1
00:00:00,770 --> 00:00:06,350
In this lesson we're going to talk about named scopes and the reason we're gonna talk about name scopes

2
00:00:06,710 --> 00:00:13,040
is so that we can start grouping different types of operations together in our tensor flow graph.

3
00:00:13,040 --> 00:00:18,740
And while we're doing that we'll also get a chance to talk about some Python programming concepts in

4
00:00:18,740 --> 00:00:19,310
particular.

5
00:00:19,610 --> 00:00:23,340
We're going to make use of something called the context manager.

6
00:00:23,540 --> 00:00:29,960
Wouldn't it be nice to group certain operations together in our graph for example this set of operations

7
00:00:29,960 --> 00:00:33,220
here all happens in the first hidden layer.

8
00:00:33,320 --> 00:00:38,000
We've got our weights that are multiplied with our inputs and then we've got our biases that are added

9
00:00:38,000 --> 00:00:43,920
to our weights after the multiplication and then we have all of that feed into our activation function.

10
00:00:43,940 --> 00:00:47,690
All of this could be represented as a single unit.

11
00:00:47,990 --> 00:00:53,660
If we do that for every single layer we can see how the different layers are connected on our tensor

12
00:00:53,660 --> 00:01:01,190
flow graph the way we can accomplish this intense flow is with something called a name scope the name

13
00:01:01,190 --> 00:01:03,770
scope will help us organize our calculations.

14
00:01:03,770 --> 00:01:08,100
And it does this using something called the context manager.

15
00:01:08,120 --> 00:01:13,190
So if you think about it all the different calculations that are part of say the first hidden layer

16
00:01:13,450 --> 00:01:16,450
they should really belong to the same context.

17
00:01:16,490 --> 00:01:29,180
So with T F dot name on a score scope open parentheses single quotes output underscore layer colon at

18
00:01:29,180 --> 00:01:37,230
the end and then select all of these lines of code and hit tab on my keyboard to invent them.

19
00:01:37,340 --> 00:01:42,460
And what this does is it makes all of these lines part of the same block.

20
00:01:42,800 --> 00:01:49,730
All of these lines of code now have the same context and that context has a name and that name is going

21
00:01:49,730 --> 00:01:51,380
to be output layer.

22
00:01:51,860 --> 00:01:58,880
So so all of these operations that I've got here that all belong to my output layer should not be grouped

23
00:01:59,300 --> 00:02:02,300
in tensor bold so let's see if this works.

24
00:02:02,300 --> 00:02:11,480
So let's see if I can come here to set up tensor flow graph and go to cell run all below and update

25
00:02:11,750 --> 00:02:15,030
my graph looks the same as before.

26
00:02:15,080 --> 00:02:22,600
Well it's because it is the same as before I found my model at 15 0 6 and at 15 25.

27
00:02:22,790 --> 00:02:27,530
So that's why I have to change this to my most recent run.

28
00:02:28,250 --> 00:02:34,650
And now you can see all the operations that have the same name scope grouped in my output layer.

29
00:02:34,910 --> 00:02:40,790
If I wanted to peep inside and see what operations there are actually inside all I need to do is double

30
00:02:40,790 --> 00:02:47,030
click and then I can see all the operations all the calculations that I've grouped under the same name

31
00:02:47,030 --> 00:02:47,960
scope.

32
00:02:48,080 --> 00:02:55,400
So what I'd like you to do is pause the video real quick and group all the other operations that belong

33
00:02:55,400 --> 00:02:57,220
to the second hidden layer.

34
00:02:57,620 --> 00:03:01,730
And the first hidden layer to clean up our graph.

35
00:03:01,910 --> 00:03:05,650
I'll give you a few seconds to pause the video and give this a go.

36
00:03:10,870 --> 00:03:12,140
Ready.

37
00:03:12,160 --> 00:03:14,050
Here's the solution.

38
00:03:14,050 --> 00:03:19,600
So we have to do a little bit of reorganization in our code to add a name scope to our second hidden

39
00:03:19,600 --> 00:03:20,290
layer.

