1
00:00:01,340 --> 00:00:09,260
In the last couple of videos, we learn how to create new little network will get US sequential EPA,

2
00:00:10,760 --> 00:00:16,790
the sequential EPA was quite easy to use, but it has some limitation.

3
00:00:19,010 --> 00:00:26,840
With sequential EPA, you can only create neural network with simply sequential architecture.

4
00:00:28,130 --> 00:00:37,070
You cannot create complex topologies like this where you have multiple inputs or multiple outputs.

5
00:00:38,180 --> 00:00:40,910
For that, you have to use functional LaPier.

6
00:00:42,350 --> 00:00:43,370
And functionally functional.

7
00:00:44,150 --> 00:00:48,530
We create each of these layers in the form of functions.

8
00:00:49,270 --> 00:00:52,520
Or you can say a building block off your neural network.

9
00:00:53,330 --> 00:00:57,930
And you can use this functions to create a complex structure.

10
00:00:58,250 --> 00:01:01,520
By joining them according to your structure need.

11
00:01:03,600 --> 00:01:12,230
Before going into details, let's first building the model that we created in our last lecture just

12
00:01:12,230 --> 00:01:12,750
very well.

13
00:01:12,960 --> 00:01:14,150
And you're more than name.

14
00:01:14,700 --> 00:01:19,100
And then you also need to clear the session off your get us.

15
00:01:19,740 --> 00:01:23,700
This will free up the resources for our next model training.

16
00:01:25,680 --> 00:01:27,710
So just start get us back.

17
00:01:28,050 --> 00:01:29,400
Not clear session.

18
00:01:32,470 --> 00:01:37,350
Then this, too, now virally to create and create our next modern.

19
00:01:39,780 --> 00:01:43,020
For functional EPA, we will be using this example.

20
00:01:44,540 --> 00:01:53,750
This kind of neural network is known as wide and deep neural network deep because our input is going

21
00:01:53,750 --> 00:02:01,490
through two layer of soft, dense, hidden layer and vital because of what input is directly going for

22
00:02:01,580 --> 00:02:02,630
output as well.

23
00:02:04,040 --> 00:02:13,240
So along with the output of this hidden layers, it also connect all parts of our input that include

24
00:02:13,260 --> 00:02:14,360
to the output layer.

25
00:02:17,050 --> 00:02:24,740
This linkage, which I have my best wide, is not possible when we are using sequentially pure.

26
00:02:27,670 --> 00:02:35,050
The advantage of this architecture is that it makes possible foreign neural network to learn both the

27
00:02:35,050 --> 00:02:36,030
deep patterns.

28
00:02:36,490 --> 00:02:42,490
But this deep linkage and the simple rules by this wide linkage.

29
00:02:45,040 --> 00:02:53,710
In our regular MLP, more than all the data flows through this full stacks of dense layers and thus

30
00:02:54,010 --> 00:03:00,460
some simple patterns in the data may end up being distorted by the sequence of this transformation.

31
00:03:03,800 --> 00:03:07,300
So now let's create this lid's one by one.

32
00:03:08,870 --> 00:03:14,330
We will be using the same data set as we use for our last lecture.

33
00:03:18,000 --> 00:03:20,780
First, we need to clear it up and put layer.

34
00:03:21,270 --> 00:03:30,510
We are calling it input and we are creating it would get us not leers, not input, and then we have

35
00:03:30,510 --> 00:03:31,490
to provide the ship.

36
00:03:32,250 --> 00:03:32,940
We can either.

37
00:03:32,950 --> 00:03:33,360
Right.

38
00:03:34,620 --> 00:03:40,830
Just it in decades because we have eight independent variables or we can write it in this way.

39
00:03:41,130 --> 00:03:45,630
Extreme dot ship and then the first and so on attribute.

40
00:03:46,230 --> 00:03:47,760
So this is our input layer.

41
00:03:48,480 --> 00:03:51,700
We will create this kind of list.

42
00:03:51,930 --> 00:03:55,680
And then we will connect these lists accordingly.

43
00:03:56,760 --> 00:04:04,200
So next we create a dense layer with 30 neurons using RELU activation.

44
00:04:05,970 --> 00:04:07,770
And you can also notice that.

45
00:04:08,790 --> 00:04:12,210
We are calling this input player like a function.

46
00:04:12,990 --> 00:04:16,020
So this input layer is the input for this head on layer one.

47
00:04:16,620 --> 00:04:18,600
And we are calling it like a function.

