1
00:00:00,720 --> 00:00:06,570
We've got a predictions array but now we need to get it in the Kaggle submission format so that's what

2
00:00:06,570 --> 00:00:08,150
we'll be working towards in this video.

3
00:00:08,940 --> 00:00:12,060
Let's go make ourselves a little heading.

4
00:00:12,060 --> 00:00:21,810
Preparing test data set predictions for Kaggle and if we look at the sample submission we kind of know

5
00:00:21,810 --> 00:00:27,330
what we need and we know we need an I.D. column which is the file I.D. from our test images.

6
00:00:27,330 --> 00:00:34,110
We know we need columns for each of the different dog breeds and we know we need the prediction probabilities

7
00:00:34,110 --> 00:00:37,170
which is what we've got at the moment we've got an array with all of these numbers.

8
00:00:37,200 --> 00:00:43,670
We basically need to just set up our predictions probability array into a data frame like this.

9
00:00:43,820 --> 00:00:48,300
Export it as a CSB and then upload that to Kaggle.

10
00:00:48,300 --> 00:00:49,850
So let's write that down.

11
00:00:50,040 --> 00:00:55,910
Actually we'll put a little link in here of what Kaggle wants a valuation.

12
00:00:55,910 --> 00:00:56,480
Copy that.

13
00:00:56,510 --> 00:01:00,350
This is just what the sample submission has to look like.

14
00:01:00,350 --> 00:01:16,550
Looking at the Kaggle sample submission we find that it wants our models prediction probability outputs

15
00:01:17,270 --> 00:01:30,020
in a data frame with an idea column and a column for each different dog breed.

16
00:01:30,040 --> 00:01:35,270
Well it's not really a data frame it's a CSA but we can be a little bit incorrect there in our wording

17
00:01:35,930 --> 00:01:38,090
and how we're gonna do this.

18
00:01:38,420 --> 00:01:40,220
Let's write ourselves some guides.

19
00:01:40,370 --> 00:01:49,640
So to get the data in this format well this is a testament to our favorite little graphic here of preparing

20
00:01:49,670 --> 00:01:51,170
inputs and outputs.

21
00:01:51,170 --> 00:01:57,420
So now we've worked with the inputs we've saved and reloaded our train model.

22
00:01:57,520 --> 00:02:00,140
Now we're making a bunch of predictions that we have made predictions.

23
00:02:00,160 --> 00:02:05,290
We're preparing our outputs more specifically preparing our outputs for Kaggle.

24
00:02:05,750 --> 00:02:20,570
We will create a panda's data frame with an I.D. idea column as well as a column for each dog breed

25
00:02:21,920 --> 00:02:32,240
and then we're going to add data to the idea column by extracting the test image ideas from their file

26
00:02:32,240 --> 00:02:36,870
parts and then we're going to add data.

27
00:02:37,180 --> 00:02:51,910
The prediction probabilities to each of the dog breed columns and then we want to export the data frame

28
00:02:52,840 --> 00:03:00,790
as a CSP to submit it to Kaggle.

29
00:03:00,820 --> 00:03:02,530
This is what we're after.

30
00:03:02,830 --> 00:03:11,640
And probabilities is spelled wrong standard Daniel type poking some nice little fool stops all right.

31
00:03:11,700 --> 00:03:12,470
So the first step.

32
00:03:12,480 --> 00:03:17,780
Let's go through and create a panel is data frame we've had experience of this way back in the panda

33
00:03:17,790 --> 00:03:20,460
section with empty columns.

34
00:03:20,490 --> 00:03:28,830
So we'll go to parades DLF equals PD data frame so just the standard nomenclature for prediction data

35
00:03:28,830 --> 00:03:36,770
frame parades IDF parades data frame all predictions data frame columns is going to be equal to I D

36
00:03:37,820 --> 00:03:46,330
plus we need a column for each different dog breed so we'll do one list unique breeds and I just want

37
00:03:46,330 --> 00:03:50,270
to show you what list unique breeds looks like list unique breeds

38
00:03:53,640 --> 00:03:54,300
so there we go.

39
00:03:54,840 --> 00:04:09,690
So if we added I.D. to that let's give that about I date IPS trigger happy once again plus there we

40
00:04:09,700 --> 00:04:11,980
go I.D. plus you list unique breeds

41
00:04:14,970 --> 00:04:25,320
that is some column names for us we'll comment out that and we will check out the head of our spreads

42
00:04:25,320 --> 00:04:32,340
IDF should be empty with just a bunch of column names that's what we're after I.D. and all of our unique

43
00:04:32,340 --> 00:04:35,700
breeds in order we've done this.

