Original subtitles

1

Hello and welcome to the last step of this module to object detection before we start the homework.

2

But here we go with the last step and this last step is going to be about the training of the SSD.

3

So we're not going to implement the whole code that we'll try and the SSD simply because it's a huge

4

code.

5

And besides you're going to see that you're going to need specific requirements to be able to run that

6

training.

7

So this tutorial is going to be about explaining how you could train the SSD.

8

So I'm going to show you the file how to execute it.

9

And mostly I'm going to show you the data set on which the SSD was trained.

10

So in case you want to train the SSD with some other object you would understand the approach of how

11

to do it.

12

But remember you're going to see that you're going to need some pretty advanced system.

13

So as you can see right now I'm on a web page and this is the web page that contains the data sets on

14

which the SS was trained.

15

This dataset is not image net.

16

If you were hoping for that but the Pascal visual object classes dataset.

17

So if you remember in the code we saw Virk the OSI you know for classes to have the mapping between

18

the classes and integers.

19

Well Varg stands for visual object classes and this is already a huge dataset not as huge as image net

20

but a huge one that you will see contains lots and lots of images.

21

So you can find this page at this address but don't worry the next it will be an article which you will

22

find the homework folder that will contain all the tools to train the SSD Plus the address of this website

23

so that you can have a look at this better.

24

All right so let's go through this web page so as you can see the Pascals Virk project provides standardized

25

image data sets for object recognition.

26

So exactly what we're doing and besides provides a common set of tools for accessing the data set and

27

annotations.

28

So it's not only a data set it's also an advanced structure of data set that helps a lot for the training.

29

And also you can see that there were some challenges but now finished.

30

But anyway it is definitely useful data set to train the SSD.

31

And so now if we have a closer look at what these days that are well we need to scroll down to that.

32

What challenges from 2005 to 2012 and here are the data sets you can see there's several of them from

33

2005 to 2012 and the ones we have the ones on which the SSD was trans and I'm going to show you the

34

code and the data sets themselves and how you can execute the code to train the SSD on these two data

35

set.

36

These two data sets are the Vark 2007 and the Vok 2012.

37

So let's have a closer look at where they are right now and we need to scroll down even more.

38

Actually at the bottom of the page there we go.

39

So here you have this table that's the most important part of this page.

40

That's where you can see how the data sets are structured and made and what they contain exactly.

41

So let's have a look at the first one.

42

Remember we have the 2007 data set and 2012 sort of the 2007 data set contains 20 classes.

43

Remember I told you that the pre-trained the model we have can detect between 30 to 40 object.

44

Well among these objects we have some persons Well the S's can detect any person in a video or in some

45

images then several animals some birds cats cows dogs horses sheeps some vehicles airplanes bicycles

46

boats buses cars you can do some computer vision for some driving car if you have a self-driving car

47

business I would highly recommend to integrate as is the model for the computer vision inside yourself

48

in cars are also motorbikes and trains of course.

49

And then we have some other objects like some bottles chairs dining table potted plants sofas and TV

50

monitors.

51

All right.

52

And I think we have even more but there you go.

53

You can see you have plenty of objects to detect.

54

So really really useful.

55

And now if we go to the 2012 dataset of 2012 we can see we also have 20 classes and the data contains

56

eleven thousand five hundred and thirty images containing 27 450 animated objects so we're going to

57

have a look at these objects specifically because as I told you I've prepared for you a folder to training

58

as is the folder that contains not only the code but also these two data sets.

59

We're going to have a clear look at images.

60

So we're going to look at some of them.

61

And so what I'm going to do right now is walk you through this folder that I've prepared.

62

So let's open Anaconda up.

63

There we go.

64

We need to connect the virtual platform of course because the training is made with by torch again and

65

some other tools that are preinstalled individual them.

66

So there we go.

67

Let's not forget to connect to the virtual platform applications on virtual platform.

68

There we go.

69

And now we launch spider so spiders launching and I'm going to show you this training as is the four

70

that I've prepared for you.

71

I put it in the module to object detection color which I'm going to go to right now so I'll explore

72

it on my desktop that's the computer vision it is that folder and if we go to module 2 you'll find a

73

new folder.

