1 00:00:00,240 --> 00:00:01,230 Hi and welcome back. 2 00:00:01,470 --> 00:00:06,510 In this section, we'll take a look at using carrots to build a one or two encoder. 3 00:00:06,540 --> 00:00:09,330 So open Notebook 29 and we'll get started. 4 00:00:09,780 --> 00:00:15,780 So firstly, this lesson was taken from the official Keros blog tutorial on building also incluidas 5 00:00:15,780 --> 00:00:16,500 in Keros. 6 00:00:17,010 --> 00:00:22,320 And here's a nice little diagram to basically show what's happening and also include this here we have 7 00:00:22,320 --> 00:00:23,610 a bottleneck layer here. 8 00:00:24,060 --> 00:00:29,850 And basically what we're doing, we're just trying to encode the input in a smaller dimensionality similar 9 00:00:29,850 --> 00:00:35,640 to principal components analysis and generating an output that looks a lot like the input and try to 10 00:00:36,570 --> 00:00:37,260 try to encode it. 11 00:00:37,950 --> 00:00:40,560 So basically, let's get started with this lesson. 12 00:00:40,560 --> 00:00:44,670 So here are some definitions of what an encoder and decoder is. 13 00:00:45,030 --> 00:00:51,270 Remember an encoded compresses of dung samples, the input image and to a lesser number of bits or lower 14 00:00:51,270 --> 00:00:52,020 dimensionality. 15 00:00:52,560 --> 00:00:56,370 And this last symptom of this is called a living in space or bottleneck. 16 00:00:56,880 --> 00:01:02,310 And what the decoder does, the decoder takes this lesson number of bits or the smaller dimensional 17 00:01:02,310 --> 00:01:09,120 representation of the input and then tries to reconstruct the input using only the including of that 18 00:01:09,120 --> 00:01:09,630 input. 19 00:01:09,840 --> 00:01:13,800 So that's basically a summary of what an auto encode is right here. 20 00:01:14,190 --> 00:01:18,420 So now let's get started with building a very simple, basic auto encoder. 21 00:01:18,720 --> 00:01:22,140 So firstly, let's import some packages that we'll be using. 22 00:01:22,140 --> 00:01:29,040 Those are the Keros TensorFlow packages, as well as MATLAB, and NumPy and knowThis will be using the 23 00:01:29,040 --> 00:01:30,120 amnesty dataset. 24 00:01:30,420 --> 00:01:36,210 So we'll be including we'll be treating auto encoder on the M this handwritten digit dataset. 25 00:01:36,660 --> 00:01:42,000 So the helper functions we're creating here is firstly a pre-processed function that normalizes the 26 00:01:42,000 --> 00:01:47,700 supplied array and then reshapes it into the appropriate format so that we can pass the input to our 27 00:01:47,730 --> 00:01:48,570 neural network. 28 00:01:49,170 --> 00:01:52,410 Then we have a function that adds random noise to an image. 29 00:01:52,950 --> 00:01:57,420 And then lastly, by the way, we'll be using this later and later in this section. 30 00:01:57,420 --> 00:01:58,350 So don't worry about it. 31 00:01:58,800 --> 00:02:00,950 If you're wondering why we're adding noise, don't worry. 32 00:02:00,960 --> 00:02:01,830 I'll explain later. 33 00:02:02,520 --> 00:02:09,570 And then we have a display function that takes these two areas and displays 10 random images from each 34 00:02:09,570 --> 00:02:10,950 one of the supplied arrays. 35 00:02:11,190 --> 00:02:11,940 So that's it. 36 00:02:11,940 --> 00:02:18,810 So let's run this and we'll be importing our functions and declaring our helper functions. 37 00:02:20,400 --> 00:02:21,100 So there we go. 38 00:02:21,120 --> 00:02:22,440 So it's going to be run now. 39 00:02:22,860 --> 00:02:26,340 Next, we have to load and pre process our endless data set. 40 00:02:26,700 --> 00:02:32,370 So we just use the amnestied load function to get only to set up the training data and test it to notice. 41 00:02:32,370 --> 00:02:39,740 We're not using the limo's here because we just assign into this dummy variable, but an auto encoder. 42 00:02:39,750 --> 00:02:41,130 We don't need the variables at all. 43 00:02:41,430 --> 00:02:45,450 We just need to learn the representation, the underlying representation of the data. 44 00:02:45,780 --> 00:02:47,640 That's why we just look at the data itself. 45 00:02:48,240 --> 00:02:53,460 So then we normalize and reshape the data using the preprocessor pre-processed helper function we created 46 00:02:53,460 --> 00:02:53,850 above. 