1
00:00:00,300 --> 00:00:08,220
All right so in this lesson the goal is to make predictions based on what it is that we've been drawing

2
00:00:08,310 --> 00:00:09,830
on our canvas.

3
00:00:10,200 --> 00:00:16,170
So we're now getting to the point where we're moving away from the data processing and getting to the

4
00:00:16,260 --> 00:00:23,280
exciting part which is using the process data to make predictions using the tenth floor light model

5
00:00:23,430 --> 00:00:25,380
that we've trained.

6
00:00:25,380 --> 00:00:33,530
And that means all we need to do is create a tensor and then make predictions on our data.

7
00:00:33,570 --> 00:00:36,830
So how do we create a tensor.

8
00:00:36,880 --> 00:00:44,190
Do you remember if we look at this page on the documentation then we can scroll down and we get a little

9
00:00:44,190 --> 00:00:49,750
code snippet here and what we see is that creating a tensor is very very simple.

10
00:00:49,770 --> 00:00:57,690
We just use the tensor flow module so TAF put a doctorate and then called us tensor method and inside

11
00:00:57,690 --> 00:01:06,780
the parentheses we supply our data and this data here is in the form of an array and we've got two ways

12
00:01:06,780 --> 00:01:14,400
that we can create a tensor either using the square bracket notation like so or providing a shape of

13
00:01:14,400 --> 00:01:20,850
the tens of first and then providing the data in form of an array like so but also providing the shape

14
00:01:21,570 --> 00:01:28,230
when we tested our tensor flow J.S. model in the very beginning of this module just after converting

15
00:01:28,230 --> 00:01:33,150
it to tensor flow jazz we use the double square bracket notation.

16
00:01:33,240 --> 00:01:39,450
So at the very top of our processing jazz file we had some sample data and we had the double square

17
00:01:39,450 --> 00:01:43,440
brackets and then we used that to create our tensor.

18
00:01:43,440 --> 00:01:50,220
So we're gonna stick with this notation but what we're going to do next is we're gonna use the pixel

19
00:01:50,220 --> 00:01:55,920
values that we're grabbing from the canvas instead of the sample data.

20
00:01:56,670 --> 00:02:04,590
So let's create our tensor down here and I want to stored a constant called Capital X to go with the

21
00:02:04,590 --> 00:02:13,330
convention for naming feature vectors and I'm going to use tensor flow with adult and then tensor and

22
00:02:13,380 --> 00:02:21,200
in the parentheses going to supply the pixel values within a pair of square brackets.

23
00:02:21,210 --> 00:02:24,660
That's because we want our tensor to be two dimensional.

24
00:02:24,660 --> 00:02:25,270
Right.

25
00:02:25,320 --> 00:02:32,130
So I've got two pairs of square brackets the one inside pixel values and then one just outside of it

26
00:02:32,280 --> 00:02:34,050
like so.

27
00:02:34,050 --> 00:02:40,980
So now that we have our tensor why don't we log the shape and the data type for this tensor for let's

28
00:02:40,980 --> 00:02:51,580
add two lock statements and the first one is gonna read shape of tensor and that's going to be X dot

29
00:02:51,910 --> 00:02:55,780
shape which is a method from tensor flow jars.

30
00:02:55,930 --> 00:03:02,020
And we're also going to log the data type just so we see what kind of data is actually stored inside

31
00:03:02,200 --> 00:03:03,730
our tensor.

32
00:03:03,790 --> 00:03:10,120
Now if you remember we're hoping that it's gonna be a floating point number but we should verify this

33
00:03:10,210 --> 00:03:13,070
in the logs like so.

34
00:03:13,120 --> 00:03:13,720
All right.

35
00:03:13,780 --> 00:03:19,660
So now that we have our two log statements and we are logging this property here the shape of property

36
00:03:19,780 --> 00:03:29,220
of our tensor and the data type property of our tensor let's save and refresh and head back into our

37
00:03:29,230 --> 00:03:33,910
Web site which should be running on your local host server.

38
00:03:33,940 --> 00:03:40,450
So when I refresh this and I'm going to have my logs here up on the right hand side and when I draw

39
00:03:40,450 --> 00:03:48,490
something in here and then just click check and so now I get a lot of logs but I don't actually see

40
00:03:48,550 --> 00:03:52,850
the log statements that I've just added so I've saved this file.