74

This folder training as is d that contains all the tools and data sets for you to train the SSD.

75

But I have to specify something very important now.

76

Remember at the beginning of this oil I told you that you need an advanced system to train the SSD while

77

this advance system is simply to have kuda on your machine.

78

Kuda is like an accelerator that is connected to every geographic cards and that allows to speed up

79

considerably the training computations.

80

Therefore the training implementations that were made are only compatible with kuda.

81

I'm going to tell you a fun fact now if there was an implementation of the training implemented without

82

kuda I can tell you for sure the training would take several months or even years.

83

This is not a joke.

84

The training of the SSD is first of all made with thousands and thousands of images.

85

But if you train that without kuda it would definitely take several months.

86

You can not trained the SSD on your CPQ you have to train it with a GP you and the best you can find

87

is kuda.

88

So just to set things clear this implementation is only kuda compatible and therefore you will only

89

be able to execute the code if you have an NDA Geographic card with kuda and labels on your machine.

90

But even if that's the case.

91

While it would take some time and you would not have your SSD trained in the next hour.

92

But I think it's still interesting for you to see how you can run the training and therefore that's

93

exactly what I'm going to show you right now.

94

But first let me walk you through this training SSD folder starting with the data so you can see we

95

have this data folder that contains several files.

96

Here you can have the scripts which are shellscript and which allow to download the datasets Vogue 2007

97

and Vok 2012.

98

So to do that you simply need to open a terminal and then type S H space Virk 2007 dot S H then execute

99

and this will download word 2007 and then you can type another command as a space walk 2012 that s h

100

and then press enter to the queue to download.

101

Also 2012 but I've already done that for you.

102

I've downloaded this dataset and here they are there in this folder Vark dev kit and you have Virk 2007

103

and Vark 2012.

104

So let's have a look at Vocht 2007.

105

First you have several subfolder but the one I want to show you is this one.

106

GPL images which contains lots and lots of images.

107

As you can see let's open randomly.

108

Some of them so let's see this one for example number 8.

109

It's a chair.

110

All right so that's the ground truth for the chair.

111

Remember Kyrios intuition lectures need to start with the ground truth to ground truth is the truth

112

that there is in fact a chair right here on this image.

113

So that's the ground truth that will be compared to the prediction of the SS doing the training just

114

before a possible error is back propagated into the neural network.

115

All right so let's close this let's open another one.

116

Number 14.

117

It's a car a yellow cab probably in New York I would say maybe.

118

Or some other cities in the US.

119

But there we go.

120

That's to detect a car so it's a ground truth for a car number 31 then a train.

121

There we go.

122

You can even detect train.

123

And let's open a few more.

124

So that's a pigeon to detect birds.

125

Ground Truth for the bird.

126

And now 53 a cat cute little cat.

127

And one last one hundred.

128

Well let's go further.

129

That's open the number one thousand five hundred eighty six.

130

And that's a beautiful horse jumping.

131

And speaking of horses well that's going to be the homework for this section.

132

It's going to be it's going to be a fun homework for you to relax and enjoy playing some object detection

133

and some really really nice videos.

134

We're going to do some action on beautiful horses running on a field.

135

So there we go.

136

You have the object.

137

And now let's have a look at the other data set because the training is made on the two data sets at

138

the same time.

139

So have to scroll back up.

140

There we go.

141

Varg 2017 and now let's have a look at 2012.

142

Same subfolder is.

143

But this time with different images.

144

All right.

145

Here they are.

146

So the 2012 we actually have images of several 2007 2008 and probably 2009 there we go and up to 2012.

147

So let's open any of them for example this one from 2010.

148

Beautiful dog with a somewhat suspicious look.

149

But I love dogs.

150

I hope you liked the dog bouncing on the field as we detected during this much too.

151

I really like this Doug.

152

So let's have let's open another one let's open one image from 2011.

153

There we go.

154

This one.

155

Well two nice persons probably a mother with her daughter.

156

So very nice.

157

Let's open one from 2012 now and that will be our last.

158

So there we go the less ground truth we have.

159

Oh another kid and we can see several persons.

160

This one is actually interesting.

161

There are several ground truth in this image.

162

So maybe let's open the last one to look for a last object.

