Original subtitles

1

Hello and welcome to this new editorial in the British this editorials we described the challenge that

2

we have.

3

We have to detect a funny dog on a two seconds video and we will do it through a computer vision based

4

on deep learning neural networks.

5

So we already found the right folder now.

6

This quick to toile I'm going to explain the libraries that we're going to use.

7

They're already all installed I already prepared the code that will import them.

8

So there is nothing to do but I think it's important that you understand what we will be using them

9

for.

10

So let's start with the first one as you can see the first library when port is torche that's of course

11

the torch library that contains PI torch which is definitely by far our best weapon to build a new one

12

that work and do some computer vision and that's for a specific reason it's because by torch content

13

the dynamic graphs things to which we are able to compute very efficiently the gradients of composition

14

functions in backward propagation.

15

You know when we have to update day to wait through stochastic gradient descent Well we have to compute

16

the gradient of some composition functions because we have several layers.

17

You know it's a deep neural network so we have several layers.

18

And therefore it's like one has some functions of the PRI's layer which has some functions of the previous

19

previous layer so that generates some composition functions.

20

We have to compute the gradient of these composition functions to have data weights according to how

21

much they are responsible for the error between the target and the predictions.

22

So that's where it plays.

23

And the dynamic graphs is a highly advanced graph structure that allows to have some very fast and efficient

24

computation of the gradients.

25

So that's why torture is our first choice then from torche undergrad which is the module responsible

26

for graden descent.

27

We are importing the variable class which will be used to convert the tensors into some torche variables

28

that will contain both the tensor and a gradient.

29

And then the storage variable containing the tensor in the gradients will be one element of the graph.

30

Then of course we import CB2 and that even if we're not going to implement a model based on open Hargus

31

gate we're just importing CB2 because we will be drawing some rectangles around the deck.

32

But the detection of the dog will not be based on open city Harker's Cate's.

33

They will be based on as is the neural network that is single shot multa box detection so opens we just

34

to draw the rectangles then hear from Data Import base transform the classes as label map data is just

35

a folder that contains the classes based transform and classes then base transform is a class that will

36

do the required transformations so that the image the input images will be compatible with the neural

37

network.

38

You know when we feed the neural network with the input images they have to have a certain format and

39

base transform will be used to transform the images in this format so that they can be accepted into

40

the network.

41

And then what Les's.

42

Well look Les's is just a dictionary that will do the encoding of the classes.

43

So for example planes will be encoded as one Dug's will be included as to is just an example it's not

44

exactly just mapping but that's the idea.

45

We'll do a mapping because of course we want to work with numbers and not text.

46

So that's just a very simple dictionary doing the mapping between the text fields of the classes and

47

some integers.

48

All right then from the import build SSD.

49

So first SSD is the library of the single shot multi-book action model and then build as that we import

50

from the SSD library will be the constructor of the SSD neural network.

51

And so if you want to have a look you can have a look in this as is digitized and fell to see how it

52

works.

53

But it is just a constructor that will build the architecture of this single shot not box detection

54

model.

55

And finally image I know is just the library that we'll use to process the images of the video and applying

56

the detect function that will implement on the images.

57

So at first we wanted to import pill P L which is another library but image I O actually turns out to

58

be a much better choice in terms of lines of code.

59

You'll see we will only have to type two or three lines of code to be able to process the images of

60

the video.

61

That is funny Doug and before and apply to detect function that will implement to detect the dog and

62

the humans on the video.

63

All right.

64

So I hope you have now a clear understanding of the libraries that will be used.

65

It's important to know how they work.

66

And now with you're going to do is define the detect function that will do the detections.

67

So let's take a fresh start in the next tutorial.

68

And so until then enjoy computer vision.