47 00:02:54,570 --> 00:03:00,480 And then we add some noise to the data because what we'll be doing here will be training an auto encoder 48 00:03:00,870 --> 00:03:03,840 that basically can remove noise from images. 49 00:03:04,320 --> 00:03:06,210 So let's take a look and see how we do that. 50 00:03:06,210 --> 00:03:12,600 So we just use a noise helper function to add random noise to the training data and test data, and 51 00:03:12,600 --> 00:03:13,830 then we can display it here. 52 00:03:13,860 --> 00:03:16,400 Let's display those sample images here. 53 00:03:16,420 --> 00:03:18,100 So it's downloaded data set first. 54 00:03:18,120 --> 00:03:19,260 That's what it's doing here. 55 00:03:20,370 --> 00:03:21,210 And there we go. 56 00:03:21,480 --> 00:03:25,620 So we have the the clean images on top and then the noisy ones below. 57 00:03:27,480 --> 00:03:30,450 So now we're ready to build or to include a model. 58 00:03:30,450 --> 00:03:35,450 So let's take a look at firstly, the encoder model you can see it's quite simple. 59 00:03:35,460 --> 00:03:39,510 We just have to convolutional is and each one followed by a maximal. 60 00:03:39,870 --> 00:03:45,060 So what you can imagine as we build this model, as the input goes through the network, it gets smaller. 61 00:03:45,370 --> 00:03:47,640 In fact, yes, it's getting a lot smaller right now. 62 00:03:48,120 --> 00:03:51,720 And then that's what we used to convert 2D transport layers. 63 00:03:52,080 --> 00:03:55,740 These are basically upsampling layers, but they're smart upsampling layers. 64 00:03:55,740 --> 00:04:00,930 And because they learn the feature representation of the image of this, their previous input here. 65 00:04:01,890 --> 00:04:09,870 So we have the two clung to the upscaling layers to produce basically back the output right here, and 66 00:04:09,870 --> 00:04:10,610 that's what we get. 67 00:04:10,620 --> 00:04:17,610 So that's how we can create or to encode a model, which is basically the input here and X, which basically 68 00:04:17,610 --> 00:04:21,600 strings along all of these sequential operations here. 69 00:04:22,140 --> 00:04:28,380 And that gives us the final or to encode a model where we use Adam as the optimizer and binary cross 70 00:04:28,380 --> 00:04:30,000 entropy as the lost function. 71 00:04:30,420 --> 00:04:36,510 So let's run this and we can take a look at the architecture here so you can see how it gets bottlenecked 72 00:04:36,720 --> 00:04:40,290 as we go down here and then suddenly it gets back bigger. 73 00:04:40,680 --> 00:04:44,430 That's because of the continuity transport layers here. 74 00:04:45,150 --> 00:04:46,110 So that's it. 75 00:04:46,620 --> 00:04:48,710 And then we are now ready to trino quarter on quarter. 76 00:04:49,080 --> 00:04:51,660 So notice labels? 77 00:04:51,870 --> 00:04:52,800 No, the training data. 78 00:04:53,160 --> 00:04:59,900 That's because the output of this network is a 28 by 28 dimension effective and. 79 00:04:59,970 --> 00:05:04,790 No, previously it used to be at one by 10, and those were the outputs for probabilities of which classes. 80 00:05:05,150 --> 00:05:07,040 That's when we were building classifiers. 81 00:05:07,430 --> 00:05:11,690 Now we're building a neural net to learn the image and output the image back again. 82 00:05:11,900 --> 00:05:17,210 So and that's why we have separate kind of separate encoder and decoder models. 83 00:05:18,260 --> 00:05:19,870 So we're not ready to try this. 84 00:05:19,880 --> 00:05:25,430 And also notice to test data is what we'll be evaluating one after to see how well it performs. 85 00:05:26,000 --> 00:05:30,020 So let's train this and let's see how long it takes to train. 86 00:05:30,260 --> 00:05:35,000 I actually don't remember we're training for 50 epochs with a bad set of 128. 87 00:05:36,500 --> 00:05:40,220 OK, it's pretty quick, but we'll have to wait a bit still. 88 00:05:40,400 --> 00:05:41,420 So let's wait it out. 89 00:05:51,190 --> 00:05:57,610 OK, so we're done training now, and as you can see, we trained 50 bucks and examining the validation 90 00:05:57,610 --> 00:05:58,300 laws here. 91 00:05:58,840 --> 00:06:01,770 You can see 50 folks was probably a good choice. 92 00:06:01,780 --> 00:06:02,740 It seemed to have. 93 00:06:02,950 --> 00:06:06,700 It seemed to actually keep decreasing here, so we could have actually trained us for a bit longer. 