WEBVTT 00:00.080 --> 00:01.120 Welcome back. 00:01.200 --> 00:04.680 We finished creating the model. 00:04.680 --> 00:06.360 Now let's compile it. 00:06.400 --> 00:11.520 We use model the name of our model dot compile. 00:11.520 --> 00:19.520 And here we have to pass three parameters the optimizer the loss and the metrics. 00:19.560 --> 00:24.320 Ma let me explain step by step the model compilation. 00:24.440 --> 00:34.120 What model compilation does compilation configures the model for training by specifying how to learn. 00:34.200 --> 00:43.360 Optimizer, what to minimize loss function how to measure progress, which is the metrics. 00:43.400 --> 00:52.000 Okay, so compile the model means configuring the model for training by specifying the optimizer loss 00:52.000 --> 00:54.200 function and the metric. 00:54.200 --> 00:58.320 How to learn, what to minimize, how to measure progress. 00:58.320 --> 01:03.600 So we're going to specify them using the compile function. 01:03.640 --> 01:12.480 Here we'll start by the optimizer equals to tf TensorFlow dot optimizer dot s g SGD. 01:12.760 --> 01:21.120 The algorithm that updates the model's weight during training, reducing errors and the learning rate 01:21.120 --> 01:24.600 which is the parameter inside this function. 01:24.600 --> 01:30.560 Learning rate controls how big each weight update step is. 01:30.840 --> 01:41.160 The learning rate zero point if I specify 0.1, which is too high, might overshoot the optimal weights, 01:41.160 --> 01:49.640 and if I specify 0.001, which is too low, training becomes very slow. 01:49.840 --> 01:56.000 So in order to balance this problem we use 0.01. 01:56.040 --> 01:58.600 Not very fast and not very slow. 01:58.840 --> 02:05.370 Training would be Accurate and not optimal and not very slow. 02:05.570 --> 02:08.930 Then we need to specify the loss. 02:09.130 --> 02:09.850 Loss. 02:09.850 --> 02:19.090 Here we have MC MSE mean squared error measures how wrong the predictions are. 02:19.130 --> 02:29.130 Using this formula MSE equals to the average of true price minus predicted price to power squared. 02:29.170 --> 02:29.690 Okay. 02:29.890 --> 02:39.370 So we are calculating this using the the the training and the testing data. 02:39.370 --> 02:41.410 So we are comparing those. 02:41.450 --> 02:47.570 And we get by this formula this average in order to track our loss. 02:47.610 --> 02:49.450 Why MSE for regression. 02:49.450 --> 02:53.690 Because it penalizes large errors more heavily. 02:53.850 --> 03:00.490 Smooth curve helps gradient descent and the standard for regression problems. 03:00.490 --> 03:04.450 So we use it with standard regression problems. 03:04.450 --> 03:05.890 Exactly like this problem. 03:05.930 --> 03:10.650 M, m or m a mean absolute error. 03:10.690 --> 03:13.130 Alternative way to measure errors. 03:13.330 --> 03:20.930 The formula is average of the absolute value of true price minus predicted price. 03:21.210 --> 03:31.450 Okay, the last MSE used for training and may used for monitoring human interpretation. 03:31.570 --> 03:37.770 As a quick recap, we need to pass three parameters in order to compile the model. 03:37.770 --> 03:41.970 For the compile function, we have the optimizer. 03:42.330 --> 03:55.130 The optimizer is simple and is used in simple problems that didn't need complex optimizers. 03:55.170 --> 03:57.810 Learning rate 0.01. 03:57.850 --> 03:59.970 Fast enough to learn it quickly. 04:00.010 --> 04:02.100 Slow enough to be stable. 04:02.140 --> 04:09.420 MSE loss standard for regression matches your squared error in data generation. 04:09.580 --> 04:12.660 M easy to interpret. 04:12.900 --> 04:14.540 Average color error. 04:14.580 --> 04:15.500 For example. 04:15.500 --> 04:18.540 In summary, compilation tells the model. 04:18.580 --> 04:20.980 Learn using gradient descent. 04:21.020 --> 04:27.020 Measure errors with MSE and show me as a progress report. 04:27.020 --> 04:32.860 I want from you to write all those notes down because those are very important. 04:32.860 --> 04:42.100 By the way, don't, uh worry, I'm gonna provide you with all of those notes, all of those notebooks 04:42.100 --> 04:44.580 with detailed explanation. 04:44.580 --> 04:54.820 So as a quick recap, compilation tells the model, learn using gradient descent, measure errors with 04:54.860 --> 04:59.780 MSE and show me Ma as a progress report.