1 00:00:11,680 --> 00:00:14,060 In this lecture we are going to build our model. 2 00:00:14,080 --> 00:00:17,250 The lost function in train our Siamese network. 3 00:00:17,530 --> 00:00:23,140 To start we're going to define our network which actually itself contains an entire neural network. 4 00:00:23,200 --> 00:00:28,260 As usual you can feel free to play around with these hyper parameters and choose your own architecture. 5 00:00:28,270 --> 00:00:33,160 This CNN is just what I happened to copy from the fashion amnesty example earlier in the course 6 00:00:37,690 --> 00:00:42,910 notice that instead of a final soft Max we end with a dense layer that has no activation. 7 00:00:42,910 --> 00:00:44,740 Since we are not doing classification 8 00:00:48,780 --> 00:00:51,210 in the forward function it's very simple. 9 00:00:51,330 --> 00:00:54,110 First as input we accept two batches of images. 10 00:00:54,120 --> 00:00:55,850 I am one and I am too. 11 00:00:56,070 --> 00:01:01,860 We pass both of these through the same CNN to get feet 1 and feet to the embedding for both sets of 12 00:01:01,860 --> 00:01:02,490 images. 13 00:01:03,210 --> 00:01:08,530 Finally we calculate the Euclidean distance between two feet one in feet two for each row. 14 00:01:08,700 --> 00:01:15,120 So if we add any input pairs will have an output distances. 15 00:01:15,250 --> 00:01:21,540 Next we create an instance of our model and set the feature dimension to be 50. 16 00:01:21,620 --> 00:01:23,420 Next we move the model to the GP you 17 00:01:27,460 --> 00:01:29,400 Next we create our custom lost function. 18 00:01:29,410 --> 00:01:35,410 The contrast of loss the first argument is the prediction of Y and the second argument is the target 19 00:01:35,410 --> 00:01:37,810 t inside this function. 20 00:01:37,810 --> 00:01:42,080 We start by calculating the squared distance for the non matching pairs. 21 00:01:42,220 --> 00:01:47,140 This is that funny calculation you saw in the theory lecture where we subtract y from the margin and 22 00:01:47,140 --> 00:01:47,630 so forth. 23 00:01:48,550 --> 00:01:51,890 So here you see we're using a margin value of 1. 24 00:01:51,940 --> 00:01:56,820 This means that for non matching pairs we want the distance to be at least 1. 25 00:01:56,980 --> 00:02:00,220 But as long as it's bigger than that we don't really care how big it is. 26 00:02:02,330 --> 00:02:08,780 Now from the formula you know that we want to take the max between the margin minus Y and 0 but you'll 27 00:02:08,780 --> 00:02:11,900 recognize that this is exactly the value function. 28 00:02:11,900 --> 00:02:16,290 So we can just use the built in real you in PI to urge. 29 00:02:16,400 --> 00:02:21,680 Next we pass this calculation into the rest of the loss which is relatively simple and take the mean 30 00:02:21,680 --> 00:02:22,550 over each sample 31 00:02:26,820 --> 00:02:28,860 next we create our optimizer. 32 00:02:28,860 --> 00:02:32,630 Usually I say loss and optimizer but that is not the case in the script. 33 00:02:32,730 --> 00:02:34,950 Since our loss isn't one of the built in losses 34 00:02:38,880 --> 00:02:40,320 next we have our training function. 35 00:02:40,920 --> 00:02:43,160 So what's different about this. 36 00:02:43,350 --> 00:02:48,810 First you'll notice that in addition to the train generator and the test generator we also pass in the 37 00:02:48,810 --> 00:02:51,720 values for the number of steps per epoch. 38 00:02:51,720 --> 00:02:56,940 This is essential when you use a generator because if you don't know when an epoch is finished you'll 39 00:02:56,940 --> 00:02:59,970 just end up in an infinite loop. 40 00:02:59,970 --> 00:03:03,620 Recall that our generators have infinite while loops inside of them. 41 00:03:03,630 --> 00:03:05,780 This carries over from tends to flow and carries. 42 00:03:05,790 --> 00:03:08,760 So if you've ever used these libraries it should seem familiar 43 00:03:14,080 --> 00:03:14,830 inside the loop. 44 00:03:14,830 --> 00:03:20,950 It's mostly the usual stuff we loop over the train generator and on each iteration we get x1 x2 and 45 00:03:20,950 --> 00:03:24,560 the targets since that's what is yielded by our generator. 46 00:03:24,670 --> 00:03:28,760 We move that to the GP U and so on and so forth. 47 00:03:28,780 --> 00:03:33,430 One minor difference is that I have this variable called steps that counts the number of steps I've 48 00:03:33,430 --> 00:03:34,660 completed so far. 49 00:03:37,300 --> 00:03:40,590 So at the end of our inner loop I incremental steps by 1. 50 00:03:40,840 --> 00:03:44,950 And if it's greater than or equal to the steps per epoch I break out of the inner loop 51 00:03:48,490 --> 00:03:52,510 so it's the same basic idea for the test generator except no gradient descent. 52 00:03:57,760 --> 00:04:00,420 Okay so next we call a training function. 53 00:04:00,850 --> 00:04:04,270 As you can see this train is quite fast. 54 00:04:04,270 --> 00:04:05,470 Pretty good for a CNN 55 00:04:09,250 --> 00:04:12,050 X. We plot the last pre iteration which looks OK 56 00:04:17,210 --> 00:04:23,220 and that's it for now in the next lecture we'll focus on how to evaluate our train to Siamese network.