1 00:00:00,320 --> 00:00:04,136 The last few you train horse versus humans, cats versus dogs, 2 00:00:04,136 --> 00:00:08,300 would bind the classification problems with just two output. 3 00:00:08,300 --> 00:00:12,380 For a lot of applications, you might want more than two output courses such as 4 00:00:12,380 --> 00:00:15,880 Rock Paper Scissors where you would have three possible courses. 5 00:00:15,880 --> 00:00:19,660 So in this week, you learn the details of how to implement this. 6 00:00:19,660 --> 00:00:23,425 And in fact, you've actually seen one version of this already in the first 7 00:00:23,425 --> 00:00:27,513 course where you did the ten course classification problem of fashion. 8 00:00:27,513 --> 00:00:30,943 But in this week, you'll learn the important coding difference of how to 9 00:00:30,943 --> 00:00:33,440 actually build these multi-class classifiers. 10 00:00:33,440 --> 00:00:34,310 >> Yes, exactly, 11 00:00:34,310 --> 00:00:38,060 and it's one of the things that's really exciting about the image data generator is 12 00:00:38,060 --> 00:00:42,480 how your images get automatically labeled based on the directories that they're in. 13 00:00:42,480 --> 00:00:45,650 So we'll take a look at having with rock, paper, scissors, for example, we'll 14 00:00:45,650 --> 00:00:49,230 have three sub-directories of rock, paper, and scissors with the images in them. 15 00:00:49,230 --> 00:00:51,930 And we'll learn how the image data generator then works in 16 00:00:51,930 --> 00:00:54,660 outside of binary mode and in multi-class mode. 17 00:00:54,660 --> 00:00:58,420 We'll also look at the coding changes that you would have within the DNN part 18 00:00:58,420 --> 00:01:00,130 of your convolutional neural network so 19 00:01:00,130 --> 00:01:02,920 that you can work more efficiently with multi-class. 20 00:01:02,920 --> 00:01:07,103 >> And one of the really cool things we'll see this week is the rock, 21 00:01:07,103 --> 00:01:11,756 paper, scissors data set that you generated using computer graphics. 22 00:01:11,756 --> 00:01:15,893 >> Yeah, so I wanted to explore that one of the bottlenecks 23 00:01:15,893 --> 00:01:20,125 I find in creating data sets is in getting access to data. 24 00:01:20,125 --> 00:01:22,619 And if I wanted to build a rock, paper, scissors, 25 00:01:22,619 --> 00:01:25,062 I wanted to have people of different races hands. 26 00:01:25,062 --> 00:01:28,022 So Asian hands, black hands, white hands. 27 00:01:28,022 --> 00:01:29,669 I wanted to have male and female hands. 28 00:01:29,669 --> 00:01:32,763 I wanted to have nail polish and pointy nails on some and 29 00:01:32,763 --> 00:01:35,258 I wanted to have big hands and little hands. 30 00:01:35,258 --> 00:01:39,493 And if I were to do that with real images, I'd have to hire a lot of models and 31 00:01:39,493 --> 00:01:41,554 I didn't have the budget for that. 32 00:01:41,554 --> 00:01:44,420 And so, I thought well, what if I do it with CGI? 33 00:01:44,420 --> 00:01:47,396 And if I generate CGI hands for all of these, and then to see. 34 00:01:47,396 --> 00:01:50,850 Because convolutions are really all about extracting features. 35 00:01:50,850 --> 00:01:52,040 What does a finger look like? 36 00:01:52,040 --> 00:01:53,730 What does a finger nail look like? 37 00:01:53,730 --> 00:01:54,420 That type of thing. 38 00:01:54,420 --> 00:01:57,850 And with photo real CGI, the convolutions that make for 39 00:01:57,850 --> 00:02:01,380 finger might look exactly the same that it would be of a photograph of a hand. 40 00:02:01,380 --> 00:02:03,350 So I wanted to explore that and do some training on that. 41 00:02:03,350 --> 00:02:05,123 And it works surprisingly well. 42 00:02:05,123 --> 00:02:09,468 >> Yeah, one of the exciting developments in deep learning is that computer graphics 43 00:02:09,468 --> 00:02:11,453 is advanced to the point where more and 44 00:02:11,453 --> 00:02:15,735 more groups are considering using computer graphics to generate images to feed to 45 00:02:15,735 --> 00:02:18,791 these incredibly data hungry deep learning algorithms. 46 00:02:18,791 --> 00:02:23,124 Several years ago, when computer graphics was less advanced, it was like 47 00:02:23,124 --> 00:02:27,406 a slightly fishy thing to do to use computer graphics to synthesize data. 48 00:02:27,406 --> 00:02:31,567 And even though today, using graphics, not a slam dunk, it does not always work, and 49 00:02:31,567 --> 00:02:33,452 there's still problems quite often. 50 00:02:33,452 --> 00:02:38,054 I thought it was really cool that you were able to generate a pretty decent data set, 51 00:02:38,054 --> 00:02:40,510 certainly good enough for a learning tool. 52 00:02:40,510 --> 00:02:44,747 And in other context, I see companies starting to explore these ideas for 53 00:02:44,747 --> 00:02:46,601 more and more settings as well. 54 00:02:46,601 --> 00:02:49,749 >> Yeah, I think as an industry, we're only getting started with it. 55 00:02:49,749 --> 00:02:53,529 And that's really, really good for us to experiment with these things. 56 00:02:53,529 --> 00:02:58,002 And hopefully, the students in this course being able to experiment using these CGI 57 00:02:58,002 --> 00:03:00,840 based data sets will be a nice introduction for them. 58 00:03:00,840 --> 00:03:05,064 >> Yeah, so more data to feed to this giant hungry maw of our data hungry deep 59 00:03:05,064 --> 00:03:06,472 learning algorithms. 60 00:03:06,472 --> 00:03:07,220 >> Exactly. 61 00:03:07,220 --> 00:03:11,220 >> So in this week, you get to play with this cool new data set and 62 00:03:11,220 --> 00:03:15,810 use it to practice implementing building multi-class classifiers in TensorFlow. 63 00:03:15,810 --> 00:03:16,310 Please dive in.