1
00:00:00,210 --> 00:00:03,480
Now this is exactly what I was talking about.

2
00:00:03,540 --> 00:00:07,860
If we go back to the KEYNOTE things you might see in collab.

3
00:00:07,860 --> 00:00:09,070
This is one of them.

4
00:00:09,090 --> 00:00:11,970
So what it means is that I took a little break just now.

5
00:00:11,970 --> 00:00:18,710
And if we come back to collab because of the limited amount of resources that Google is able to provide.

6
00:00:18,840 --> 00:00:25,470
If you leave your notebook in collab for a long enough time it will eventually disconnect from a runtime

7
00:00:25,710 --> 00:00:31,340
which is a runtime is again just a computer on Google server that powers you know.

8
00:00:31,710 --> 00:00:34,680
So this is what you do if you have to take a break.

9
00:00:34,740 --> 00:00:36,740
You should take a break every so often.

10
00:00:36,780 --> 00:00:39,270
You're not spending too much time sitting down go for a walk.

11
00:00:39,270 --> 00:00:40,400
Nature is beautiful right.

12
00:00:40,850 --> 00:00:41,970
Well that's what I just did.

13
00:00:41,970 --> 00:00:47,860
I just took my dog for a walk so if you do get this little pop up you're going to have to reconnect.

14
00:00:47,860 --> 00:00:51,310
But you'll see up here the runtime is reconnecting.

15
00:00:51,310 --> 00:00:57,440
So it's initializing mounting Google Drive which is where our files are maybe it takes a little while

16
00:00:57,440 --> 00:00:58,820
to appear.

17
00:00:58,880 --> 00:00:59,480
There we go.

18
00:00:59,480 --> 00:01:00,970
Drive.

19
00:01:00,980 --> 00:01:03,030
We'll have google drive back.

20
00:01:03,050 --> 00:01:06,260
There's dog vision our beautiful little project photo.

21
00:01:06,350 --> 00:01:06,770
Excellent.

22
00:01:06,770 --> 00:01:08,470
We can close this out we don't need that.

23
00:01:08,570 --> 00:01:12,800
But what we're going to have to do is Oh better take off these right.

24
00:01:12,800 --> 00:01:14,890
Pay attention to my own notes.

25
00:01:15,320 --> 00:01:19,850
I'll put a little green tick emoji we've imported import intensive not hub making sure we're using a

26
00:01:19,850 --> 00:01:21,990
GP you now.

27
00:01:22,010 --> 00:01:28,730
The beautiful thing is the notebook itself will save the fact that we are using a GP you.

28
00:01:28,890 --> 00:01:34,610
So when it reloads we won't have to change the hardware accelerate it to you it will automatically load

29
00:01:34,610 --> 00:01:36,050
it as a GP you.

30
00:01:36,560 --> 00:01:41,230
But we do have to rerun any cells that we may have ran before.

31
00:01:42,140 --> 00:01:43,300
So there we go.

32
00:01:44,390 --> 00:01:46,430
Check the GP import tensor flow.

33
00:01:46,460 --> 00:01:52,540
Make sure we use intensive load to point X and I'll delete this one.

34
00:01:52,600 --> 00:01:54,040
Wonderful.

35
00:01:54,040 --> 00:01:54,640
Okay.

36
00:01:54,760 --> 00:02:00,430
So now we've got our workspace ready and you might be thinking Daniel that took a fairly long time to

37
00:02:00,430 --> 00:02:01,020
get set up.

38
00:02:01,090 --> 00:02:05,620
But as always a person is only as good as the tool they use.

39
00:02:05,620 --> 00:02:05,890
Right.

40
00:02:05,890 --> 00:02:07,620
Or as the tools they use.

41
00:02:07,660 --> 00:02:13,250
So that's why we spent a bit of time getting our code lab workspace ready getting our new project folder

42
00:02:13,250 --> 00:02:15,420
together getting our data there.

43
00:02:15,430 --> 00:02:19,230
So now we go back What step are we up to.

44
00:02:19,240 --> 00:02:22,030
There should be a zero here for forgetting our workspace ready.

45
00:02:22,300 --> 00:02:28,030
But now we're up to step 1 which is getting our data ready a.k.a. turning it into tenses.

46
00:02:28,030 --> 00:02:31,000
Now this is the premise of tensor flow in itself.

