1
00:00:00,510 --> 00:00:03,120
Now we've seen what an image looks like intensive form.

2
00:00:03,360 --> 00:00:07,990
Let's build this little function that's gonna do these steps here.

3
00:00:08,040 --> 00:00:13,940
So what we might do is just copy them so we can see them down below.

4
00:00:14,020 --> 00:00:18,480
Am I gonna just write a little sell here.

5
00:00:18,480 --> 00:00:34,640
Now we've seen what an image looks like as a tensor let's make a function to pre process them already.

6
00:00:34,740 --> 00:00:42,610
And so what we might do is define a variable here image size to 24 and you might be wondering like I

7
00:00:42,610 --> 00:00:48,790
always do when I see things like this that aren't explained what is 224 whether Look get that will we're

8
00:00:48,790 --> 00:00:54,440
going to see that in a later video when we pick out a machine learning model that we can use.

9
00:00:54,550 --> 00:01:00,130
And so a lot of machine learning is lining up the size of your inputs or the shape of your inputs in

10
00:01:00,130 --> 00:01:03,880
our case images with the model that you're using.

11
00:01:03,880 --> 00:01:09,730
Again we'll cover this later on but essentially we're making sure that our images that we feed into

12
00:01:09,730 --> 00:01:17,620
a machine learning model are of this shape so to 24 high to 24 wide because that is what the machine

13
00:01:17,620 --> 00:01:22,080
learning model that we're going to use which trained on images in that same shape.

14
00:01:22,090 --> 00:01:28,810
Remember how it's important to feed data in the same format and shape to your machine learning model

15
00:01:29,170 --> 00:01:31,830
as it was trained on when you're trying to make predictions.

16
00:01:31,900 --> 00:01:33,600
But we'll see that in practice later on.

17
00:01:34,330 --> 00:01:35,620
But let's write this function.

18
00:01:35,620 --> 00:01:41,850
So create a function for pre processing images.

19
00:01:42,050 --> 00:01:46,950
So I call this nice and simple process image.

20
00:01:46,960 --> 00:01:51,730
It's going to take an image path as input and

21
00:01:54,430 --> 00:02:06,340
little doctoring could be takes an image file path and turns the image into a tensor something simple

22
00:02:06,340 --> 00:02:12,070
like that we could make it a little bit more intuitive later on but to begin with that's enough to explain

23
00:02:12,190 --> 00:02:13,680
it for ourselves.

24
00:02:13,720 --> 00:02:21,120
So we want to read in an image file and to do that we'll save it to a variable called image.

25
00:02:21,490 --> 00:02:27,790
And if we read through loading images that you can do this for other types of data to write there's

26
00:02:27,790 --> 00:02:29,000
text data there.

27
00:02:29,050 --> 00:02:35,740
This pan this data frame num pi CSA but reading through there we see that we can go t after Io for input

28
00:02:35,740 --> 00:02:38,130
output read file.

29
00:02:38,260 --> 00:02:39,850
There we go.

30
00:02:39,970 --> 00:02:40,750
What's this say.

31
00:02:40,870 --> 00:02:43,900
Reads and outputs the entire contents of the input file name.

32
00:02:43,990 --> 00:02:45,170
Well that makes sense.

33
00:02:45,310 --> 00:02:48,040
Returns a tensor of type String.

34
00:02:48,040 --> 00:02:49,270
Beautiful.

35
00:02:49,420 --> 00:02:52,600
Let's do image path but don't take my word for it.

36
00:02:52,810 --> 00:02:56,800
What we'll do is and I want you to adopt this practice too.

37
00:02:57,280 --> 00:03:01,430
If you ever see a function that you're not entirely sure of what it does.

38
00:03:01,570 --> 00:03:04,880
Rewrite it line by line so let's check this out.

39
00:03:05,170 --> 00:03:15,220
We'll go tensor equals t after Io read read file and then we need an image path.

40
00:03:15,220 --> 00:03:17,070
So what do file names.

41
00:03:17,170 --> 00:03:23,100
I'm 26 so we'll get file names 26 and what's that look like

42
00:03:27,070 --> 00:03:27,550
all.

43
00:03:28,280 --> 00:03:35,880
So it's shape of a string and then a whole bunch of characters I don't understand but that's an alright

44
00:03:35,930 --> 00:03:38,800
thing because the computer understands them better than I do.

