WEBVTT 00:00.120 --> 00:01.120 Welcome back. 00:01.160 --> 00:05.760 In our model we have nine features cylinders displacement. 00:05.760 --> 00:06.480 Horsepower. 00:06.480 --> 00:06.880 Weight. 00:06.880 --> 00:07.720 Acceleration. 00:07.720 --> 00:10.240 Model year one, two and three. 00:10.280 --> 00:17.720 Now those are the features that the input layer takes as parameters and as features. 00:17.720 --> 00:26.760 Also in the hierarchy of our model we have hidden layer number one with 64 neurons learning complex 00:26.760 --> 00:27.600 patterns. 00:27.800 --> 00:37.360 Then we have another layer, another 64 neurons called hidden layer number two that learn more patterns. 00:37.360 --> 00:42.800 And then we have the output layer which is the mileage and the mpg. 00:42.920 --> 00:43.480 Okay. 00:43.800 --> 00:46.720 We talked about this hierarchy in the previous videos. 00:46.720 --> 00:48.600 Now let's code. 00:48.640 --> 00:56.080 Let's create a function that creates a reusable model that follows this hierarchy. 00:56.120 --> 01:03.400 We have input layer hidden layer number one, hidden layer number two and the input output layer under 01:03.440 --> 01:05.080 the feature scaling. 01:05.240 --> 01:10.970 Here we need to start building the neural network and build the model architecture. 01:11.170 --> 01:12.370 New code. 01:12.530 --> 01:15.650 Let's start defining a new function. 01:15.850 --> 01:18.330 Def build model. 01:18.450 --> 01:26.170 And to build a model we start with model equals to TensorFlow Keras dot sequential. 01:26.370 --> 01:30.090 And we talked about this in the previous videos. 01:30.250 --> 01:32.770 But don't worry we're going to lecture it. 01:32.810 --> 01:34.530 This is the first layer. 01:34.530 --> 01:36.370 This is the input layer. 01:36.370 --> 01:40.850 So here let me make it as input layer. 01:40.970 --> 01:46.170 We have TF dot dot layers dot dance. 01:46.170 --> 01:48.370 We are creating a dense layer. 01:48.370 --> 01:59.130 So this code means that we are going to create a TensorFlow layer a dense layer 64 neurons first hidden 01:59.130 --> 02:03.890 layer with 64 units and activation ReLU. 02:04.010 --> 02:10.090 ReLU introduces non-linearity returns maximum zero and x. 02:10.090 --> 02:13.570 So here we have nonlinear. 02:13.810 --> 02:19.940 The input shape automatically adapts to the number of feature in scaled data. 02:20.140 --> 02:29.540 We can use train X train keys, but I prefer using the exit train X train scaled. 02:29.580 --> 02:31.980 We we used the scaled. 02:31.980 --> 02:34.900 We're going to use the scaled data. 02:34.900 --> 02:40.620 So here get the shape and pass the parameter one. 02:40.660 --> 02:46.460 So this will automatically adapt to number of feature in scaled data. 02:46.460 --> 02:50.700 So we are using the scaled data not the exit train. 02:50.860 --> 02:52.460 Please pay attention guys. 02:52.740 --> 02:55.060 Don't use X train data. 02:55.100 --> 02:59.140 Use the exit train underscore scaled okay. 02:59.180 --> 03:05.660 Because we uh we introduced the feature scaling in the uh, previous videos. 03:05.660 --> 03:09.860 And we talked about exit train scaled okay. 03:09.900 --> 03:13.820 So our data now becomes X train scaled. 03:13.860 --> 03:20.100 We need to deal and train our model with this trained and scaled data. 03:20.260 --> 03:24.820 The second layer we need to create is the hidden layer. 03:24.900 --> 03:31.900 So hidden layer number one we start with tf dot dot layer dot dense. 03:31.900 --> 03:41.100 In order to create any layer we start with tensorflow.keras dot layers dot dense 64 activation ReLU 03:41.420 --> 03:49.820 and another hidden layer df dot dot layer dot dense. 03:49.900 --> 03:53.980 Here we can use not dense layer. 03:53.980 --> 03:58.380 We can use drop out 0.2. 03:58.420 --> 04:10.420 Why we don't use dense because we need to prevent overfitting and dropout 0.2 randomly drops 20% of 04:10.700 --> 04:14.020 neurons during training to prevent overfitting. 04:14.020 --> 04:19.500 So here we have 20% for the training data, for the testing data. 04:19.740 --> 04:27.300 And here we set the randomly drops 20% of neurons during training to prevent overfitting. 04:27.540 --> 04:35.310 The third and or the fourth and last layer is the output layer. 04:35.310 --> 04:37.110 So output layer. 04:37.310 --> 04:38.510 How to create it. 04:38.550 --> 04:40.390 Start with df dot. 04:41.590 --> 04:42.430 Dot layer. 04:42.470 --> 04:49.750 Dot dense equals to one one neuron for regression and no activation. 04:49.750 --> 04:58.270 So here one neuron okay so we have only one neuron for regression and no activation. 04:58.430 --> 05:01.710 Don't miss to close this sequential. 05:01.710 --> 05:05.350 So this is our um our sequential. 05:05.390 --> 05:07.230 How to create the model layers. 05:07.230 --> 05:11.030 We start by Keras tf.keras dot sequential. 05:11.070 --> 05:15.350 This is the function used for creating the the layers. 05:15.350 --> 05:25.310 We created the input layer with one with 64 neurons activation, ReLU and input shape. 05:25.350 --> 05:28.230 We use the scaled trained x. 05:28.270 --> 05:30.910 Then we create the hidden layer number one. 05:30.950 --> 05:35.310 We create the hidden layer number two and the output layer.