47
00:02:31,000 --> 00:02:32,590
That's why it's called tensor flow.

48
00:02:32,590 --> 00:02:37,930
The premise is to get your data into tenses a.k.a. remember all machine learning models like to deal

49
00:02:37,930 --> 00:02:38,910
with numbers.

50
00:02:38,950 --> 00:02:45,610
So we have to get our data into tenses which is just a fancy word for Matrix a.k.a. a name pi array

51
00:02:45,820 --> 00:02:48,490
with some dimensions but a tensor.

52
00:02:48,610 --> 00:02:51,730
The key thing here is that it can be run on a Jeep you.

53
00:02:51,790 --> 00:02:58,330
So if we go back here at tensor flow is turn your data into tenses and then it flows through a workflow

54
00:02:58,360 --> 00:02:59,350
like this.

55
00:02:59,560 --> 00:03:01,000
You get it tensor flow

56
00:03:03,860 --> 00:03:07,360
doing a machine learning goes and you get a comedy show for free.

57
00:03:07,360 --> 00:03:08,260
All right.

58
00:03:08,260 --> 00:03:10,590
So step one we're going to access the data.

59
00:03:10,600 --> 00:03:16,820
So right now it's over here in drive in the dog vision.

60
00:03:16,820 --> 00:03:21,690
So now we have to somehow get it from this folder into this code space.

61
00:03:21,770 --> 00:03:26,270
So let's do that we'll get out of that make a little heading.

62
00:03:26,420 --> 00:03:32,100
Actually we can go to we're getting our data ready.

63
00:03:32,570 --> 00:03:41,300
Turning it into tenses remember we're communicating with ourselves here we go here with all machine

64
00:03:41,300 --> 00:03:56,430
learning models our data has to be in numerical format so that's what we'll be doing first turning our

65
00:03:56,490 --> 00:03:59,820
images into tenses

66
00:04:04,140 --> 00:04:04,980
numerical

67
00:04:08,150 --> 00:04:09,410
representation

68
00:04:13,920 --> 00:04:20,520
and now this is a code cell and in that collab a special way to transform something into markdown is

69
00:04:20,520 --> 00:04:27,600
to hold command and double tap M if you want to see a list of all shortcuts and collab hold command

70
00:04:27,720 --> 00:04:34,750
M H there we go have a look and I far out there is a lot there.

71
00:04:34,780 --> 00:04:40,900
So this takes some practice again because some of the commands we've been using in Jupiter are slightly

72
00:04:40,900 --> 00:04:41,770
different.

73
00:04:41,830 --> 00:04:44,710
If you do need to look one up have a look in here.

74
00:04:44,890 --> 00:04:49,930
Otherwise I'll just kind of keep running with it if I need to add a code so I can just press this little

75
00:04:49,930 --> 00:04:55,440
plus button here or up here and the main one will will be using is shifting into right.

76
00:04:55,450 --> 00:04:59,800
Just to run our code let's look at the labels what does that look like.

77
00:04:59,800 --> 00:05:00,640
So if we come back

78
00:05:04,080 --> 00:05:12,600
let's start by accessing our data and checking out the labels.

79
00:05:13,850 --> 00:05:16,170
So to do that it's in a CSB format.

80
00:05:16,230 --> 00:05:21,780
So if we want to check it out we might have to or we will have to import pandas because it's a lot easier

81
00:05:21,780 --> 00:05:28,770
to look at things that I see format using pandas so check out the labels of our data.

82
00:05:28,770 --> 00:05:35,530
We're gonna import pandas as payday our favorite labels CSP equals.

83
00:05:36,510 --> 00:05:42,840
So what we're going to do is just pass in maybe we could set up a little a little PA but we'll type

84
00:05:42,840 --> 00:05:44,110
it out to begin with.

85
00:05:44,430 --> 00:05:50,210
We could just copy this copy path but we're gonna practice.

86
00:05:50,230 --> 00:05:51,140
What would that look like.

87
00:05:51,160 --> 00:05:51,820
Something like that.

88
00:05:51,820 --> 00:05:52,990
Let's practice typing it out.

89
00:05:52,990 --> 00:06:01,300
Remember you don't need content drive my drive slash dogged vision.

90
00:06:02,080 --> 00:06:04,010
Oh yeah.

