1 00:00:11,120 --> 00:00:16,730 So in this video, we will continue looking at the notebook for human activity recognition, we just 2 00:00:16,730 --> 00:00:22,160 learned about how the data is organized so we have a decent enough understanding in order to pass it 3 00:00:22,160 --> 00:00:23,510 into the format we need. 4 00:00:24,590 --> 00:00:26,780 So let's start by defining a few constants. 5 00:00:27,350 --> 00:00:30,470 Note that you can get these values by looking at the data files. 6 00:00:31,940 --> 00:00:34,790 So we'll start with big tea, which we know is one twenty eight. 7 00:00:35,750 --> 00:00:38,660 Next, we have entraining and test, which are as follows. 8 00:00:39,110 --> 00:00:40,940 Basically you can find these by counting. 9 00:00:40,940 --> 00:00:47,930 The number of rows in any of the data files is equal to nine since we have nine parallel time series 10 00:00:48,170 --> 00:00:51,290 and K is equal to six since we have six classes. 11 00:00:56,240 --> 00:00:59,810 OK, so the next step is to define a function and call the load data. 12 00:01:00,440 --> 00:01:05,490 Again, it helps to think about what we need so that the code we write will go towards that goal. 13 00:01:06,200 --> 00:01:10,630 And of course, this just goes back to what we learned in the section about Time series basics. 14 00:01:11,120 --> 00:01:16,110 We'd like to have an X array of size and by TBD Biddy's and we'd like to have a wide array of length 15 00:01:16,110 --> 00:01:16,500 then. 16 00:01:17,420 --> 00:01:19,310 And this goes for both training and test. 17 00:01:22,440 --> 00:01:28,290 OK, so we'll start by reallocating our eggs train a next test to raise, both of which are invited. 18 00:01:31,100 --> 00:01:36,170 The next step is to instantiate a list containing the file names of each file, which contain our Time 19 00:01:36,170 --> 00:01:36,710 series. 20 00:01:37,100 --> 00:01:38,480 Note that this is for train. 21 00:01:41,670 --> 00:01:46,770 The next step is due to find a far path prefix, which is just the folder where these files are located. 22 00:01:50,070 --> 00:01:54,640 The next step is to live through each of the file names, since we use the enumerate function. 23 00:01:54,690 --> 00:01:56,340 This gives us the index as well. 24 00:01:57,180 --> 00:02:02,280 OK, so inside the loop, we begin by using pandas to load the file into the data frame. 25 00:02:02,730 --> 00:02:04,410 You should recall this code from earlier. 26 00:02:07,880 --> 00:02:14,360 The next step is to immediately convert the data frame into an empire which we call Little X. Now, 27 00:02:14,360 --> 00:02:16,720 again, it's always helpful to think about shapes. 28 00:02:17,030 --> 00:02:18,800 So what is the shape of little X? 29 00:02:19,280 --> 00:02:25,880 Well, it has nine rows and rows a time series of length, Big T, therefore, little X has the shape. 30 00:02:25,880 --> 00:02:32,330 And by T this makes sense since we have a D of these files and if we join them all together, we get 31 00:02:32,330 --> 00:02:33,330 invited. 32 00:02:34,640 --> 00:02:41,420 OK, so this means we have to index X train in the third dimension with the index little D, we assign 33 00:02:41,420 --> 00:02:46,640 this portion of X train to be a little X, OK, so hopefully this all makes sense. 34 00:02:47,120 --> 00:02:51,800 By the time this loop is over you will have all nine time series stored inside X train. 35 00:02:54,920 --> 00:03:00,890 The next step is to load in our labels again, we can use pennies for convenience and immediately convert 36 00:03:00,890 --> 00:03:02,020 it into an empire. 37 00:03:03,170 --> 00:03:08,090 Since the labels start from one, we are going to subtract one so that they all start from zero. 38 00:03:08,840 --> 00:03:14,120 We also want to flatten the labels, since using a data frame will give us to the objects of shape and 39 00:03:14,120 --> 00:03:16,180 buy one for this code. 40 00:03:16,190 --> 00:03:18,440 It's better to just have a onesie array of length n. 41 00:03:21,470 --> 00:03:27,260 OK, so the next step is to do the exact same steps, but for the test set, since this code is pretty 42 00:03:27,260 --> 00:03:29,600 much all the same, there is nothing to explain. 