163

So that's another person and another person so probably all the images from 2012 are persons.

164

Now here we have a horse and a person so that's basically combining several different classes for the

165

day to be able to recognize several objects in an image.

166

All right so we're going to stop here and now now that we're done with the dataset I'm going to show

167

you the code and I'm going to show you how to run the code.

168

First let me just scroll back up to find my way back to the folder.

169

There we go.

170

Scripts done data done and going back to the main training as the folder.

171

All right.

172

So in this folder you also have besides the data you also have the layers which was the same folder

173

as before when we made the detection.

174

So that's because in order to train the SSD we need some of the functions and modules of the SSD.

175

Then we have the SSD implementation that contains all the architecture of the SSD Anchorage to have

176

a look at this.

177

And we have our trained file which I'm going to open right now because data from this file that we're

178

going to execute the training of the SSD on these two datasets Word 2007 and book 2012.

179

All right.

180

And then we have you Teal's which contains some tools for the training like augmentations so augmentations

181

are some ways to increase the amount of images so that we can have even more material to train our SD

182

on even if we have already lots and lots of images.

183

You can actually check that in our deep learning course we have some tutorials explaining in more details

184

how Image augmentations work and then finally we have this for that that contains some weights.

185

So make sure to understand that these are not some weights of some pre-trained model because actually

186

right now we're doing some training but these are just some initialized weights.

187

You imagine that they need to be initialized in some way a way that should be compatible with the future

188

of data the way weights during the training.

189

So these are the weights and now we're going to go through this train file and mostly we're going to

190

execute the code to show you how the training is done.

191

It's actually going to be very easy we simply need to select the whole code and then press command or

192

control press and to to execute.

193

But before we do that I would just like to show you some of the arguments here.

194

So basically these are some parcels containing all the arguments that you can change if you want to

195

change the way the training will be done.

196

So for example you can change the argument related to the learning rate here the default learning rate

197

is 0.01 but you can change it to 0.05 for example if you want to try a different training then you can

198

also change the parameter for the momentum.

199

The way to the gamma parameter.

200

Well these are all the parameters that you can change the hyper parameters to do some type of parameter

201

tuning of the training.

202

But again remember that training is taking lots and lots of time so it's mostly for the research and

203

developers working on the state of the art the models.

204

That's not for us to do right now anyway.

205

But then you have some other parameters I mentioned.

206

Image net which is another big huge dataset on which you can train your DeBerry morals and computer

207

vision models and therefore if you want to train your SSD on a different data set then Virk 2007 invoked

208

2012.

209

Well here is the parameter related to that which contains the path to the dataset data if get that exactly

210

the path data as this father and then there it is this folder containing Vok 2007 invoke 2012.

211

All right so let me go back.

212

So basically that's how it works.

213

You know you have several hyper parameters you can change them to experiment different trainings and

214

the rest is some implementation of other functions like.

215

Exactly and wait and it function that initializers the way it's the proper way.

216

So we're going to see that in more details in module 3 because in modules we will implement from scratch

217

the training of the Ganns.

218

Generally our research on that works because this will be possible to do without kuda and without thousands

219

of lines of code.

220

So you have several functions and mostly we have.

221

There you go the train function that does the training of the SSD on all the images that are contained

222

in the Vok folders.

223

So there we go now to finish the tutorial.

224

I'm going to show you how you can execute this file so very simply you just select the whole code again.

225

Make sure that you have an empty Geographic card and kuda enabled and that's not the case you will get

226

an error and be relieved.

227

Even if we had a non a version of this training.

228

Well it would take months or years.

229

And therefore this would be completely useless.

230

And that's exactly the reason why the developers have not implemented and none could have version of

231

the training.

232

That is because it would be completely useless.

233

So I'm going to you now.

234

There we go.

235

The train is on its way and it's basically going to take several hours so we're just going to stop here.

236

It was just to show you how the training works how the data sets were structured and to show you a little

237

behind the scene how the SS is trained.

238

And now we hope you understand why we had to use a pre-trained model to do our detection.

239

All right so now we're going to move on to the homework.

240

It's going to be a fun and exciting homework about detecting some beautiful horses running on a field.

241

I can't wait to show you this.

242

And until then enjoy computer vision.