91
00:06:04,180 --> 00:06:05,560
This is what projects need right.

92
00:06:05,590 --> 00:06:09,500
I need cool named labels dot CSB.

93
00:06:09,690 --> 00:06:10,060
All right.

94
00:06:10,300 --> 00:06:12,520
And we're going to can that.

95
00:06:13,320 --> 00:06:23,610
And then what we might check out is we'll get labels CSB we wanted to describe it describe would help

96
00:06:23,610 --> 00:06:31,970
if I could type I'd be full time lucky and we go and print we'll also get the head.

97
00:06:32,040 --> 00:06:37,690
So the first five rows let's check it out again more may this mall space.

98
00:06:37,820 --> 00:06:39,500
Wonderful.

99
00:06:39,500 --> 00:06:45,090
What do we have here count ten thousand two hundred twenty two I.D..

100
00:06:45,170 --> 00:06:48,990
Grade ten thousand two hundred twenty two unique value.

101
00:06:49,010 --> 00:06:49,470
So.

102
00:06:49,540 --> 00:06:50,290
OK.

103
00:06:50,450 --> 00:06:53,530
So there's ten thousand two hundred twenty two unique ideas.

104
00:06:53,870 --> 00:07:00,010
And one hundred and twenty unique grades and then this is what the I.D. column looks like.

105
00:07:01,280 --> 00:07:03,950
Again it looks a bit different because we've printed it.

106
00:07:03,950 --> 00:07:07,390
So if we just add labels dot CSB dot head.

107
00:07:07,390 --> 00:07:12,810
This is what you might be used to looking at something like that Navy guys are going to I.D. column

108
00:07:13,110 --> 00:07:17,590
which is the file names I believe of our images.

109
00:07:17,610 --> 00:07:23,260
So this is of train and the brain associated with that I.D..

110
00:07:23,290 --> 00:07:30,880
Now a little tidbit here if you do have a lot of files in one of your folders in Google Drive it actually

111
00:07:30,880 --> 00:07:33,710
does take a fairly long time to load here.

112
00:07:33,790 --> 00:07:38,060
So rather than wait for that let's just go back to Kaggle.

113
00:07:38,110 --> 00:07:41,560
Treat this as our way of looking at things.

114
00:07:41,570 --> 00:07:42,200
There we go.

115
00:07:42,380 --> 00:07:49,430
So if we get a label CSB ten thousand two hundred twenty two unique values Yes that's correct.

116
00:07:49,520 --> 00:07:50,370
Unique.

117
00:07:50,640 --> 00:07:53,560
And then train.

118
00:07:53,790 --> 00:08:01,010
We've got unique ideas so does that look like what we've got in our day.

119
00:08:01,040 --> 00:08:02,210
Yes it does.

120
00:08:02,210 --> 00:08:03,380
Wonderful.

121
00:08:03,380 --> 00:08:03,800
Okay.

122
00:08:03,800 --> 00:08:05,270
Now what we might check out.

123
00:08:05,450 --> 00:08:07,020
We've got 120 different brains.

124
00:08:07,020 --> 00:08:14,240
So that means that there's ten thousand or so images hopefully there's about 100 hundred or so images

125
00:08:14,690 --> 00:08:15,590
per class.

126
00:08:16,340 --> 00:08:17,650
Well actually let's just figure that out.

127
00:08:17,660 --> 00:08:19,810
Let's figure out how many images there are.

128
00:08:19,910 --> 00:08:25,080
Her class first and then we'll see why that's important.

129
00:08:25,130 --> 00:08:29,420
How many images are there of each breed.

130
00:08:29,540 --> 00:08:36,010
And we could probably do this pretty easily by going labels since the grade some accessing the braid

131
00:08:36,020 --> 00:08:37,900
column and then I'm going to go.

132
00:08:37,900 --> 00:08:38,900
Value counts.

133
00:08:38,900 --> 00:08:44,220
So count the values in this column actually let's just see what that does.

134
00:08:44,220 --> 00:08:49,280
This is a dog braid Scottish deer hound Maltese dog and to butcher.

135
00:08:49,320 --> 00:08:50,540
Not even sure what that is.

136
00:08:50,550 --> 00:08:53,280
Shout out if if your dog's on here asking my dog.

137
00:08:53,500 --> 00:08:55,540
Yeah it's really cool OK.