41
00:03:52,900 --> 00:03:55,290
This is saved but I still don't see it.

42
00:03:55,300 --> 00:04:03,870
So you might have to once again clear browsing data just to be sure that that's not the reason why I

43
00:04:03,880 --> 00:04:06,120
refresh again draw something in here.

44
00:04:07,470 --> 00:04:10,980
Click check answer then I see my two log statements here.

45
00:04:10,980 --> 00:04:17,190
So this is the shape that I'm expecting to see one by seven hundred and eighty four and my tensor is

46
00:04:17,190 --> 00:04:20,730
indeed of data type float 32.

47
00:04:20,730 --> 00:04:22,750
Fantastic.

48
00:04:22,770 --> 00:04:26,560
So now we can use our tensor to make a prediction.

49
00:04:26,640 --> 00:04:31,130
We've actually already seen how to make predictions at the very top of this file.

50
00:04:31,200 --> 00:04:35,970
So the line that I'm talking about is this one right here.

51
00:04:35,970 --> 00:04:41,190
Model don't predict and then the tensor in the parentheses.

52
00:04:41,190 --> 00:04:46,390
That's how we made a prediction based on our sample data at this point.

53
00:04:46,410 --> 00:04:53,350
What I'd like you to do is try and make a prediction based on the data coming from the canvas as a challenge.

54
00:04:53,400 --> 00:05:02,910
What I'd like you to do is reuse this model here and make the prediction inside are predict image method.

55
00:05:02,910 --> 00:05:11,100
I'll give you a few seconds to pause the video before I show you the solution.

56
00:05:11,110 --> 00:05:11,450
All right.

57
00:05:11,470 --> 00:05:12,240
Ready.

58
00:05:12,280 --> 00:05:22,180
So we've loaded our model inside the load model function but we can't use this model inside our predict

59
00:05:22,180 --> 00:05:27,680
image function because of the scope of this variable.

60
00:05:27,700 --> 00:05:33,730
So at the moment it has a local scope it only exists inside load model.

61
00:05:33,730 --> 00:05:40,840
If we wanted to access the model variable inside predict image we would only be able to do that if our

62
00:05:40,840 --> 00:05:43,230
model had a global scope.

63
00:05:43,240 --> 00:05:50,440
So what we need to do is change this variable scope and we can do that simply by defining or declaring

64
00:05:50,440 --> 00:05:51,600
our variable appear.

65
00:05:51,670 --> 00:05:59,440
So VaR model and instead of creating the model down here on line 7 I'm instead assigning it a value

66
00:05:59,920 --> 00:06:01,100
at this point.

67
00:06:01,120 --> 00:06:09,030
So I've created it up here and it's getting the value inside Load model and that promotes it to a global

68
00:06:09,030 --> 00:06:14,200
variable allowing me to use it down here to make my prediction.

69
00:06:14,220 --> 00:06:22,490
So now I can see model not predict and I can supply my capital X feature vector as an input.

70
00:06:22,590 --> 00:06:28,990
I'm actually going to store the result of this prediction in a constant called result.

71
00:06:29,060 --> 00:06:36,240
So I'll set result equal to model not predict and then what I'll do is I'm going to call it result dot

72
00:06:36,270 --> 00:06:43,860
print in order to show what the result is in the console and that completes the challenge.

73
00:06:43,920 --> 00:06:48,350
Let's say the changes come back into chrome and see this works.

74
00:06:48,540 --> 00:06:50,080
So I'm going to refresh.

75
00:06:50,130 --> 00:06:51,700
I'm going to draw something in here.

76
00:06:51,730 --> 00:06:55,340
Click check answer and there at the very bottom.

77
00:06:55,440 --> 00:06:59,540
I should see a log output which gives the prediction of what I drew.

78
00:06:59,580 --> 00:07:00,920
So I draw three.

79
00:07:01,000 --> 00:07:02,700
Let's try again.

80
00:07:02,700 --> 00:07:06,660
So it seems to be working which is great news for us right.

81
00:07:06,780 --> 00:07:08,070
Let's try a few more numbers.

82
00:07:08,160 --> 00:07:12,760
Let's try a 1 Let's try a 4.