44
00:04:35,700 --> 00:04:37,910
So putting in some emojis.

45
00:04:38,070 --> 00:04:43,290
All right add data to the idea come on by extracting the test image I.D. from their file paths I believe

46
00:04:43,290 --> 00:04:51,900
we have already done that but what we might do is just do it again test file names but the way we can

47
00:04:51,900 --> 00:05:00,270
get each I.D. is by we want to extract this out so we could do a little bit of fancy regex on this but

48
00:05:00,450 --> 00:05:09,450
I'll show you another way append test image I.D. to predictions data frame.

49
00:05:09,870 --> 00:05:17,650
So we want to go test path look did we already create test path we did.

50
00:05:17,680 --> 00:05:18,640
Okay.

51
00:05:18,790 --> 00:05:19,720
Wonderful.

52
00:05:19,720 --> 00:05:24,380
So let's get test I.D..

53
00:05:24,430 --> 00:05:33,640
So this is just the path to our test file and test I.D. is going to be equal to a list comprehension

54
00:05:33,640 --> 00:05:35,320
of our stock path.

55
00:05:35,320 --> 00:05:37,940
And then what is it split text.

56
00:05:37,990 --> 00:05:48,810
I'll show you what this does in the second split text and for path in always don't list the test path.

57
00:05:48,820 --> 00:05:52,910
That should be correct how can I show you an example.

58
00:05:53,150 --> 00:05:56,030
Always step path dot split text.

59
00:05:56,030 --> 00:06:01,330
We want to get test file names 0.

60
00:06:01,650 --> 00:06:04,160
Well this work.

61
00:06:04,190 --> 00:06:04,670
There we go.

62
00:06:04,950 --> 00:06:05,230
OK.

63
00:06:05,270 --> 00:06:09,980
So it splits off the last little bit of the path but we only want this.

64
00:06:10,100 --> 00:06:19,790
So that's why we have to do it for just the test path so we want to just go through this test folder.

65
00:06:19,820 --> 00:06:22,500
Now it's not going to show up here is it because there's a lot of images in there.

66
00:06:22,500 --> 00:06:26,650
That'll take a long time to load and Colette you're just going to have to trust me with what this does

67
00:06:28,780 --> 00:06:30,620
let's just see it in action anyway.

68
00:06:30,760 --> 00:06:32,150
Test ideas.

69
00:06:32,560 --> 00:06:39,410
Will this work if in doubt run the code.

70
00:06:39,470 --> 00:06:48,660
All righty actually we don't need the JPEG bit.

71
00:06:48,710 --> 00:06:53,720
So what we're going to have to do we only want the first element of each of these.

72
00:06:54,320 --> 00:07:01,850
So we're going to have to put a little zero index there and let's see what this does.

73
00:07:01,890 --> 00:07:03,120
That's what we're after.

74
00:07:03,120 --> 00:07:04,080
Test days.

75
00:07:05,030 --> 00:07:06,390
I don't want a large print out there.

76
00:07:06,900 --> 00:07:10,100
So what I'm gonna do is just go Fred's DNF.

77
00:07:10,200 --> 00:07:12,590
So I'm taking my data frame from up here.

78
00:07:12,710 --> 00:07:18,430
I'm going to make an I.D. column with the test I.D..

79
00:07:18,900 --> 00:07:24,930
So just taking that and putting that there little bit of overkill running this cell three times but

80
00:07:25,290 --> 00:07:26,430
that is all right.

81
00:07:26,460 --> 00:07:32,860
Let's check out the head of our spreads ADF as we always do when we manipulate our data okay.

82
00:07:32,880 --> 00:07:34,230
So you've got some ideas in there.

83
00:07:34,230 --> 00:07:40,770
Now we need to just append all of our prediction probabilities for each dog breed.

84
00:07:40,770 --> 00:07:42,440
Now how might we do that.

85
00:07:42,440 --> 00:07:43,290
Mm hmm.

86
00:07:43,320 --> 00:07:47,700
I want you to have a think about this for a second.

87
00:07:47,830 --> 00:07:49,310
What do we have.

