WEBVTT 00:00.120 --> 00:03.360 We created the layers of our model. 00:03.400 --> 00:05.680 Now let's compile this model. 00:05.680 --> 00:13.400 Compiling means preparing the model for training by specifying how to optimize the model by the optimizer, 00:13.600 --> 00:19.840 how to measure performance, the loss function, and what metrics to track during training. 00:20.080 --> 00:21.560 Which are the metrics. 00:21.680 --> 00:30.720 Okay, so in order to compile it, we're going to use the same function inside the same function here 00:30.760 --> 00:34.360 compiling the model model dot compile. 00:34.520 --> 00:39.640 And we have the optimizer Adam Adam. 00:40.720 --> 00:43.040 Not a name of a friend. 00:43.160 --> 00:44.760 It's adaptive. 00:44.760 --> 00:53.880 Moment estimation is a popular optimization algorithm combining benefits of two other methods Adagrad 00:53.880 --> 00:56.000 and Rmsprop. 00:56.040 --> 00:58.840 Don't worry, you're not going to use them. 00:58.880 --> 01:07.880 Automatically adjusts learning rates for each parameter, and it's good default choice for most problems. 01:08.080 --> 01:12.980 The loss function here We talked about it. 01:13.020 --> 01:17.700 Mean squared error standard for regression problems. 01:17.820 --> 01:24.540 Calculates or calculated as average of true values minus predictions. 01:24.540 --> 01:25.340 Square. 01:25.540 --> 01:29.260 So true true values minus predictions squared. 01:29.300 --> 01:29.620 Get. 01:29.620 --> 01:39.140 The average of these values heavily penalizes large errors due to squaring, and works well when you 01:39.140 --> 01:41.820 want to predict continuous values. 01:41.980 --> 01:47.940 And the metrics that we need are the Em and MSE. 01:48.300 --> 01:50.140 The mean absolute error. 01:50.140 --> 01:58.460 Additional performance metric calculated as average of absolute value of true values minus predictions. 01:58.500 --> 02:05.660 So if you remember, we talked about those metrics in the previous videos and in the introductory section. 02:05.660 --> 02:09.460 So please go back if you want to learn more about them. 02:09.460 --> 02:12.460 But don't worry it's very simple. 02:12.620 --> 02:14.340 Um formula. 02:14.780 --> 02:23.290 So MSE mean squared error average of true values minus prediction squared May mean absolute error. 02:23.330 --> 02:27.210 Average of the absolute value of true values minus predictions. 02:27.210 --> 02:29.530 It's very, very simple. 02:29.970 --> 02:34.970 The less sensitive to outliers than MSE. 02:35.330 --> 02:37.730 Easier to interpret than MSE. 02:37.770 --> 02:41.570 Same units as target values during training. 02:41.930 --> 02:49.050 Adam, our friend Adam will adjust weights to minimize MSE. 02:49.330 --> 02:54.890 You'll see both loss MSE and Ma in training logs. 02:54.930 --> 02:58.930 Lower values for both indicate better performance. 02:59.330 --> 03:06.290 This configuration is ideal for predicting continuous numerical values like prices, temperatures, 03:06.450 --> 03:11.730 or any quantity where the output is not categorical. 03:11.770 --> 03:13.490 The main point you point. 03:13.490 --> 03:23.610 You need to understand that as MSE and Ma get lower values, this indicates that we have a better performance 03:23.610 --> 03:25.050 for our model.