83
00:07:12,950 --> 00:07:20,150
Like so let's try a 4 maybe drawn like so that's not bad.

84
00:07:20,150 --> 00:07:21,410
That's pretty encouraging right.

85
00:07:21,800 --> 00:07:26,810
So it seems to be correctly interpreting what it is that we're drawing into our canvas.

86
00:07:26,810 --> 00:07:32,810
Now do you remember the cleanup code that we added with open C.V. at the bottom of our predict image

87
00:07:32,810 --> 00:07:33,530
function.

88
00:07:33,530 --> 00:07:40,070
We had these lines here that deleted all of our open CV objects.

89
00:07:40,070 --> 00:07:46,400
Let's investigate if there's something similar that we have to do with tensor flow light just below

90
00:07:46,400 --> 00:07:47,990
where we're printing our results.

91
00:07:47,990 --> 00:07:58,900
I'm going to add a log statement and this one is going to simply log P.F. dot memory and that's it.

92
00:07:58,910 --> 00:08:04,310
So going to save my changes and then what I'll do is I'll coming out a few of the things that are clogging

93
00:08:04,310 --> 00:08:12,350
up my console someone coming out this line right here where we're logging our scaled array and and then

94
00:08:12,380 --> 00:08:15,770
coming out the pixel values right here.

95
00:08:16,460 --> 00:08:20,400
And I'm comin out these two lines right here.

96
00:08:20,400 --> 00:08:26,540
Now if you're wondering how I'm doing this so quickly it's because I'm using a keyboard shortcut I'm

97
00:08:26,810 --> 00:08:29,570
passing command and then the forward slash.

98
00:08:29,660 --> 00:08:36,770
But if you're working on Windows it'll be the control key and the forward slash on your keyboard.

99
00:08:36,770 --> 00:08:37,040
All right.

100
00:08:37,310 --> 00:08:43,150
So I'm going to save my changes head back into him and refresh.

101
00:08:43,280 --> 00:08:50,710
Now let's see what happens as we are using this Web site and predicting our answer over and over again.

102
00:08:50,720 --> 00:08:58,310
So I want to write a little two here click check and then here's the log coming from TFT memory at the

103
00:08:58,310 --> 00:08:58,790
moment.

104
00:08:58,790 --> 00:09:04,680
I've got 13 tenses and I've got this many bytes in my memory.

105
00:09:04,700 --> 00:09:06,160
Let's try another number.

106
00:09:06,240 --> 00:09:14,060
Quick check answer again and what we see is that I've got two more tenses and the number of bytes has

107
00:09:14,150 --> 00:09:22,400
increased so I'm using more and more memory as I'm checking my answer again and again and hitting that

108
00:09:22,400 --> 00:09:25,700
predict image function in my code.

109
00:09:25,700 --> 00:09:34,070
So that's a problem actually we are bleeding memory as the site is being used more and more and this

110
00:09:34,070 --> 00:09:40,190
suggests that we should manage our memory just like we did with open CV.

111
00:09:40,190 --> 00:09:46,340
If we take a look at this documentation here on tensor flow and we scroll down a little bit you've actually

112
00:09:46,340 --> 00:09:55,010
got this section here on memory and here what it says is that we must manage our memory explicitly and

113
00:09:55,010 --> 00:09:59,870
we can do that by calling this disposed method on our tenses.

114
00:10:00,110 --> 00:10:03,920
In other words calling dispose will free up our memory.

115
00:10:03,920 --> 00:10:09,030
Looking back at our code let's investigate how many tenses we've actually got.

116
00:10:09,080 --> 00:10:16,310
We've got this one right here which we've just created but result actually is also a tensor.

117
00:10:17,300 --> 00:10:22,580
So one of the things we should do is maybe we're we're adding our cleanup code at the end is called

118
00:10:22,640 --> 00:10:31,460
dispose on capital X which is our tensor holding onto our features and we should also call dispose on

119
00:10:31,460 --> 00:10:38,930
the other tensor that we've got like result but this isn't the only place that we've got tense it's

120
00:10:38,930 --> 00:10:39,860
right.

121
00:10:39,860 --> 00:10:46,670
We also have tenses up here namely this one right here and this one right here.

122
00:10:46,670 --> 00:10:53,220
Now given that this is sample data we are actually not going to call this pose on these tenses.