138
00:08:55,790 --> 00:09:03,010
And now let's visualize this is probably going to look a bit better if we visualize this pig size equals

139
00:09:03,090 --> 00:09:03,990
2010

140
00:09:08,010 --> 00:09:15,310
Wall see the beauty of collab is that we didn't even have to import map Gottlieb to see that.

141
00:09:15,360 --> 00:09:20,870
So my notebooks a bit zoomed in but this is pretty cool if we were to draw a line.

142
00:09:21,180 --> 00:09:25,680
So look we got dog braids here all of 120 different dog rates.

143
00:09:25,680 --> 00:09:28,830
And then this is the number of images that they have.

144
00:09:28,830 --> 00:09:37,160
So if we were to draw a line somewhere across here we'd say about maybe seventy five on average we can

145
00:09:37,160 --> 00:09:40,440
figure that out pretty easily anyway so labels CSA.

146
00:09:40,460 --> 00:09:44,390
How do you find the main of a column.

147
00:09:44,390 --> 00:09:53,930
Actually we need to count them first because otherwise they're just strings dot main average maybe it's

148
00:09:53,930 --> 00:09:55,820
probably a little bit more robust in the main

149
00:09:59,580 --> 00:10:00,830
so the middle number 82.

150
00:10:00,910 --> 00:10:01,130
OK.

151
00:10:01,160 --> 00:10:02,180
So we said seventy five.

152
00:10:02,210 --> 00:10:03,950
But that is a good thing.

153
00:10:04,190 --> 00:10:11,030
You can imagine right if you had let's say Scottish deer hound if we had 500 images of that but this

154
00:10:11,030 --> 00:10:11,920
bad boy over here.

155
00:10:11,930 --> 00:10:13,940
What is this one Eskimo dog.

156
00:10:14,230 --> 00:10:20,750
If we only had three images of Eskimo dog we could imagine our machine learning model might do very

157
00:10:20,750 --> 00:10:25,260
well at figuring out what a Scottish D ham looks like because it has so many examples.

158
00:10:25,490 --> 00:10:30,440
But as for an Eskimo dog if there's only three images it's probably going to be a bit hard to figure

159
00:10:30,440 --> 00:10:36,910
out what it looks like when you're asking yourself how many images should I have per class Google recommends

160
00:10:37,030 --> 00:10:45,250
a bare minimum ten Google minimum number of images per class

161
00:10:48,710 --> 00:10:50,000
preparing your training data.

162
00:10:50,000 --> 00:10:50,450
Here we go.

163
00:10:50,450 --> 00:10:56,090
So this is Google's auto m l so which is which is their service for automatically building machine learning

164
00:10:56,090 --> 00:10:57,820
models.

165
00:10:57,920 --> 00:11:00,560
This is some extra curricular if you want to check this out.

166
00:11:00,560 --> 00:11:05,180
Pretty cool preparing your data.

167
00:11:05,180 --> 00:11:07,550
There we go for model training purposes.

168
00:11:07,550 --> 00:11:12,140
It is recommended you use about 100 annotations per label.

169
00:11:12,140 --> 00:11:19,800
So in our case it would be great if you had 100 examples of each image class with at least 10.

170
00:11:20,100 --> 00:11:24,320
So one hundred is great but at least 10 is a good start.

171
00:11:24,710 --> 00:11:29,450
So if you wanted to build your own image classifier if it's a binary classification you'd want to be

172
00:11:29,450 --> 00:11:32,780
looking at minimum 10 images per class.

173
00:11:32,780 --> 00:11:40,670
So our data set out about 75 to 85 images per class is pretty well spread out.

174
00:11:40,670 --> 00:11:41,630
This is a good start.

175
00:11:42,140 --> 00:11:42,780
OK.

176
00:11:42,920 --> 00:11:45,800
So done a little bit of exploration of the labels.

177
00:11:45,800 --> 00:11:53,460
Now what we might do is start to give out image paths so labels is pretty easy to deal with in a CSA

178
00:11:53,480 --> 00:11:58,370
but all of our images again we can't see that are in this format here.

179
00:11:58,460 --> 00:12:01,790
So we need a way to get them into our CO lab notebook.

180
00:12:01,790 --> 00:12:04,180
Let's have a look at how to do that in the next video.