94 00:06:07,180 --> 00:06:10,630 But nevertheless, this is a fairly low validation loss here. 95 00:06:11,080 --> 00:06:13,780 So let's take a look at our predictions now. 96 00:06:14,260 --> 00:06:19,960 So to get to predictions here, what we're doing, we're going to get all of the predictions out by 97 00:06:19,960 --> 00:06:23,590 just feeding into test data here into the quarter to include a model. 98 00:06:23,950 --> 00:06:29,950 Don't predict and we get all the predictions here and then we just display some random turnaround and 99 00:06:29,950 --> 00:06:30,600 predictions here. 100 00:06:30,610 --> 00:06:35,170 So this is this display as the input data here and this display as the output. 101 00:06:35,560 --> 00:06:42,940 So remember what we're doing here, but the output is it's going to be the input that was reduced in 102 00:06:42,940 --> 00:06:43,720 dimensionality. 103 00:06:44,260 --> 00:06:50,260 And then we're going to we going to blow that back up to the full representation of the full decoding 104 00:06:50,260 --> 00:06:52,840 model so we can compare how well it did. 105 00:06:52,990 --> 00:06:59,950 So let's take a look at that and you can see this is actually quite good, actually really good. 106 00:07:00,190 --> 00:07:01,360 You can see the snow. 107 00:07:01,810 --> 00:07:05,170 Remember, let's take a look at the auto included here. 108 00:07:05,860 --> 00:07:12,310 We basically shrunk this data down all the way to a seven by seven by 282 representation at a time. 109 00:07:12,640 --> 00:07:20,410 And now we can use that input to blow back using the decoder, and you can see it gives us almost the 110 00:07:20,410 --> 00:07:21,300 exact image. 111 00:07:21,370 --> 00:07:26,970 Looking at it carefully, it seems like a disk encoded the exact image itself, so that is actually 112 00:07:26,980 --> 00:07:27,700 pretty -- good. 113 00:07:28,510 --> 00:07:30,050 So now let's take a look at this. 114 00:07:30,070 --> 00:07:35,740 Remember, we also created a noisy version of our dataset and we said we're going to use our auto encoder 115 00:07:35,750 --> 00:07:36,460 to clean that up. 116 00:07:36,850 --> 00:07:38,320 Well, let's take a look at how we do that. 117 00:07:38,320 --> 00:07:41,770 So now we have the noisy data here as input. 118 00:07:42,040 --> 00:07:46,930 However, notice what's different here, we don't use a noisy version of the training data here. 119 00:07:47,260 --> 00:07:53,500 We used a clean vision because what we want to auto included to do is to clean that image data set up. 120 00:07:53,890 --> 00:07:58,930 So we are taking noisy inputs and we're going to feed it into the network and then we're going to learn 121 00:07:59,050 --> 00:08:01,510 how to decode a clean vision out of it. 122 00:08:02,080 --> 00:08:03,130 So let's run that. 123 00:08:03,130 --> 00:08:05,650 No, let's put this for 50 bucks here. 124 00:08:05,650 --> 00:08:09,760 So it won't take that long, and let's wait for our results. 125 00:08:19,190 --> 00:08:23,210 OK, so we're done training, that's about three minutes of training this network. 126 00:08:23,630 --> 00:08:30,110 And you can firstly notice that a loss function here isn't as low as a previous one, and that's understandable 127 00:08:30,110 --> 00:08:33,320 because given how noisy our input images were. 128 00:08:33,890 --> 00:08:38,990 So let's go back to this again, and you can see we're training with noisy test data. 129 00:08:39,080 --> 00:08:40,880 So this is a validation data set here. 130 00:08:41,240 --> 00:08:44,210 And this is a target data here says, remember that. 131 00:08:44,930 --> 00:08:49,800 So now let's evaluate its performance on some noisy data. 132 00:08:49,820 --> 00:08:57,750 So let's see how well our auto encoder can clean up images and it's remarkably well, it gives it. 133 00:08:57,770 --> 00:09:03,140 We can take these inputs that these noisy inputs and clean them up and get these almost perfect images 134 00:09:03,140 --> 00:09:03,440 here. 135 00:09:03,860 --> 00:09:09,140 The only one that doesn't look too great is perhaps maybe this tree, because it's a bit of a streak 136 00:09:09,140 --> 00:09:13,910 right here and the five has some noise added to it. 137 00:09:14,720 --> 00:09:17,450 Maybe this eight as well, but I mean, this is quite good. 138 00:09:17,930 --> 00:09:21,500 So I hope you enjoy this lesson and know what we're going to do. 139 00:09:21,860 --> 00:09:25,730 We're going to build an auto encoder in PyTorch, so stay tuned for that. 140 00:09:25,880 --> 00:09:26,270 Thank you.