43 00:03:30,410 --> 00:03:34,130 Basically anywhere you previously saw train, you now see test. 44 00:03:39,630 --> 00:03:44,880 OK, so the next step is to use our function, which gives us X and Y for both training and test. 45 00:03:51,430 --> 00:03:57,370 Now, just as a sanity check, let's plot the multivariate time series of a single sample of chosen 46 00:03:57,370 --> 00:04:00,080 sample seven thousand for no particular reason. 47 00:04:00,880 --> 00:04:06,490 OK, so let's think about what the shape of this data will be since we index to an array of size and 48 00:04:06,490 --> 00:04:10,830 by TBD at some row, the result will be of shape TBD. 49 00:04:11,470 --> 00:04:16,720 So this is a single multivariate time series of length T with dimensions. 50 00:04:20,700 --> 00:04:26,220 OK, so this is a plot of all the details series at once, nothing special, but it gives you a sense 51 00:04:26,220 --> 00:04:27,840 of what the data looks like. 52 00:04:32,720 --> 00:04:38,960 OK, so the next step is to build our model, we'll begin by creating two empty lists to store inputs 53 00:04:38,960 --> 00:04:39,820 and features. 54 00:04:40,730 --> 00:04:45,350 The features are just the output of each mini neural network that will process different dimensions 55 00:04:45,500 --> 00:04:46,700 of the Time series. 56 00:04:49,890 --> 00:04:55,060 The next step is to enter a loop that goes these times inside the loop. 57 00:04:55,080 --> 00:05:01,410 We declare an input object with the shape T. So this is what we have for processing a univariate time 58 00:05:01,410 --> 00:05:01,960 series. 59 00:05:02,670 --> 00:05:06,870 Basically, we're just processing a univariate time series D Times. 60 00:05:08,220 --> 00:05:10,330 The next step is to create some dense layers. 61 00:05:10,920 --> 00:05:16,410 Note that I've chosen the number of hidden units in the number of hidden layers arbitrarily, so please 62 00:05:16,410 --> 00:05:18,150 feel free to try different things. 63 00:05:19,470 --> 00:05:20,310 At this point. 64 00:05:20,340 --> 00:05:24,410 The final X variable refers to the output of this many neural network. 65 00:05:24,810 --> 00:05:30,120 So we append AI to the list of inputs and we appendix to the list of features. 66 00:05:32,430 --> 00:05:38,610 OK, so once that loop is done, the features list will contain the outputs of all the mini neural networks. 67 00:05:39,180 --> 00:05:43,260 The next step is to create the concatenate layer, which joins other features together. 68 00:05:44,160 --> 00:05:50,220 So since you have nine feature vectors each of size 16, the result will be a big feature vector of 69 00:05:50,220 --> 00:05:51,980 length, nine times 16. 70 00:05:53,160 --> 00:05:58,630 Once we've done our concatenation, we have one final dense layer with K outputs as usual. 71 00:05:59,610 --> 00:06:04,800 The last step is to create our model objects passing in the list of inputs and the output x. 72 00:06:11,140 --> 00:06:16,270 OK, so when you have a complicated model like this, it's useful to create a picture of your model, 73 00:06:16,720 --> 00:06:20,170 the tensor flow plot model function can be used to do this. 74 00:06:24,940 --> 00:06:30,490 So as you can see, our model is basically nine parallel mini neural networks, all taking in a different 75 00:06:30,490 --> 00:06:32,680 part of the multivariate time series. 76 00:06:33,130 --> 00:06:38,470 After the data goes through these many neural networks, they are combined together into a single feature 77 00:06:38,470 --> 00:06:38,950 vector. 78 00:06:39,640 --> 00:06:43,060 That single feature vector then goes through one final dense layer. 79 00:06:43,750 --> 00:06:46,440 OK, so this is a multi tailed neural network. 80 00:06:49,430 --> 00:06:51,230 The next step is to create a checkpoint. 81 00:06:51,770 --> 00:06:57,210 This will be used to save the best model according to some validation metric during the training process. 82 00:06:57,770 --> 00:06:59,440 So we have three arguments. 83 00:06:59,840 --> 00:07:05,570 The first argument is the file name to save the best model to the second argument is which metric to 84 00:07:05,570 --> 00:07:06,170 monitor. 85 00:07:06,680 --> 00:07:11,350 You could choose the laws here, but since we are paying attention to accuracy, that's what I've chosen. 86 00:07:12,410 --> 00:07:16,880 The third argument is used so that it doesn't save every model, but only the best model. 