WEBVTT 00:00.120 --> 00:03.040 We created the model architecture. 00:03.200 --> 00:07.640 The input layer, the hidden layers and the output layer. 00:07.840 --> 00:11.560 Now let's train our model here. 00:11.880 --> 00:18.480 Create a new cell and let's set up early stopping to prevent overfitting. 00:18.760 --> 00:24.920 In the previous videos we learned how to use Model.fit function to train the model. 00:25.000 --> 00:32.000 But in this example we're going to use a new thing called early stopping. 00:32.000 --> 00:34.400 Also we're going to use Model.fit function. 00:34.400 --> 00:39.480 But I'm going to use something called early stopping. 00:39.640 --> 00:48.640 Early stopping equals to Tf.keras dot callback dot early stopping. 00:48.800 --> 00:52.440 Pass the monitor equals to val loss. 00:52.520 --> 00:58.160 Patients 20 and restore best weights equals to true. 00:58.280 --> 01:00.600 What does this code do? 01:00.640 --> 01:09.470 Early stopping automatically halts training when the model stops, improving on validation data. 01:09.630 --> 01:13.070 Preventing overfitting to the training set. 01:13.270 --> 01:25.070 So if you have large data and your model, for example, get trained with the first 200 epochs and you 01:25.070 --> 01:28.510 specified, for example, 500 epochs. 01:28.790 --> 01:36.310 Why you should make unnecessary 300 epochs and get the same results. 01:36.470 --> 01:42.910 So here the early stopping to prevent overfitting to the training set. 01:42.950 --> 01:51.270 When the model stops improving on validation data, it's very important we're going to depend on the 01:51.310 --> 01:53.230 Mae and MSE. 01:53.630 --> 01:57.870 We're going to see the results here in this video. 01:57.870 --> 02:08.670 And you should see how early stopping is important to prevent memory leakage, overfitting, um, resource 02:08.710 --> 02:10.350 heating and memory. 02:10.510 --> 02:16.270 Okay, so there are a lot of things we can prevent using the early stopping. 02:16.470 --> 02:16.990 Don't worry. 02:17.030 --> 02:23.910 We're going to show you, uh, to see all those in the next couple of minutes. 02:24.150 --> 02:24.990 Monitor. 02:25.390 --> 02:27.670 Tell us what to watch. 02:27.910 --> 02:31.510 Tracks the validation loss during the training. 02:31.510 --> 02:33.430 So go and see the value. 02:33.710 --> 02:34.790 Val loss. 02:34.910 --> 02:37.390 You can also use val accuracy. 02:37.430 --> 02:38.710 Loss accuracy. 02:38.710 --> 02:42.150 So you can you can specify accuracy. 02:42.190 --> 02:44.310 You can specify val accuracy. 02:44.310 --> 02:51.910 But for this example we're going to focus on val loss val loss is preferred for regression tasks. 02:51.950 --> 02:54.070 Patience equals to 20. 02:54.110 --> 02:55.110 How long to wait. 02:55.150 --> 02:58.990 Number of epochs with no improvement before stopping. 02:59.230 --> 03:04.780 And if the validation loss doesn't improve for 20 consecutive epochs. 03:04.820 --> 03:14.620 Training stops balanced to best weights and crucial feature reverts model weights to the apex with the 03:14.620 --> 03:15.660 best performance. 03:15.660 --> 03:22.460 Without this, you'd get weights from the final, possibly overfitted epoch. 03:22.700 --> 03:26.500 Ensures you get the optimal model from training. 03:26.620 --> 03:32.420 Now let's start training the model so history equals to model dot fit. 03:32.460 --> 03:38.940 We use this fit function to train the model we used in the previous videos and previous sections. 03:38.980 --> 03:42.620 Passing the exit train scaled. 03:42.620 --> 03:44.580 This is very important guys. 03:44.580 --> 03:52.660 So here we need to pass the scaled because we talked about feature scaling. 03:52.820 --> 04:00.860 And here x train scaled parameter is passed passed Y train apex 1000. 04:00.860 --> 04:03.860 You can make it 200 400, 300. 04:03.900 --> 04:07.660 I'll start 200 and see the parameters. 04:07.660 --> 04:10.020 If there is a stop early stop. 04:10.260 --> 04:14.060 So there is no need to use 1000. 04:14.100 --> 04:18.020 Okay, so let me start with 200 and see the results. 04:18.060 --> 04:20.220 Validation split 0.2. 04:20.260 --> 04:23.980 Batch size 32 callbacks early stopping. 04:23.980 --> 04:29.660 So I'm gonna use this early stopping with the defined parameters here. 04:29.820 --> 04:32.860 And set the verbose equals to one. 04:32.900 --> 04:37.060 Let's click run and start training our model. 04:37.060 --> 04:40.300 Don't worry we're going to analyze all the results. 04:40.300 --> 04:42.100 And there we go guys. 04:42.100 --> 04:43.700 We trained our model. 04:44.020 --> 04:48.820 But if you notice we didn't complete the 200 epochs. 04:49.100 --> 04:56.980 We stopped at 169 epoch okay so let me study those parameters. 04:57.020 --> 05:04.050 Loss and Ma are very low values is 7.15. 05:04.610 --> 05:09.450 Val e 1.9 and then Ms. 7.1. 05:09.690 --> 05:15.090 If we go up for the first epoch, scroll up. 05:15.250 --> 05:18.730 We see that those values were high. 05:19.410 --> 05:26.290 So as we go down, those values becomes very small become very small. 05:26.330 --> 05:26.850 Okay. 05:27.210 --> 05:37.770 So as we see they are decreasing until we get the same results and there's no improvement if we continue 05:37.930 --> 05:38.690 the epochs. 05:38.810 --> 05:43.530 So for this purpose we used the early stop. 05:43.570 --> 05:44.090 Okay. 05:44.450 --> 05:49.530 So go and train the model whenever you see there is no improvement. 05:49.810 --> 05:52.050 Stop this training. 05:52.090 --> 05:52.650 Okay. 05:52.690 --> 05:58.970 This is how we train the model and how we use the early stopping feature.