1 00:00:00,000 --> 00:00:02,310 Okay. So now we will look at 2 00:00:02,310 --> 00:00:04,860 the code for the neural network definition. 3 00:00:04,860 --> 00:00:06,360 Remember last time we had a 4 00:00:06,360 --> 00:00:08,390 sequential with just one layer in it. 5 00:00:08,390 --> 00:00:10,575 Now we have three layers. 6 00:00:10,575 --> 00:00:12,750 The important things to look at are 7 00:00:12,750 --> 00:00:14,775 the first and the last layers. 8 00:00:14,775 --> 00:00:17,355 The last layer has 10 neurons in it 9 00:00:17,355 --> 00:00:20,580 because we have ten classes of clothing in the dataset. 10 00:00:20,580 --> 00:00:22,245 They should always match. 11 00:00:22,245 --> 00:00:24,930 The first layer is a flatten layer 12 00:00:24,930 --> 00:00:27,465 with the input shaping 28 by 28. 13 00:00:27,465 --> 00:00:30,855 Now, if you remember our images are 28 by 28, 14 00:00:30,855 --> 00:00:33,290 so we're specifying that this is 15 00:00:33,290 --> 00:00:36,110 the shape that we should expect the data to be in. 16 00:00:36,110 --> 00:00:38,390 Flatten takes this 28 by 28 17 00:00:38,390 --> 00:00:41,585 square and turns it into a simple linear array. 18 00:00:41,585 --> 00:00:44,390 The interesting stuff happens in the middle layer, 19 00:00:44,390 --> 00:00:46,445 sometimes also called a hidden layer. 20 00:00:46,445 --> 00:00:48,750 This is a 128 neurons in it, 21 00:00:48,750 --> 00:00:50,480 and I'd like you to think about 22 00:00:50,480 --> 00:00:52,475 these as variables in a function. 23 00:00:52,475 --> 00:00:55,735 Maybe call them x1, x2 x3, etc. 24 00:00:55,735 --> 00:00:58,010 Now, there exists a rule that 25 00:00:58,010 --> 00:01:00,170 incorporates all of these that 26 00:01:00,170 --> 00:01:02,240 turns the 784 values of 27 00:01:02,240 --> 00:01:05,060 an ankle boot into the value nine, 28 00:01:05,060 --> 00:01:07,875 and similar for all of the other 70,000. 29 00:01:07,875 --> 00:01:10,100 It's too complex a function for you to 30 00:01:10,100 --> 00:01:12,379 see by mapping the images yourself, 31 00:01:12,379 --> 00:01:14,425 but that's what a neural net does. 32 00:01:14,425 --> 00:01:17,030 So, for example, if you then say the function 33 00:01:17,030 --> 00:01:20,000 was y equals w1 times x1, 34 00:01:20,000 --> 00:01:21,710 plus w2 times x2, 35 00:01:21,710 --> 00:01:23,465 plus w3 times x3, 36 00:01:23,465 --> 00:01:27,095 all the way up to a w128 times x128. 37 00:01:27,095 --> 00:01:29,555 By figuring out the values of w, 38 00:01:29,555 --> 00:01:30,785 then y will be nine, 39 00:01:30,785 --> 00:01:32,725 when you have the input value of the shoe. 40 00:01:32,725 --> 00:01:34,910 You'll see that it's doing something very, 41 00:01:34,910 --> 00:01:37,010 very similar to what we did earlier 42 00:01:37,010 --> 00:01:39,560 when we figured out y equals 2x minus one. 43 00:01:39,560 --> 00:01:41,255 In that case the two, 44 00:01:41,255 --> 00:01:42,605 was the weight of x. 45 00:01:42,605 --> 00:01:46,150 So, I'm saying y equals w1 times x1, etc. 46 00:01:46,150 --> 00:01:49,370 Now, don't worry if this isn't very clear right now. 47 00:01:49,370 --> 00:01:51,320 Over time, you will get the hang of it, 48 00:01:51,320 --> 00:01:52,745 seeing that it works, 49 00:01:52,745 --> 00:01:54,710 and there's also some tools that will allow you 50 00:01:54,710 --> 00:01:57,025 to peek inside to see what's going on. 51 00:01:57,025 --> 00:01:59,810 The important thing for now is to get the code working, 52 00:01:59,810 --> 00:02:02,765 so you can see a classification scenario for yourself. 53 00:02:02,765 --> 00:02:05,480 You can also tune the neural network by adding, 54 00:02:05,480 --> 00:02:08,405 removing and changing layer size to see the impact. 55 00:02:08,405 --> 00:02:10,705 You'll do that in the next exercise. 56 00:02:10,705 --> 00:02:13,300 Also, if you want to go further, 57 00:02:13,300 --> 00:02:16,115 checkout this tutorial from Andrew on YouTube, 58 00:02:16,115 --> 00:02:19,235 which will clarify how dense connected layer's work 59 00:02:19,235 --> 00:02:22,100 from the theoretical and mathematical perspective. 60 00:02:22,100 --> 00:02:25,350 But for now, let's get back to the code.