Original subtitles

1

Hello and welcome to this new tutorial now that we've done our four transformations the input is ready

2

to be fed into the neural network.

3

It has the authorization to get in.

4

And therefore that's exactly what we're going to do.

5

We're going to get it.

6

And so there is nothing more simple you know we already have our pre-trained model SSD and it is already

7

pre-trained because we could load the weight thanks to this file but that's not actually what we'll

8

do.

9

Now we will load the weight in the end.

10

Right now that is just a variable so we will just use the variable but then at the end of this implementation

11

we will load the weights to get our pre-trained model.

12

So to feed x r towards variable that contains both the torch tensor of the input frame and the gradient

13

into the neural network net.

14

Well we simply need to take our neural network net and then apply X and that's it.

15

That's how we feed X to the neural network.

16

But then since this neural network nets applied to the input X will return the output y.

17

Well we're going to get this output y right now and therefore and adding y equals net X. That gives

18

us the output way we will of course describe what is why directly you can already start to try to think

19

what it is exactly.

20

But now we have the output that's great.

21

And so we can move on to the next step.

22

So the next step what is the next step.

23

Well we just got our output y Why doesn't contain directly what we're interested in that is the result

24

of the detection whether we have a dog or a human in the input frame to get that specific information

25

we're interested in.

26

Well we need to take the data attribute from Y.

27

And so what we're going to do now is create a new sensor that we're going to call detections.

28

So detection is a new tensor and that's a tensor contained in the output y.

29

And that will contain the values we're interested in and to get this tensor while we take our output

30

y.

31

And then we add that and we take our attribute data and then we get the values of the output.

32

Perfect.

33

Now we have what we want the next step now is to create a new tensor object which will have the dimensions

34

width height width height.

35

So I didn't say twice.

36

It's just a tensor of four dimensions.

37

The first dimension is with the second dimension is height the third dimension is width and the fourth

38

dimension is height.

39

And now of course most of you must be thinking why do we have to create such a tensor.

40

Well that's because the position of the detected objects inside the image has to be normalized between

41

0 and 1.

42

And to do this normalization will need this scale tensor with these four dimensions.

43

Basically the Newtons are we're about to create right now.

44

Scale will be just use to do this normalization between zero and one of the positions of the object

45

detected in the image.

46

That's the only purpose.

47

And now why do we have with height width height.

48

That's because the first two with height will correspond to the scale of values of the upper left corner

49

of the rectangle detector.

50

And the second with height will correspond to the scale of values of the lower right corner of this

51

same rectangle detector.

52

That's why we have a double with height.

53

So let's create this scale sensor so that you can visualize it.

54

So on a general rule to create a tensor in Torch Well we need to take our torche library and then we

55

use the tensor class so scale will be an object of the tenso class which therefore will be a tensor

56

a torch tensor.

57

But as the arguments of this tensor class we need to specify the four dimensions of the tensor and these

58

four dimensions are it's hights what's heights.

59

Perfect.

60

So this first with hied correspond to the upper left corner of the rectangle and this second with height

61

corresponds to the lower right corner of the rectangle.

62

And we're doing this to normalize the scale of values of the position of the detected objects between

63

0 and 1.

64

Perfect.

65

So another good thing done.

66

Don't worry about the warnings here.

67

That's just because we haven't use these detections and scale variables yet we will do it very quickly.

68

But before we do that I highly recommend to take a break because what we're about to do now will be

69

slightly more complicated than what we've been doing so far.

70

So we're going to finish with this tutorial now.

71

Take a good break.

72

Possibly a little nap or good coffee and then we'll attack more.

73

The heart of the ass is tomorrow.

74

Hope that didn't sound too aggressive.

75

But yeah we're going to get into the heart of the as is the neural network.

76

So have a good break and I'll see you in the next tutorial.

77

Until then enjoy computer vision.