1 00:00:00,280 --> 00:00:03,780 Last week you looked at creating a synthetic seasonal data set that 2 00:00:03,780 --> 00:00:07,310 contained trend, seasonality, and a bit of noise. 3 00:00:07,310 --> 00:00:11,310 You also looked at some statistical methods for analyzing the data set and 4 00:00:11,310 --> 00:00:13,320 making predictions from it. 5 00:00:13,320 --> 00:00:15,140 Some of the results you got were actually quite good, 6 00:00:15,140 --> 00:00:18,020 but there was no machine learning applied yet. 7 00:00:18,020 --> 00:00:21,180 This week, you're going to look at using some machine learning methods with 8 00:00:21,180 --> 00:00:22,200 the same data. 9 00:00:22,200 --> 00:00:23,465 Let's see where machine learning can take us. 10 00:00:25,872 --> 00:00:28,278 First of all, as with any other ML problem, 11 00:00:28,278 --> 00:00:31,760 we have to divide our data into features and labels. 12 00:00:31,760 --> 00:00:35,757 In this case our feature is effectively a number of values in the series, 13 00:00:35,757 --> 00:00:37,580 with our label being the next value. 14 00:00:37,580 --> 00:00:41,545 We'll call that number of values that will treat as our feature, 15 00:00:41,545 --> 00:00:45,545 the window size, where we're taking a window of the data and 16 00:00:45,545 --> 00:00:48,645 training an ML model to predict the next value. 17 00:00:48,645 --> 00:00:53,085 So for example, if we take our time series data, say, 30 days at a time, 18 00:00:53,085 --> 00:00:57,630 we'll use 30 values as the feature and the next value is the label. 19 00:00:57,630 --> 00:00:59,000 Then over time, 20 00:00:59,000 --> 00:01:03,470 we'll train a neural network to match the 30 features to the single label. 21 00:01:05,000 --> 00:01:07,930 So let's, for example, use the tf.data.Dataset 22 00:01:07,930 --> 00:01:12,100 class to create some data for us, we'll make a range of 10 values. 23 00:01:13,320 --> 00:01:16,219 When we print them we'll see a series of data from 0 to 9. 24 00:01:17,580 --> 00:01:19,710 So now let's make it a little bit more interesting. 25 00:01:19,710 --> 00:01:24,150 We'll use the dataset.window to expand our data set using windowing. 26 00:01:25,220 --> 00:01:27,620 Its parameters are the size of the window and 27 00:01:27,620 --> 00:01:30,380 how much we want to shift by each time. 28 00:01:30,380 --> 00:01:35,175 So if we set a window size of 5 with a shift of 1 when we print it we'll 29 00:01:35,175 --> 00:01:39,637 see something like this, 01234, which just stops there 30 00:01:39,637 --> 00:01:44,283 because it's five values, then we see 12345 etc, etc,. 31 00:01:44,283 --> 00:01:49,122 Once we get towards the end of the data set we'll have less values 32 00:01:49,122 --> 00:01:51,800 because they just don't exist. 33 00:01:51,800 --> 00:01:54,390 So we'll get 6789, and then 789, etc, etc,. 34 00:01:56,130 --> 00:02:01,100 So let's edit our window a little bit, so that we have regularly sized data. 35 00:02:01,100 --> 00:02:04,125 We can do that with an additional parameter on the window called 36 00:02:04,125 --> 00:02:06,300 drop_remainder. 37 00:02:06,300 --> 00:02:07,790 And if we set this to true, 38 00:02:07,790 --> 00:02:11,230 it will truncate the data by dropping all of the remainders. 39 00:02:11,230 --> 00:02:15,110 Namely, this means it will only give us windows of five items. 40 00:02:15,110 --> 00:02:19,700 So when we print it, it will now look like this, 41 00:02:19,700 --> 00:02:23,340 starting at 01234 and ending at 56789. 42 00:02:23,340 --> 00:02:27,600 Great, now let's put these into numpy lists so 43 00:02:27,600 --> 00:02:30,130 that we can start using them with machine learning. 44 00:02:30,130 --> 00:02:35,150 Good news is, is that this is super easy, we just call the .numpy method on each 45 00:02:35,150 --> 00:02:40,260 item in the data set, and when we print we now see that we have a numpy list. 46 00:02:41,630 --> 00:02:45,830 Okay, next up is to split the data into features and labels. 47 00:02:45,830 --> 00:02:49,980 For each item in the list it kind of makes sense to have all of the values but 48 00:02:49,980 --> 00:02:54,980 the last one to be the feature, and then the last one can be the label. 49 00:02:54,980 --> 00:02:59,880 And this can be achieved with mapping, like this, where we split into everything 50 00:02:59,880 --> 00:03:06,120 but the last one with :-1, and then just the last one itself with -1:. 51 00:03:06,120 --> 00:03:08,860 Which gives us this output when we print, 52 00:03:08,860 --> 00:03:12,960 which now looks like a nice set of features and labels. 53 00:03:12,960 --> 00:03:15,760 Typically, you would shuffle their data before training. 54 00:03:15,760 --> 00:03:19,440 And this is possible using the shuffle method. 55 00:03:19,440 --> 00:03:21,670 We call it with the buffer size of ten, 56 00:03:21,670 --> 00:03:23,830 because that's the amount of data items that we have. 57 00:03:24,980 --> 00:03:26,300 And when we print the results, 58 00:03:26,300 --> 00:03:29,060 we'll see our features and label sets have been shuffled. 59 00:03:30,130 --> 00:03:34,360 Finally, we can look at batching the data, and this is done with the batch method. 60 00:03:35,830 --> 00:03:38,259 It'll take a size parameter, and in this case it's 2. 61 00:03:38,259 --> 00:03:42,369 So what we'll do is we'll batch the data into sets of two, and 62 00:03:42,369 --> 00:03:45,550 if we print them out, we'll see this. 63 00:03:45,550 --> 00:03:49,070 We now have three batches of two data items each. 64 00:03:49,070 --> 00:03:52,960 And if you look at the first set, you'll see the corresponding x and y. 65 00:03:52,960 --> 00:03:57,076 So when x is four, five, six and seven, our y is eight, or 66 00:03:57,076 --> 00:04:01,460 when x is zero, one, two, three, you'll see our y is four. 67 00:04:02,950 --> 00:04:07,930 Okay, now that you've seen the tools that let us create a series of x and y's, or 68 00:04:07,930 --> 00:04:10,995 features and labels, you have everything you need to work on 69 00:04:10,995 --> 00:04:13,980 a data set in order to get predictions from it. 70 00:04:13,980 --> 00:04:18,010 We'll take a look at a screen cast of this code next, before moving on to creating 71 00:04:18,010 --> 00:04:20,590 our first neural networks to run predictions on this data.