48
00:04:19,110 --> 00:04:22,250
And that's what we call it, functional EPA.

49
00:04:22,680 --> 00:04:31,020
We create this kind of layers and we use these layers as functions for our next players.

50
00:04:33,330 --> 00:04:37,040
So we have connected our input layer of what he'd done to.

51
00:04:39,560 --> 00:04:41,670
Now we create our second hidden lair.

52
00:04:43,700 --> 00:04:50,690
Again, we are using cameras, dot layers, dot dense, and we are creating it with 13 neurons and activation

53
00:04:50,690 --> 00:04:51,440
is a loop.

54
00:04:52,280 --> 00:04:57,380
And now for this hidden layer to our input should be the layer one.

55
00:04:57,710 --> 00:05:01,250
So we are passing our hidden layer one as a function here.

56
00:05:02,990 --> 00:05:05,480
Now, the next step is this.

57
00:05:05,900 --> 00:05:09,410
Here we want the output of this hidden layer, too.

58
00:05:10,130 --> 00:05:13,590
And also we want all lowered inputs variable here.

59
00:05:14,270 --> 00:05:15,560
You can see this linkage.

60
00:05:15,770 --> 00:05:19,330
We want these two to go into this contact layer.

61
00:05:20,860 --> 00:05:26,560
Concat land is just muddying the output of the left and all the inputs.

62
00:05:29,760 --> 00:05:37,240
So we can ride this Laria as concat and then we'll use guitars, dot, Lear, dot concat.

63
00:05:39,570 --> 00:05:43,610
And we are passing the list off our put off her or two.

64
00:05:44,360 --> 00:05:45,430
And put Claire.

65
00:05:47,210 --> 00:05:55,490
If we wanted to put off her down layer one as what we can right then layer one here and it will add

66
00:05:55,580 --> 00:06:04,190
a linkage like this also, so you can see you can customize all this linkages very easily using a functional

67
00:06:04,190 --> 00:06:04,670
LaPier.

68
00:06:05,920 --> 00:06:12,720
Not next step is to create our output layer, our output should get input from the concrete layer.

69
00:06:13,570 --> 00:06:15,280
And this should be a single neuron.

70
00:06:15,850 --> 00:06:21,260
So we are creating output equal to give us not lietz, not Bence.

71
00:06:22,060 --> 00:06:29,340
And then a single neuron without any activation function and input as the output of contact layer.

72
00:06:29,680 --> 00:06:32,390
So we are passing concrete layer as a function here.

73
00:06:34,470 --> 00:06:36,470
Now, we have created all the layers.

74
00:06:37,410 --> 00:06:41,700
The next step is to combine all this layers and create a modern.

75
00:06:44,100 --> 00:06:49,910
So we are creating a model object and then we are calling it our start model start model.

76
00:06:50,460 --> 00:06:56,100
And here we are mentioning what we want as our input and what we want as our output.

77
00:06:56,730 --> 00:07:01,090
So what input is this first input layer and our output?

78
00:07:01,320 --> 00:07:02,960
Is this last output layer?

79
00:07:04,710 --> 00:07:12,800
Think sequentially appear we first create a model and then we create each layer, layer by layer, but

80
00:07:12,830 --> 00:07:13,910
then functionally appear.

81
00:07:15,170 --> 00:07:21,350
We create this layers and then we join this layers to create this whole network.

82
00:07:22,750 --> 00:07:29,680
Just run this now again, just to look at the structure of the model that you have created.

83
00:07:29,950 --> 00:07:31,750
You can call not somebody.

84
00:07:33,070 --> 00:07:34,850
So what object name is Morton?

85
00:07:35,380 --> 00:07:36,780
And we are calling somebody.

86
00:07:38,650 --> 00:07:40,960
You run this, you will get all the details.

87
00:07:41,440 --> 00:07:45,030
First, we have input layer with eight input variables.

88
00:07:45,940 --> 00:07:50,530
Then we have a dense layer with 30 neurons than we ever were a second.

89
00:07:50,650 --> 00:07:53,140
Then slier again with 13 neurons.

90
00:07:54,430 --> 00:08:01,130
Then we have a concrete layer where we are contacting the input of the second non-slip.

91
00:08:01,600 --> 00:08:02,890
And our input layer.

92
00:08:03,790 --> 00:08:08,320
So you can see we're 30 less a so her beaird neurons here.

93
00:08:10,300 --> 00:08:15,110
And then we have our output that we have on leveland neuron.

