1
00:00:00,300 --> 00:00:01,166
Hello my friends.

2
00:00:01,166 --> 00:00:04,933
Welcome back to this implementation
and mostly welcome to part three

3
00:00:04,933 --> 00:00:07,800
training,
the CNN. In the previous tutorial,

4
00:00:07,800 --> 00:00:11,700
we built the brain,
which contained the eyes of the eye.

5
00:00:11,700 --> 00:00:14,066
You know,
thanks to the convolutional layers.

6
00:00:14,066 --> 00:00:17,100
And now we're going to make the brain
smart with that

7
00:00:17,100 --> 00:00:21,100
training of the CNN
on all our training set images.

8
00:00:21,333 --> 00:00:24,200
And at the same time, as you see,
we're going to evaluate

9
00:00:24,200 --> 00:00:27,366
our same model on the test
set over the epochs.

10
00:00:27,366 --> 00:00:30,600
You know, we're going to train our CNN
over 25 epochs.

11
00:00:30,800 --> 00:00:31,866
And at each epoch

12
00:00:31,866 --> 00:00:36,833
we'll actually see how our model
is performing on our test set images.

13
00:00:37,166 --> 00:00:40,866
So this is a different kind of training
as what we before because we always

14
00:00:40,866 --> 00:00:43,866
used to separate the training
and the evaluation.

15
00:00:44,133 --> 00:00:46,200
But here
this will happen at the same time.

16
00:00:46,200 --> 00:00:50,300
That's because we're doing some specific
application which is computer vision.

17
00:00:50,900 --> 00:00:51,433
All right.

18
00:00:51,433 --> 00:00:52,266
Are you ready?

19
00:00:52,266 --> 00:00:53,200
Let's do this.

20
00:00:53,200 --> 00:00:57,000
Let's start with the first step
compiling the CNN.

21
00:00:57,566 --> 00:00:57,866
All right.

22
00:00:57,866 --> 00:01:00,800
So in a new code cell
we're going to compile the CNN.

23
00:01:00,800 --> 00:01:04,966
Meaning we're going to connect it
to an up to miser a loss function.

24
00:01:04,966 --> 00:01:07,500
And some metrics.

25
00:01:07,500 --> 00:01:11,700
And well you know here once again
we're doing binary classification.

26
00:01:11,700 --> 00:01:15,433
And so very simply
we're going to compile our CNN

27
00:01:15,700 --> 00:01:19,266
exactly the same way
as how we build our A.

28
00:01:19,266 --> 00:01:21,000
And then in the previous section,

29
00:01:21,000 --> 00:01:24,066
because indeed we're going to choose
once again an Adam optimizer

30
00:01:24,066 --> 00:01:27,700
to, you know, perform stochastic gradient
descent to update the weights

31
00:01:27,966 --> 00:01:32,833
in order to reduce the loss error
between the predictions and the target.

32
00:01:33,300 --> 00:01:35,733
Then we're going to choose the same loss.

33
00:01:35,733 --> 00:01:38,933
You know,
the binary cross-entropy loss once again

34
00:01:38,933 --> 00:01:42,366
because we're doing executives
same task binary classification.

35
00:01:42,833 --> 00:01:44,533
And then same for the metrics.

36
00:01:44,533 --> 00:01:47,666
We're going to choose the accuracy metrics
because that's, you know,

37
00:01:47,666 --> 00:01:50,800
the most relevant way
to measure the performance

38
00:01:50,800 --> 00:01:54,433
of a classification model,
which is exactly the case of our CNN.

39
00:01:55,000 --> 00:01:56,733
And therefore, to compile the CNN.

40
00:01:56,733 --> 00:01:58,633
Well, it's going to be a piece of cake,

41
00:01:58,633 --> 00:02:01,700
because indeed we're going to do exactly
the same as before.

42
00:02:01,700 --> 00:02:04,700
So we're going to start by taking our CNN,

43
00:02:04,800 --> 00:02:08,533
from which we're going to call the compile
method,

44
00:02:09,000 --> 00:02:13,433
which will take as inputs
well first or up to miser,

45
00:02:13,766 --> 00:02:17,466
which will choose to be
the Adam optimizer.

46
00:02:18,000 --> 00:02:23,000
Then the loss function,
which will choose to be the binary

47
00:02:24,000 --> 00:02:26,566
cross entropy.

48
00:02:26,566 --> 00:02:27,466
All right.

