering we've got 50000 training images.

179
00:15:16,560 --> 00:15:23,840
How many testing images do we have on those for test that shape and shift into shows us that we've got

180
00:15:24,150 --> 00:15:33,590
10000 testing images now that we know what our dataset consists of and how our information is formatted.

181
00:15:33,650 --> 00:15:38,360
It's time to pre process our data for our neural network.

182
00:15:38,360 --> 00:15:44,420
Considering that we loaded this data directly from cameras there's actually not much need for cleaning

183
00:15:44,870 --> 00:15:50,210
but we will do some pre processing for our neural network to make our lives a little easier.

184
00:15:50,560 --> 00:15:54,730
And we're going to cover all that and more in the next lesson.

185
00:15:54,740 --> 00:15:55,520
I'll see you there.