94
00:08:16,540 --> 00:08:18,480
Now the next step is to compile.

95
00:08:19,480 --> 00:08:29,260
And just as in our previous lecture, we will be using mini skirt to edit as lost function as Judy with

96
00:08:29,260 --> 00:08:39,580
learning rate of zero point zero zero one as our optimizer and Emmi or mean absolute edit as our additional

97
00:08:39,760 --> 00:08:40,870
metrics to calculate.

98
00:08:42,760 --> 00:08:43,450
From this.

99
00:08:45,650 --> 00:08:51,410
Now, fitting the model, the same will just say more than Lord fit for weight training.

100
00:08:51,550 --> 00:08:54,270
Does it provide a bit of epoch values?

101
00:08:55,290 --> 00:09:02,350
And then the relevation dare to say, since last time we ran that tradition, more than 440 books.

102
00:09:04,080 --> 00:09:06,240
So changing it for four equals.

103
00:09:07,770 --> 00:09:11,730
So that's just been more than.

104
00:09:16,460 --> 00:09:22,910
Again, you will see the lost functions, whether the loss and the value of my kid that you have up

105
00:09:23,090 --> 00:09:23,360
here.

106
00:09:26,460 --> 00:09:27,640
You can see the lost value.

107
00:09:27,660 --> 00:09:30,510
We are getting on our training set at zero point three six.

108
00:09:30,540 --> 00:09:35,410
This is Tamasi value and the validation loss is zero point three six three eight.

109
00:09:37,070 --> 00:09:40,780
Let's calculate the value on our test data.

110
00:09:44,000 --> 00:09:52,020
You can see the lost value here is zero point three zero six four are near in our last case.

111
00:09:56,560 --> 00:10:00,060
I guess we were getting law says zero point two five one five.

112
00:10:01,630 --> 00:10:10,350
So this model is not performing that good, but in some situations, this kind of deep and wide net

113
00:10:10,380 --> 00:10:13,450
cut but falls better than a normal MLP.

114
00:10:13,510 --> 00:10:14,170
Martin's.

115
00:10:15,930 --> 00:10:21,640
In this case of a normal MLP model, is platforming better than this wide and deep net for.

116
00:10:23,310 --> 00:10:26,390
Again, you have all the parameters available to you.

117
00:10:26,610 --> 00:10:33,090
So if you just read model underscored a student history, you will get all the lowest values.

118
00:10:34,850 --> 00:10:43,640
And the Emmy, you and Lou, that we got during their training, and you can also load this on the just

119
00:10:43,640 --> 00:10:44,840
like we did earlier.

120
00:10:47,460 --> 00:10:56,370
So the important thing here is that what reputation, loss or validation and me value is still decreasing.

121
00:10:56,880 --> 00:11:02,280
So there is a scope of further improvement in the accuracy of our model.

122
00:11:03,390 --> 00:11:05,280
So let's just run this.

123
00:11:05,470 --> 00:11:08,550
More than 440 more epoch as well.

124
00:11:20,320 --> 00:11:22,540
Now, let's conclude the.

125
00:11:23,420 --> 00:11:24,840
Yes, well, use as one.

126
00:11:27,840 --> 00:11:32,730
You can see now the laws has decreased to zero point six.

127
00:11:33,630 --> 00:11:36,270
Earlier, we will getting around zero point three.

128
00:11:37,260 --> 00:11:40,890
Now we have zero point two six as a loss when you.

129
00:11:42,980 --> 00:11:50,190
You can also put out this graph again to see whether the model of converts this time or not.

130
00:11:54,220 --> 00:12:00,250
You can see this is almost so straight line, so we can see that model has converged.

131
00:12:00,940 --> 00:12:07,940
And the best value for the word modern and the masc value what our model is.

132
00:12:08,470 --> 00:12:10,880
Zero one two six one on the test site.

133
00:12:13,030 --> 00:12:18,730
This is somewhat similar to the accuracy we bought earlier with our normal MLP model.

134
00:12:20,050 --> 00:12:20,210
But.

135
00:12:21,100 --> 00:12:27,010
I would say that our normal MLP be who was performing better than this deep and wide net for.

136
00:12:29,170 --> 00:12:31,770
So that's all for this lecture in the next lecture.

137
00:12:31,800 --> 00:12:39,210
We will see how to say what what model and how to save check points at the end of each epochs.

138
00:12:39,840 --> 00:12:40,260
Thank you.