49
00:02:27,466 --> 00:02:30,566
And finally final argument the metrics.

50
00:02:30,933 --> 00:02:32,133
We will choose only one.

51
00:02:32,133 --> 00:02:35,133
But remember we have the choice
to take several of them.

52
00:02:35,166 --> 00:02:36,666
But just accuracy is fine.

53
00:02:36,666 --> 00:02:40,800
And therefore in these pair of square
brackets we're going to input in quotes.

54
00:02:40,800 --> 00:02:42,800
Well that's accuracy.

55
00:02:42,800 --> 00:02:43,800
And done.

56
00:02:43,800 --> 00:02:47,833
This compiles
successfully the CNN to the optimizer

57
00:02:47,833 --> 00:02:49,600
the loss function and the metric.

58
00:02:49,600 --> 00:02:51,633
So exactly the same as before.

59
00:02:51,633 --> 00:02:55,200
However now
to train the CNN on the training set

60
00:02:55,200 --> 00:02:58,433
and evaluating it
at the same time on the test set.

61
00:02:58,800 --> 00:03:01,500
Well it will not be exactly
the same as before, but

62
00:03:01,500 --> 00:03:03,700
once again, very, very similar.

63
00:03:03,700 --> 00:03:04,633
Let's check this out.

64
00:03:04,633 --> 00:03:06,166
Let's create a new code cell.

65
00:03:06,166 --> 00:03:09,600
And now you can actually guess
the first two steps.

66
00:03:09,600 --> 00:03:10,900
They're always the same.

67
00:03:10,900 --> 00:03:14,733
The first step is to take
of course our CNN from which.

68
00:03:14,733 --> 00:03:16,266
And that's the second step.

69
00:03:16,266 --> 00:03:18,600
What method do we need to call here.

70
00:03:18,600 --> 00:03:21,000
Well once again this never changes.

71
00:03:21,000 --> 00:03:24,800
This is of course to fit method,
the fit method,

72
00:03:24,800 --> 00:03:28,566
which as always will train
the CNN on the training set.

73
00:03:29,133 --> 00:03:29,666
Okay.

74
00:03:29,666 --> 00:03:32,866
So this time
what are going to be the inputs.

75
00:03:33,300 --> 00:03:35,766
Well the first input is always the same.

76
00:03:35,766 --> 00:03:36,400
It's going to be

77
00:03:36,400 --> 00:03:39,733
of course the set, you know, the data set
on which you're going to train

78
00:03:39,866 --> 00:03:43,600
your model here, the CNN and that's
of course the training set.

79
00:03:43,600 --> 00:03:46,633
And the name of the parameter for
that is simply x

80
00:03:46,900 --> 00:03:51,400
and therefore will specify x to be,
you know, our training set,

81
00:03:51,900 --> 00:03:56,933
that exact same training set
which we created in part one.

82
00:03:57,133 --> 00:04:02,766
You know right here that training set,
that training set to which we applied

83
00:04:02,766 --> 00:04:07,500
this image, data generated to
to indeed perform image augmentation.

84
00:04:08,000 --> 00:04:08,400
All right.

85
00:04:08,400 --> 00:04:09,000
That training set.

86
00:04:09,000 --> 00:04:13,400
That's the first input of our parameter
in this fit method

87
00:04:13,833 --> 00:04:16,500
okay then next parameter.

88
00:04:16,500 --> 00:04:16,766
All right.

89
00:04:16,766 --> 00:04:20,200
So the next parameter
is this time you know the difference.

90
00:04:20,200 --> 00:04:22,466
The difference with what we did before.

91
00:04:22,466 --> 00:04:26,100
So it has to do of course with the fact
that we're not only training the CNN

92
00:04:26,100 --> 00:04:30,733
on the training set, but also evaluating
it at the same time on the test set.

93
00:04:31,000 --> 00:04:34,033
And that second parameter corresponds
exactly to this.

94
00:04:34,200 --> 00:04:37,800
We have to specify here
the validation data.

95
00:04:37,933 --> 00:04:39,266
That's the name of the parameter.

96
00:04:39,266 --> 00:04:44,433
But that's of course the set on
which we want to evaluate our CNN.

97
00:04:44,433 --> 00:04:47,433
And as of course the test set
that will be the value of the parameter.

98
00:04:47,700 --> 00:04:50,433
But the name of that parameter is,
as I just said,

99
00:04:50,433 --> 00:04:54,000
validation underscore data.