45
00:03:38,810 --> 00:03:47,570
So what it's done is it just taken the file name and turned the contents into this long tensor.

46
00:03:47,960 --> 00:03:48,880
Wonderful.

47
00:03:48,920 --> 00:03:49,900
So the next step.

48
00:03:50,080 --> 00:03:52,940
Use tens flow to read the file save it to image.

49
00:03:52,940 --> 00:03:54,100
Yep.

50
00:03:54,190 --> 00:03:56,820
To now image a j peg into tenses.

51
00:03:56,840 --> 00:03:57,320
Okay.

52
00:03:57,470 --> 00:03:59,120
So let's do that.

53
00:03:59,120 --> 00:04:13,360
Turn the j peg image into numerical tensor with three color channels Red Rain blue

54
00:04:16,450 --> 00:04:23,260
and again checking out the loading images documentation so we're reusing this variable here.

55
00:04:23,290 --> 00:04:32,220
Now we're going to reset it to B T after image dot decode see how we've got some options here we go

56
00:04:32,270 --> 00:04:39,100
to PJ which is the type of image file we're gonna use J peg because our images are in j peg image and

57
00:04:39,100 --> 00:04:46,220
then we pass it the variable channels the outer channels indicates the desired number of coloured channels

58
00:04:46,220 --> 00:04:54,090
for the decoded image in our case we want three for red green on blue beautiful.

59
00:04:54,320 --> 00:05:02,850
Now again if we wanted to check out exactly what this is doing let's take tensor and go after the image

60
00:05:03,310 --> 00:05:10,410
that decode j peg we'll take tensor and then we'll go channels equals three

61
00:05:13,800 --> 00:05:23,140
Oh look at that you've got exactly the same sort of data as that we have up here wonderful right now

62
00:05:23,140 --> 00:05:27,640
to save in the interest of time we're going to write a couple more lines in this function but we won't

63
00:05:27,700 --> 00:05:31,600
go through these similar steps that's what I want I want you to have some practice doing.

64
00:05:31,600 --> 00:05:37,480
If you do see a function and you're curious to what it does try writing it out line by line and see

65
00:05:37,480 --> 00:05:39,150
what happens.

66
00:05:39,190 --> 00:05:41,680
So you're just gonna have to trust me with what these lines do.

67
00:05:42,100 --> 00:05:50,300
So convert the color channel values so this is called normalization I'll write the code first and then

68
00:05:50,300 --> 00:05:58,580
tell you about it from 0 to 255 to 0 2 1 values.

69
00:05:58,580 --> 00:06:10,140
So we want to go image equals t after image convert image D type image TAF float 32.

70
00:06:10,160 --> 00:06:13,340
Now what's the doctoring say here.

71
00:06:13,550 --> 00:06:17,930
Convert image to day type scaling its values if needed.

72
00:06:17,930 --> 00:06:24,050
Images that are represented using floating point values are expected to have values in the range 0 1.

73
00:06:24,050 --> 00:06:25,210
Wonderful.

74
00:06:25,220 --> 00:06:27,980
Now you might be thinking how how do you know this.

75
00:06:27,980 --> 00:06:36,430
Daniel how do you know that you have to convert it from 0 to 255 to between 0 1 Well actually I lied.

76
00:06:36,670 --> 00:06:39,580
I said we weren't going to look at that but this is a good one to investigate

77
00:06:42,100 --> 00:06:51,970
tensor just so you can see we've got values there between 0 and 255 and now I'll run this Oh no I don't

78
00:06:51,970 --> 00:07:01,900
actually want that messing myself up here rerun this rerun this so just when I demonstrate what this

79
00:07:01,900 --> 00:07:05,580
is doing so image convert.

80
00:07:06,000 --> 00:07:12,660
So basically all it's doing is it converts our values.

81
00:07:12,660 --> 00:07:15,640
It scales them instead of being between 0 and 255.

82
00:07:15,690 --> 00:07:20,490
They now are between 0 and 1 with the same scale float 32

83
00:07:26,340 --> 00:07:27,650
TFT float 32

84
00:07:37,650 --> 00:07:38,430
wonderful.

85
00:07:38,430 --> 00:07:39,890
So we've got the same shape.

86
00:07:39,960 --> 00:07:46,920
But now instead of being between zero activity five between zero one and now this type of process makes

87
00:07:46,920 --> 00:07:49,350
computing on these more efficient.

88
00:07:49,500 --> 00:07:52,190
That's the most you need to know it's called normalization.