88
00:07:49,420 --> 00:07:56,390
We have test predictions which is an array of predictions probabilities and what do we have that's a

89
00:07:56,390 --> 00:07:57,950
list of different dog breeds.

90
00:07:58,910 --> 00:08:05,330
We have unique breeds cos we do so add the prediction probabilities this is getting exciting we're getting

91
00:08:05,330 --> 00:08:11,260
so close to being out to submit our first full blown DB learning model to Kaggle.

92
00:08:11,450 --> 00:08:19,150
We've gone from head to toe the whole process the whole shebang.

93
00:08:19,430 --> 00:08:21,130
She can't tell I'm pumped.

94
00:08:21,150 --> 00:08:26,640
So this is going to access the parades DLF of all of the unique breeds.

95
00:08:26,640 --> 00:08:34,770
So all of the braids columns and then we want to just set that easy this is to test predictions.

96
00:08:34,770 --> 00:08:37,200
Parades the F dot head

97
00:08:40,830 --> 00:08:45,050
and we are done ladies and gentlemen have a go at that.

98
00:08:45,150 --> 00:08:48,360
We would compare that to the sample submission.

99
00:08:48,480 --> 00:08:54,860
Of course this goes on for a lot longer but ours is using the model we have built.

100
00:08:54,870 --> 00:08:57,250
The numbers are going to be different but that's okay.

101
00:08:57,480 --> 00:09:02,240
Now let's export it to a CSA so we can submit it to Kaggle.

102
00:09:02,310 --> 00:09:06,070
How do we export Panda's data frame to CSP.

103
00:09:06,090 --> 00:09:14,670
We use to see as vehicles we want to go drive my drive I'm going to go to dog vision and I'm just gonna

104
00:09:14,700 --> 00:09:27,930
save it as maybe full model predictions submission one CSA or maybe we'll all tell ourselves my net

105
00:09:29,160 --> 00:09:32,670
V2 don't see us.

106
00:09:32,860 --> 00:09:34,900
And we don't want index.

107
00:09:35,230 --> 00:09:39,570
So index equals false we'll go here.

108
00:09:39,570 --> 00:09:51,390
We'll just write what we're doing save our predictions data frame to see us v for submission to Kaggle

109
00:09:52,900 --> 00:09:54,610
is is gonna save.

110
00:09:54,710 --> 00:09:56,030
Fingers crossed.

111
00:09:56,240 --> 00:09:58,670
Yes of course it will.

112
00:09:58,670 --> 00:09:59,810
Now let's have a look in here.

113
00:10:01,490 --> 00:10:11,400
In our files oh maybe we've glitched out we've just saved a c v there may be lives lost the plot on

114
00:10:11,400 --> 00:10:12,570
us.

115
00:10:12,570 --> 00:10:21,700
Give me a second for my lab window to reload and I'll be back once I have access to this folder okay.

116
00:10:21,770 --> 00:10:22,240
I'm back.

117
00:10:22,520 --> 00:10:24,860
So sorry about that little interruption in broadcast.

118
00:10:24,860 --> 00:10:27,770
Seems my collab window decided to go bonkers.

119
00:10:27,770 --> 00:10:28,520
I see why.

120
00:10:28,520 --> 00:10:35,750
Because I decided to open the test folder and I need to close that and I might have to pause it again

121
00:10:36,740 --> 00:10:38,120
one second.

122
00:10:38,120 --> 00:10:38,440
All right.

123
00:10:38,450 --> 00:10:39,380
Third time's a charm.

124
00:10:39,380 --> 00:10:40,060
We are back.

125
00:10:40,070 --> 00:10:44,830
So be very careful with folders that you open in your lab window.

126
00:10:44,990 --> 00:10:48,160
If they're full of data a.k.a. 10000 plus images.

127
00:10:48,260 --> 00:10:54,430
It will freeze up your co lab notebook there we go for model predictions.

128
00:10:54,430 --> 00:10:56,370
Well we can't really see the whole thing there.

129
00:10:56,480 --> 00:10:57,130
Oh there we go.

130
00:10:57,370 --> 00:10:59,200
And that's just our CSB file there.

131
00:10:59,800 --> 00:11:02,000
So if we wanted to we can download that.

132
00:11:02,050 --> 00:11:03,530
We do want to.

133
00:11:03,550 --> 00:11:05,190
So we'll just wait for that to go through.

