1
00:00:11,010 --> 00:00:15,420
In this video, we'll be looking at yet another way to build a neural network.

2
00:00:16,350 --> 00:00:22,620
This time our goal will be to combine the two types of models we just learned how to build, specifically

3
00:00:22,620 --> 00:00:28,930
the multi input time series CNN and the feature based CNN by using concatenation.

4
00:00:29,100 --> 00:00:34,470
You'll see how easy it is to combine all different kinds of data sources into one big neural network.

5
00:00:35,970 --> 00:00:39,180
So we'll start by creating two lists for the inputs and features.

6
00:00:39,690 --> 00:00:45,330
This part is the same as the multi input enen where we use a separate mini neural network for each dimension

7
00:00:45,330 --> 00:00:46,410
of the Time series.

8
00:00:49,050 --> 00:00:54,360
The new part is the second part where we create a new input and a new mini neural network to process

9
00:00:54,360 --> 00:00:56,700
the features, so we call these Eita.

10
00:00:56,740 --> 00:00:57,570
Next to.

11
00:01:01,000 --> 00:01:07,090
The next step is to concatenate all of these features together, so previously this was just the features

12
00:01:07,090 --> 00:01:07,540
list.

13
00:01:07,810 --> 00:01:12,310
But now we have one additional item, which is next to now.

14
00:01:12,310 --> 00:01:14,390
People always ask about this syntax.

15
00:01:14,830 --> 00:01:19,300
Remember that in Python, a list plus a list just puts the two lists together.

16
00:01:20,530 --> 00:01:24,820
The next step is to pass this big feature vector into one final dense layer.

17
00:01:25,600 --> 00:01:30,820
The last line of code is where we create our model object, passing in a list of all the inputs and

18
00:01:30,820 --> 00:01:31,810
the output x.

19
00:01:39,550 --> 00:01:42,730
The next step is to call the compile function the same as before.

20
00:01:47,830 --> 00:01:49,660
The next step is to create a new checkpoint.

21
00:01:50,200 --> 00:01:51,970
Again, we use a separate file name.

22
00:01:56,190 --> 00:02:02,700
The next step is to call the Fed function, notice again that our inputs are now lists of inputs, so

23
00:02:02,700 --> 00:02:06,420
we simply concatenate the static features with what we had before.

24
00:02:14,060 --> 00:02:16,280
The next step is to plot the Lansberry Park.

25
00:02:21,070 --> 00:02:23,020
OK, so nothing out of the ordinary.

26
00:02:26,590 --> 00:02:29,230
The next step is to plot the accuracy prepack.

27
00:02:33,540 --> 00:02:37,580
So it looks like this model may perform a bit better than what we had before.

28
00:02:41,010 --> 00:02:44,520
The next step is to load up our best model and check its performance.

29
00:02:50,270 --> 00:02:55,110
OK, so we see that our combine's model does a bit better than the individual models.

30
00:02:55,910 --> 00:03:01,490
However, it still does not perform as well as a linear model that's learned only from the static features.