87 00:07:20,350 --> 00:07:25,600 OK, so the next step is to call the compile function, this is similar to what we've seen before, 88 00:07:25,630 --> 00:07:28,330 except for two things for the loss. 89 00:07:28,330 --> 00:07:34,220 Instead of specifying a string, we specify an object of type, sparse, categorical cross entropy. 90 00:07:34,720 --> 00:07:39,010 We're going to set from sequel to true since our model outputs Logits. 91 00:07:39,400 --> 00:07:44,290 If we use the default value of false, you would have to end your neural network with a softmax. 92 00:07:45,010 --> 00:07:49,070 You don't have to worry about why we do this, but the reason is for numerical stability. 93 00:07:49,510 --> 00:07:53,560 As always, you can check extra reading text for more information. 94 00:07:54,490 --> 00:07:58,630 The next thing to notice is that we've passed in accuracy as the metric to track. 95 00:07:59,320 --> 00:08:02,230 So when we're doing forecasting, this is not really needed. 96 00:08:02,230 --> 00:08:08,290 Since the mean squared error is both the loss and the metric we to track for this problem is different 97 00:08:08,290 --> 00:08:11,260 because the loss and accuracy are two separate things. 98 00:08:14,710 --> 00:08:19,240 The next step is to convert our data into a format that will work with our neural network. 99 00:08:19,780 --> 00:08:23,310 As you recall, our model is defined to take in a list of inputs. 100 00:08:23,650 --> 00:08:27,490 It does not accept an array of size N by T by D directly. 101 00:08:28,270 --> 00:08:32,950 That would be the case for CNN's and Arnon's, but it's not the case for a special, multi tailored 102 00:08:32,950 --> 00:08:33,510 model. 103 00:08:34,210 --> 00:08:39,100 So the function we built earlier is more general in that IT future proves us for the later sections 104 00:08:39,100 --> 00:08:40,610 on CNN's in Arnon's. 105 00:08:41,260 --> 00:08:46,360 So basically we're going to live through each dimension of our Time series and separate each dimension 106 00:08:46,540 --> 00:08:49,480 so that we get a list of univariate time series. 107 00:08:53,880 --> 00:08:59,460 The next step is to call our fifth function, so the arguments to this are the same as we've seen before, 108 00:08:59,490 --> 00:09:03,210 except we are now passing in a list of inputs where X would normally go. 109 00:09:03,360 --> 00:09:04,770 And we have our checkpoint's. 110 00:09:13,540 --> 00:09:15,730 The next step is to plot our lansberry prepack. 111 00:09:20,850 --> 00:09:22,950 OK, so here is our Lansberry Park. 112 00:09:23,340 --> 00:09:26,430 Notice that the train loss is much better than the loss. 113 00:09:33,570 --> 00:09:35,470 And this is our accuracy report. 114 00:09:36,030 --> 00:09:38,160 So we see the same pattern as the Los. 115 00:09:42,100 --> 00:09:47,830 The next step is to load back our best model by calling the load model function, notice that we pass 116 00:09:47,830 --> 00:09:50,530 in the filename that matches with our checkpoint. 117 00:09:53,390 --> 00:09:57,020 The next step is to call the Predix function to get our test predictions. 118 00:10:00,380 --> 00:10:02,780 The next step is to check the accuracy of our model. 119 00:10:03,680 --> 00:10:05,580 OK, so what's happening here? 120 00:10:06,260 --> 00:10:11,030 So as you recall, test is an endemic matrix of probabilities. 121 00:10:11,630 --> 00:10:15,720 In order to convert these to labels, we need to take the Amax on each row. 122 00:10:16,370 --> 00:10:19,910 The results of this will be a one degree of predictive labels. 123 00:10:20,850 --> 00:10:26,330 OK, so once we have the predicted labels, we can use the equals equals operator to compare them to 124 00:10:26,330 --> 00:10:27,180 Y test. 125 00:10:27,680 --> 00:10:33,200 So this does an element wise comparison and the result will be an array of booleans either true if we 126 00:10:33,200 --> 00:10:34,520 are correct and false. 127 00:10:34,520 --> 00:10:41,460 Otherwise, as you recall in Python, a true is treated like one and false is treated like zero. 128 00:10:41,810 --> 00:10:46,910 Therefore, when we take the mean of this array, we get the accuracy, which is ninety two point six 129 00:10:46,910 --> 00:10:47,600 percent. 130 00:10:48,440 --> 00:10:51,020 OK, so this seems like a pretty decent result. 131 00:10:51,290 --> 00:10:53,990 But then again, we don't have any point of comparison.