89
00:07:52,200 --> 00:07:56,980
And again this is just straight from the loading images documentation.

90
00:07:56,990 --> 00:07:58,750
Now what do we have to do.

91
00:07:58,760 --> 00:08:09,060
Actually I don't think we had a step here for that we might have to put in for normalize our image convert

92
00:08:10,480 --> 00:08:17,550
color channel values from 0 to 255 to 0 to 1

93
00:08:21,320 --> 00:08:23,510
5 6.

94
00:08:23,600 --> 00:08:24,370
OK.

95
00:08:24,470 --> 00:08:29,270
Now finally we can resize the image now by now.

96
00:08:30,170 --> 00:08:36,290
I believe you trust me and what this line is going to do so we not not going to type it out.

97
00:08:36,290 --> 00:08:40,640
I'm going to leave that to you two twenty four to twenty four.

98
00:08:40,640 --> 00:08:44,090
So basically a square exactly a square actually

99
00:08:47,020 --> 00:08:52,630
image dot resize resize as images to size using the specified method.

100
00:08:52,630 --> 00:08:53,110
OK.

101
00:08:53,140 --> 00:09:01,770
And now we want to resize image and the size is going to be height width.

102
00:09:01,780 --> 00:09:04,190
Now you see why we define this here.

103
00:09:04,330 --> 00:09:11,890
Maybe that if we wrote this function a bit better we could input image size as image size

104
00:09:15,010 --> 00:09:16,030
that's probably a little bit better.

105
00:09:17,620 --> 00:09:18,190
Wonderful.

106
00:09:18,190 --> 00:09:21,880
And now you'll get this kind of practice as you start to write your own functions.

107
00:09:22,150 --> 00:09:25,290
The first time you write a function are probably pretty bad pretty terrible.

108
00:09:25,360 --> 00:09:29,470
But then as you start to work out what your workflow is you'll start to understand how to write a better

109
00:09:29,470 --> 00:09:30,830
function.

110
00:09:30,830 --> 00:09:34,450
Now we can return the image.

111
00:09:34,450 --> 00:09:36,260
So what has this done.

112
00:09:36,470 --> 00:09:38,560
Has it done exactly what we just said.

113
00:09:38,750 --> 00:09:41,870
It takes an image path and an image size.

114
00:09:41,870 --> 00:09:44,470
We can define that here.

115
00:09:44,570 --> 00:09:49,820
It reads in the image file from the path named beautiful it turns the j peg into numerical tensor.

116
00:09:49,820 --> 00:09:54,950
So we're just converting it to a tensor with three color channels red green and blue because that's

117
00:09:54,950 --> 00:09:57,880
the convention for most deplaning image models.

118
00:09:57,890 --> 00:09:58,760
They want red.

119
00:09:58,910 --> 00:10:01,090
I said read that but that's a bad thing.

120
00:10:01,100 --> 00:10:08,840
Reading is a great thing we then converted the color channels from 0 to 255 to 0 to 1 values and we

121
00:10:08,840 --> 00:10:11,360
finally resize the image to our desired value.

122
00:10:11,360 --> 00:10:13,990
Now we've written this in a function.

123
00:10:14,180 --> 00:10:17,040
So the beautiful thing is that we can use this later on.

124
00:10:17,060 --> 00:10:18,170
This is what I used to do right.

125
00:10:18,170 --> 00:10:22,690
I used to run through these lines of code every time I wanted to process an image.

126
00:10:22,820 --> 00:10:24,410
Terrible workflow.

127
00:10:24,410 --> 00:10:30,160
So we can actually remove these remember if you want to check out what the function is actually doing.

128
00:10:30,160 --> 00:10:36,560
Break it down line by line takes an image path as input and then see what the rest of it does.

129
00:10:36,580 --> 00:10:37,930
Now I think that's enough for this video.

130
00:10:37,930 --> 00:10:44,620
We've got our function created ready to pre process some images what we might do now is again if we

131
00:10:44,620 --> 00:10:47,290
were to read through the loading data intensive low.

132
00:10:47,770 --> 00:10:49,370
I think it's this one yes.

133
00:10:49,570 --> 00:10:54,520
Back through here there's a little section in here that tends to flow likes to see things in the form

134
00:10:54,520 --> 00:10:55,730
of batches.

135
00:10:55,900 --> 00:10:59,310
So we'll figure what batches are in the next video.