100
00:04:54,000 --> 00:04:58,733
And this is equal,
of course, to the test set.

101
00:04:58,733 --> 00:05:03,300
Once again, that exact same test set
which we created here in part

102
00:05:03,300 --> 00:05:07,166
one when preprocessing it
to set this one to which of course

103
00:05:07,300 --> 00:05:10,833
no transformation was applied,
only features killing.

104
00:05:11,400 --> 00:05:13,900
Okay. So test set good.

105
00:05:13,900 --> 00:05:17,533
And now we have one final argument,
which is of course

106
00:05:17,533 --> 00:05:19,466
you can totally guess what it is, right?

107
00:05:19,466 --> 00:05:23,766
This is the inevitable argument
when training a deep neural network.

108
00:05:24,300 --> 00:05:29,100
I'm talking of course about the epochs
parameter, which is the number of epochs,

109
00:05:29,633 --> 00:05:34,233
and well, you know, to justify to you
which number I chose.

110
00:05:34,433 --> 00:05:37,433
Well, I actually start with ten epochs.

111
00:05:37,500 --> 00:05:40,900
And I noticed that
the accuracy was not converging.

112
00:05:41,166 --> 00:05:43,500
So then I tried with 15 epochs,

113
00:05:43,500 --> 00:05:46,000
because you're going to see that one
epoch is actually pretty slow.

114
00:05:46,000 --> 00:05:47,366
You know, it's actually a bit long.

115
00:05:47,366 --> 00:05:51,033
I mean, much longer than the epochs
in the training of our previous neural

116
00:05:51,033 --> 00:05:51,766
network, you know?

117
00:05:51,766 --> 00:05:53,800
And so I started with ten.

118
00:05:53,800 --> 00:05:57,200
It was not enough then 15 still not
enough, you know, still not converging.

119
00:05:57,300 --> 00:06:00,300
And then 25 and 25 was perfect.

120
00:06:00,700 --> 00:06:03,700
I had an accuracy
that pretty much converged

121
00:06:04,133 --> 00:06:07,066
not only on the training set,
but also on the test set, you'll see.

122
00:06:07,066 --> 00:06:10,066
So for the number of epochs
here, we're going to choose 25.

123
00:06:10,233 --> 00:06:12,066
Feel free to increase it if you have time.

124
00:06:12,066 --> 00:06:15,066
You know, if you want to let your computer
run for now.

125
00:06:15,133 --> 00:06:17,700
More with 25 epochs here, it will be fine.

126
00:06:17,700 --> 00:06:19,966
It will just take 10 to 15 minutes.

127
00:06:19,966 --> 00:06:20,933
So that will be all good.

128
00:06:20,933 --> 00:06:23,666
We'll get our results pretty fast.

129
00:06:23,666 --> 00:06:25,666
All right. And well actually that's it.

130
00:06:25,666 --> 00:06:28,666
That's only what we need to train our CNN

131
00:06:28,666 --> 00:06:32,700
on the training set
while evaluating it on the test set.

132
00:06:33,133 --> 00:06:34,033
So perfect.

133
00:06:34,033 --> 00:06:35,666
We smashed by three.

134
00:06:35,666 --> 00:06:39,133
And we can now already move on to part
four where

135
00:06:39,133 --> 00:06:43,500
we're going to make our single prediction,
which I remind will consist

136
00:06:43,500 --> 00:06:49,500
of deploying our model
on the two images of the single prediction

137
00:06:49,500 --> 00:06:53,700
folder, this one for which, of course,
our model will have to recognize

138
00:06:53,700 --> 00:06:57,966
that there is a dog, and this one
where our model will have to recognize

139
00:06:58,000 --> 00:06:59,133
there is a cat.

140
00:06:59,133 --> 00:07:00,866
So let's hope that it is right,

141
00:07:00,866 --> 00:07:03,266
but we won't have the predictions
in the next tutorial.

142
00:07:03,266 --> 00:07:05,566
We will have them in the one after that.

143
00:07:05,566 --> 00:07:08,700
Because remember,
we're going to run our implementation

144
00:07:08,700 --> 00:07:11,900
from Jupyter Notebook
because we can do it in Google Colab.

145
00:07:11,900 --> 00:07:14,466
The data set is to be all right.

146
00:07:14,466 --> 00:07:18,800
So well, as soon as you're ready for part
four, let's do this together.

147
00:07:18,966 --> 00:07:21,000
And until then, enjoy machine learning.