1 00:00:00,390 --> 00:00:01,230 And welcome back. 2 00:00:01,830 --> 00:00:06,900 What we're going to do now is we're going to actually use transfer learning by feature extraction, 3 00:00:07,320 --> 00:00:13,800 which means that all we do is we freeze all the convolutional layers, all the first block of layers 4 00:00:13,800 --> 00:00:17,430 and the network that we loaded and then just trained the top players. 5 00:00:17,520 --> 00:00:19,830 So let's take a look and see how we do that. 6 00:00:20,610 --> 00:00:28,080 So firstly, we load our net model here that we've done previously, and we set pre-trained equals true, 7 00:00:28,080 --> 00:00:30,660 which means that we're loading the image net with here. 8 00:00:31,470 --> 00:00:35,220 Then what we do, this is the part where we freeze the layers here. 9 00:00:35,760 --> 00:00:42,720 So we set for the parameter of whatever we call this variable in the model convolutional parameters, 10 00:00:43,230 --> 00:00:47,070 we can set the required ingredients here to be false. 11 00:00:47,730 --> 00:00:50,610 This is this was what how we freeze layers in part to it. 12 00:00:50,610 --> 00:00:55,410 So it's basically all the convolutional layers here in the network. 13 00:00:56,100 --> 00:00:59,220 All the trouble is we're setting it to. 14 00:00:59,850 --> 00:01:01,620 We're freezing them here effectively. 15 00:01:02,550 --> 00:01:04,030 Next, what do we do? 16 00:01:04,050 --> 00:01:06,420 We're going to create a fully connected layer here. 17 00:01:07,050 --> 00:01:11,850 So we just take the model uploaded the convolutional layers of the resonant. 18 00:01:12,540 --> 00:01:19,050 And we just take the fully connected features here, the last layer features, and then we add on to 19 00:01:19,050 --> 00:01:23,010 our output output class here, basically. 20 00:01:23,670 --> 00:01:28,020 So what this does, this is going to keep this is this these are the Trinamool is here. 21 00:01:28,740 --> 00:01:35,100 So these are newly constructed modules and by default requires graphical true meaning that they're not 22 00:01:35,100 --> 00:01:37,590 going to be frozen, they're going to be trainable at that point. 23 00:01:38,340 --> 00:01:42,810 So we just passed this model to a GPUs set across them to be lost criterion. 24 00:01:43,860 --> 00:01:49,840 And then we just use again or living with Cedillo along with us, took us to Gradient Descent Optimizer 25 00:01:49,880 --> 00:01:51,810 and put these settings here for good. 26 00:01:52,500 --> 00:01:54,770 And then let's just create that. 27 00:01:54,780 --> 00:02:00,060 Actually, this wouldn't run right now because I am still training this model. 28 00:02:00,720 --> 00:02:01,500 It's almost done. 29 00:02:01,560 --> 00:02:07,020 Hopefully, by the end of this lesson, we'll have the results and then what you can do, you can just 30 00:02:07,350 --> 00:02:09,240 finally just train that model. 31 00:02:09,240 --> 00:02:14,910 Here you pass the mode that optimizes learning, which had you lose the model criterion. 32 00:02:15,480 --> 00:02:16,460 And that's it. 33 00:02:16,470 --> 00:02:18,270 And no one can take a look at her results. 34 00:02:18,270 --> 00:02:20,740 And you can see again, it's a bit spiky. 35 00:02:20,760 --> 00:02:24,150 We put up a ninety four point seven percent accuracy. 36 00:02:24,810 --> 00:02:27,030 So that means that we're going to get that. 37 00:02:27,030 --> 00:02:32,610 This model that's been returned here will be the best model unless there's some better results in this 38 00:02:32,610 --> 00:02:34,080 25, and I don't see it. 39 00:02:34,680 --> 00:02:39,990 So that's going to be the best model, and I don't have individualized predictions for those results 40 00:02:39,990 --> 00:02:44,760 here generated so you can experiment and visualized out on your own. 41 00:02:45,540 --> 00:02:46,020 All right. 42 00:02:46,020 --> 00:02:47,640 So that's it for this lesson. 43 00:02:48,030 --> 00:02:53,810 This is quite simple and actually do train most of my image classifiers in this method here by either 44 00:02:53,850 --> 00:02:59,970 combination of transfer learning, which feature extraction or fine tuning by freezing different layers. 45 00:03:00,330 --> 00:03:06,840 It's very good to experiment with freezing different layers and PyTorch and seeing how it affects the 46 00:03:06,840 --> 00:03:08,340 results on your dataset. 47 00:03:09,180 --> 00:03:12,750 So I hope you enjoyed it and we'll stop there for now and then. 48 00:03:12,750 --> 00:03:19,350 Next, we'll continue with using a pre trained model as a as a feature extractor like we just did. 49 00:03:19,890 --> 00:03:26,070 However, instead of having a fully connected neural network on top, we're going to use a linear network 50 00:03:26,070 --> 00:03:32,760 like a logistic regression to use the inputs from our previous layer as inputs to that model. 51 00:03:33,240 --> 00:03:35,700 And then and then create a model out of that. 52 00:03:36,150 --> 00:03:37,740 So I'll see you in the next lesson. 53 00:03:37,920 --> 00:03:38,370 Thank you.