1 00:00:11,700 --> 00:00:17,370 In this lecture we are going to begin looking at our facial recognition code with Siamese networks. 2 00:00:17,490 --> 00:00:22,170 As usual you can look at the title of the notebook to determine what notebook we are currently looking 3 00:00:22,170 --> 00:00:23,340 at. 4 00:00:23,430 --> 00:00:28,560 Since this is a large piece of code it's going to span over several lectures with each lecture focused 5 00:00:28,560 --> 00:00:29,840 on a single theme. 6 00:00:30,090 --> 00:00:33,420 The theme of this lecture will be loading in the data. 7 00:00:33,420 --> 00:00:38,330 As I mentioned previously the actual theory and model behind this project are pretty simple. 8 00:00:38,520 --> 00:00:42,510 And most of the work will go into loading in the data and processing it correctly 9 00:00:48,810 --> 00:00:52,680 so the first thing you'll notice in this script is the strange imports. 10 00:00:52,680 --> 00:00:54,030 This course is about pi talk. 11 00:00:54,060 --> 00:00:56,480 But I'm importing tens of flowing cars. 12 00:00:56,550 --> 00:00:58,200 Why am I doing that. 13 00:00:58,200 --> 00:01:02,750 Well first let's start by recognizing that these are both just Python libraries. 14 00:01:02,790 --> 00:01:06,400 There's no reason you can't use multiple Python libraries simultaneously. 15 00:01:06,450 --> 00:01:11,790 As we often do for example we use an empire map plot lived side by side. 16 00:01:11,790 --> 00:01:17,430 So just because pi torch and sensor flow overlap in what they do in the sense that they're both ways 17 00:01:17,430 --> 00:01:22,290 of building neural networks it's totally acceptable to use a different library that has some useful 18 00:01:22,290 --> 00:01:26,610 functionality even if the neuron that were parts of these libraries are incompatible 19 00:01:33,190 --> 00:01:35,200 so first let's download our dataset. 20 00:01:35,500 --> 00:01:37,640 We'll be using the Yale face data set. 21 00:01:37,690 --> 00:01:42,300 It's just a bunch of different faces that have different emotions like happy sad and so forth 22 00:01:46,160 --> 00:01:46,710 next. 23 00:01:46,730 --> 00:01:51,020 Since this is a zip file we're going to unzip it using the unzip command 24 00:01:56,800 --> 00:02:03,220 next we're going to get the file paths of all the images using the glob function you'll notice that 25 00:02:03,250 --> 00:02:09,700 while all the images are gifts they don't end in the file extension that G I f they have file names 26 00:02:09,700 --> 00:02:12,540 like subjects 0 1 one happy subject. 27 00:02:12,580 --> 00:02:14,330 Oh one not sad and so forth. 28 00:02:14,770 --> 00:02:18,280 So we are looking for any file that starts with the string subject. 29 00:02:21,100 --> 00:02:26,410 Next we are going to shuffle the file pass which will be important later when we create our train and 30 00:02:26,410 --> 00:02:30,740 test sets. 31 00:02:30,750 --> 00:02:37,920 Next we're going to set the number of files to a variable called n so we have one hundred sixty six 32 00:02:37,920 --> 00:02:43,530 images. 33 00:02:43,620 --> 00:02:50,230 Next we have a function that loads in the data from a file path and returns an umpire a there are multiple 34 00:02:50,230 --> 00:02:51,210 things going on here. 35 00:02:51,240 --> 00:02:53,330 So let's break it down. 36 00:02:53,390 --> 00:02:59,270 First we call image dot load AMG to load in the file as an image object. 37 00:02:59,270 --> 00:03:03,060 This also allows us to resize the image to 60 by 80. 38 00:03:03,080 --> 00:03:09,000 Which I found leaves a smaller memory footprint and doesn't prevent us from getting good results. 39 00:03:09,020 --> 00:03:16,190 The original image size was 243 by 320 which was unnecessarily large. 40 00:03:16,190 --> 00:03:21,560 Next we call I am g to array to convert the image object into an umpire. 41 00:03:22,610 --> 00:03:29,030 And finally we convert it to a U.N. eight which takes up only eight bits per pixel rather than 32 or 42 00:03:29,110 --> 00:03:30,830 64 for a float or double 43 00:03:34,360 --> 00:03:34,850 next. 44 00:03:34,870 --> 00:03:39,140 As I always like to do I'm going to plot a random image from the data. 45 00:03:39,340 --> 00:03:40,960 Just for fun. 46 00:03:40,960 --> 00:03:43,600 This always helps me get into the flow of things when I'm coding 47 00:03:52,230 --> 00:03:53,790 if we took the shape of our image. 48 00:03:53,820 --> 00:03:58,020 You might expect to see 60 by 80 since well that's what we specify. 49 00:03:58,020 --> 00:04:04,230 We want the size to be but something weird is happening here if you look closely the actual image shape 50 00:04:04,260 --> 00:04:06,530 is 60 by 80 by 3. 51 00:04:06,690 --> 00:04:12,020 That just seem weird to you because if we look at these images they are clearly grayscale images yet 52 00:04:12,050 --> 00:04:16,590 they are stored as color images with three color channels as an exercise. 53 00:04:16,590 --> 00:04:21,480 I would recommend verifying that each of these channels actually have the same values and so they are 54 00:04:21,480 --> 00:04:22,630 redundant. 55 00:04:22,860 --> 00:04:28,220 Therefore we lose nothing by converting these technically color images integrate scale. 56 00:04:28,560 --> 00:04:31,890 We can do that by taking the mean along the last axis 57 00:04:37,490 --> 00:04:42,890 next we're going to load in the images into a single num higher array so its end by whatever the image 58 00:04:42,890 --> 00:04:44,640 shape is 60 by 80. 59 00:04:45,960 --> 00:04:51,480 So all we do here is we live through each of the image file paths we call the lower AMG function which 60 00:04:51,480 --> 00:05:02,000 we just define and then we store the image in our big array of images. 61 00:05:02,010 --> 00:05:08,180 Next we're going to create an array of labels of length n these are not our binary labels. 62 00:05:08,260 --> 00:05:13,240 These are just labels telling us which subjects the corresponding images of. 63 00:05:13,270 --> 00:05:18,610 Basically this is just a bunch of manual string passing since I don't like to play with regex. 64 00:05:18,820 --> 00:05:24,250 So first we split out the file name by splitting on the last forward slash. 65 00:05:24,250 --> 00:05:28,720 Next we split out the subject's part of the file name by splitting on the first dot. 66 00:05:29,470 --> 00:05:34,980 So now we have a string that's like subject a one a subject or two subjects or three and so forth. 67 00:05:36,620 --> 00:05:42,470 Finally we just remove the subject strain so we're left with 0 1 0 2 or 3 and so on. 68 00:05:42,710 --> 00:05:47,390 And since the IDB starts from one I just subtract one so that it starts from zero.