123
00:10:53,270 --> 00:10:58,670
Instead what we're gonna do is we're just gonna delete this code here and we're just going to be left

124
00:10:58,700 --> 00:11:03,040
with loading the model inside the load model function.

125
00:11:03,200 --> 00:11:04,550
So that solves that problem.

126
00:11:05,640 --> 00:11:12,810
Let's save the changes in processing got Jesus and head back into chrome and refresh.

127
00:11:12,810 --> 00:11:20,360
Now let's check what happens as we draw on canvas and check the answers so the first time I've got 11

128
00:11:20,360 --> 00:11:23,740
tenses and I've got this many bytes in memory.

129
00:11:23,960 --> 00:11:30,430
The second time round I still have 11 tenses and I'll have the same number of bytes in memory.

130
00:11:30,500 --> 00:11:33,830
So this is really really promising news right.

131
00:11:33,860 --> 00:11:39,320
We're no longer bleeding memory and we're using tensor flow Jesus as intended.

132
00:11:39,320 --> 00:11:43,010
In other words we're being good citizens and we're cleaning up after ourselves.

133
00:11:43,130 --> 00:11:44,760
So what's left to do now.

134
00:11:44,780 --> 00:11:48,830
Well we still have to write the logic for checking our equation.

135
00:11:48,830 --> 00:11:49,260
Right.

136
00:11:49,280 --> 00:11:53,380
Three plus two is not equal to one as far as I know.

137
00:11:53,450 --> 00:12:01,250
So let's head into this code and make sure we add the logic to actually check that what's being drawn

138
00:12:01,250 --> 00:12:04,160
into the canvas is correct or not.

139
00:12:04,250 --> 00:12:09,810
And that's exactly what we're going to be doing in the next lesson.

140
00:12:09,890 --> 00:12:15,050
But before we finish off this lesson let's do a little bit more cleaning up.

141
00:12:15,050 --> 00:12:20,900
Let's stop creating a new canvas every time we push the button and also let's clean up what's going

142
00:12:20,900 --> 00:12:23,320
on in our console here on the right.

143
00:12:23,330 --> 00:12:27,960
It's getting very very busy and starting to be a little bit distracting.

144
00:12:27,980 --> 00:12:34,430
So what we're going to do is we're simply going to comment out all the long statements that we've previously

145
00:12:34,520 --> 00:12:35,380
added.

146
00:12:35,690 --> 00:12:43,790
And we're also going to delete the testing code that we had at the bottom in order to sort of visualize

147
00:12:44,000 --> 00:12:46,730
what we're doing with the data pre processing.

148
00:12:46,730 --> 00:12:54,470
So I'm just going to comment everything out both in the processing not J.S. and inside the drawing not

149
00:12:54,570 --> 00:13:01,970
J.S. when a comment on all the log statements I encourage you to do the same so that in the next lesson

150
00:13:02,120 --> 00:13:09,560
we can look at the console and just see very very few log statements in there and focus on what's important

151
00:13:09,620 --> 00:13:12,230
and relevant for that lesson.

152
00:13:12,230 --> 00:13:12,770
All right.

153
00:13:12,830 --> 00:13:15,010
I think that's all of them.

154
00:13:15,020 --> 00:13:21,050
Let me just save the changes here and both of these files and when I refresh in hand just make sure

155
00:13:21,050 --> 00:13:27,640
that up looks like I missed this one and this one or did I.

156
00:13:27,680 --> 00:13:29,840
Let me just go to more tools.

157
00:13:29,840 --> 00:13:31,150
Clear the browsing data.

158
00:13:31,790 --> 00:13:40,970
I'm not really sure why I have to do this so often and refresh again and see if I've got a busy console

159
00:13:41,120 --> 00:13:43,780
or a clear console.

160
00:13:43,850 --> 00:13:49,630
So now all I'm seeing here are my predictions from my tensor.

161
00:13:49,790 --> 00:13:51,790
So this is all I want to be left with at this point.

162
00:13:51,950 --> 00:13:57,530
I just want to be left with the predictions from printing out the result and that's it.

163
00:13:57,800 --> 00:14:03,410
So I'll see you in the next lesson where we're gonna be using our prediction and writing our gaming

164
00:14:03,410 --> 00:14:03,830
logic.