134
00:11:05,200 --> 00:11:09,650
I'll speed up my video until this is fully downloaded Ronnie.

135
00:11:09,740 --> 00:11:11,710
So this is fully downloaded.

136
00:11:11,750 --> 00:11:15,980
All I did was just click download and then waited for it to go to about a minute or so not too long.

137
00:11:15,980 --> 00:11:17,900
This is about a 30 megabyte CSB.

138
00:11:18,160 --> 00:11:21,120
We come to my downloads folder.

139
00:11:21,120 --> 00:11:21,910
If we come here.

140
00:11:21,970 --> 00:11:26,580
Yeah about a 30 megabyte C as V so 10000 predictions.

141
00:11:26,690 --> 00:11:27,590
Now we've got that.

142
00:11:27,620 --> 00:11:33,590
Let's go to cargo and we can go late submission because this competition has already close you can see

143
00:11:33,590 --> 00:11:36,390
my past one there five days ago.

144
00:11:36,620 --> 00:11:42,140
So let's upload I want to show you that this works open.

145
00:11:42,590 --> 00:11:51,680
I'm just selecting the CSB we just downloaded it's gonna go through a briefly describe your submission.

146
00:11:51,910 --> 00:11:52,640
Let's go.

147
00:11:52,720 --> 00:11:55,180
Transfer learning that's what we'll call it because that's what we used.

148
00:11:55,180 --> 00:12:05,590
Transfer learning my ball net the to dog vision.

149
00:12:05,630 --> 00:12:12,110
Baby and then we'll make submission and it's gonna score it.

150
00:12:12,220 --> 00:12:14,440
Now there's something up with this.

151
00:12:14,570 --> 00:12:19,710
Oh 20 so that would be our goal is to beat this one on the next one.

152
00:12:20,620 --> 00:12:28,140
And if we go back to the leaderboard see where we appeared we'll probably find public leaderboard score

153
00:12:28,140 --> 00:12:30,570
0.

154
00:12:31,210 --> 00:12:36,010
I'd be very skeptical of any kind of competition where people have a score of zero.

155
00:12:36,100 --> 00:12:45,040
So what you can actually find out is you can actually janky or actually hack the score here so don't

156
00:12:45,040 --> 00:12:48,960
worry too much about submitting to here on a late submission.

157
00:12:48,970 --> 00:12:56,350
Your next goal is to reduce your own score and if you want my example submission CSA I'll just like

158
00:12:56,350 --> 00:13:00,550
this in the resource section you'll probably get the same score as me.

159
00:13:00,660 --> 00:13:06,990
Our next step in this project if we were to keep extending it and train a new model would be to beat

160
00:13:07,080 --> 00:13:12,810
our original score rather than paying attention to the leaderboard straight away is focus on when you

161
00:13:12,810 --> 00:13:15,690
first begin beach your own score.

162
00:13:15,690 --> 00:13:23,720
So now we've seen how to make full blown prediction for the cargo competition but that's not our original

163
00:13:23,720 --> 00:13:24,440
focus.

164
00:13:24,440 --> 00:13:25,610
Well it kind of was.

165
00:13:25,610 --> 00:13:27,760
I lied right back up the top.

166
00:13:27,800 --> 00:13:34,260
What's this notebook called Dog vision so our problem is identifying the breed of a dog given an image

167
00:13:34,260 --> 00:13:34,730
of a dog.

168
00:13:34,890 --> 00:13:39,720
When I'm sitting at the cafe and I take a photo of a dog I want to know what breed of dog it is.

169
00:13:39,840 --> 00:13:47,060
So it's likely you want to be able to make predictions on your own custom images.

170
00:13:47,070 --> 00:13:51,750
So rather than just images that Kaggle has given us in a test and training set.

171
00:13:51,780 --> 00:13:58,290
What if you had some images of your own dog which I happen to have or a dog that you saw at the cafe

172
00:13:58,290 --> 00:14:02,550
that looked really cute and pretty and you wanted to know what breed it was you took a photo and you

173
00:14:02,550 --> 00:14:07,530
came back to a machine learning model in this notebook and you like I want to see what breed of dog

174
00:14:07,530 --> 00:14:09,380
that was using my model.

175
00:14:09,390 --> 00:14:10,570
Well let's have a look at how to do that.

176
00:14:10,570 --> 00:14:14,160
The next video making predictions on custom images.