40
00:03:20,290 --> 00:03:27,610
It's pretty straightforward because all we need to do is write with T F name underscore scope and we

41
00:03:27,610 --> 00:03:36,290
can give us any name we want hidden the square to semicolon and use this right here and indented.

42
00:03:36,400 --> 00:03:40,360
So that's our second hidden layer done for our first hidden layer.

43
00:03:40,600 --> 00:03:44,530
We're gonna have to move some of our code to the same cell to make our life easier.

44
00:03:44,710 --> 00:03:51,150
So I'll take this one him with him take this bit of code him move it below.

45
00:03:51,160 --> 00:03:52,800
Same with this one.

46
00:03:52,990 --> 00:04:01,960
Move it him and I'll take these three cells here and I'm going to delete them and now I can add the

47
00:04:01,960 --> 00:04:13,780
same line with TAF dot name scope open parentheses single quotes hidden under square one and colon at

48
00:04:13,780 --> 00:04:14,750
the end.

49
00:04:15,050 --> 00:04:22,870
Take this indented and then what I can do is simply go to this markdown cell here and see run all below

50
00:04:23,370 --> 00:04:29,650
back in tense of board refresh my page and then I'm gonna go to my most recent model this one right

51
00:04:29,650 --> 00:04:37,510
here and here we can already see that our layers are grouped very very nicely now we've got our inputs

52
00:04:37,660 --> 00:04:43,840
our axis going to our first hidden layer and the output of our first layer is going to our second hidden

53
00:04:43,840 --> 00:04:49,900
layer and the output of the second hidden layer is going to our third hidden layer our output layer

54
00:04:50,470 --> 00:04:56,380
but I tell you what looking back at our code it's starting to look a little bit repetitive right we've

55
00:04:56,380 --> 00:05:02,350
got this block right here we've got this block right here and this block right here I think that we

56
00:05:02,350 --> 00:05:09,610
can all see that most of this code is actually the same so why not create a single function that will

57
00:05:09,610 --> 00:05:18,070
set up all our layers for us I'm going to define a function kind of call that function setup underscore

58
00:05:18,520 --> 00:05:27,040
layer and it's going to take maybe three inputs that first input I'll just call input the second input

59
00:05:27,460 --> 00:05:34,090
I'll call it Wait underscored dimensions the I am the third input I'll call bias underscore dimensions

60
00:05:34,480 --> 00:05:42,010
and that fourth input I'll call name and what I'm going to do is I'm going to half with statement with

61
00:05:42,010 --> 00:05:52,580
block here that reads TMF dot name scope open parentheses name colon and then inside that block I'm

62
00:05:52,580 --> 00:05:58,520
basically going to do this I'm basically going to take everything I normally do to set up a layer and

63
00:05:58,640 --> 00:06:01,350
I'll do that inside my function.

64
00:06:01,350 --> 00:06:07,450
I won't have to tab over a little bit and actually fix this to having him but going back a little bit

65
00:06:07,550 --> 00:06:10,000
so shift temp will go back one.

66
00:06:10,190 --> 00:06:14,160
And now I can start replacing things that will vary with my parameters.

67
00:06:14,270 --> 00:06:17,830
So this shall read wait I mentioned the.

68
00:06:17,930 --> 00:06:20,720
I'm actually just gonna call W and the biases.

69
00:06:20,750 --> 00:06:27,260
I'm just going to call a capital B the shape for my biases are gonna be my bias dimension and then I'll

70
00:06:27,260 --> 00:06:29,360
make my variable names a little bit more generic.

71
00:06:29,370 --> 00:06:35,810
This is gonna be initial w this is gonna be w and there's gonna be initial w this is gonna be initial

72
00:06:35,810 --> 00:06:36,680
B.

73
00:06:36,890 --> 00:06:43,520
This is gonna be initial B and this will just be b layer and here's where it gets interesting again.

74
00:06:43,700 --> 00:06:48,910
The input to this layer is gonna be equal to this parameter here.

75
00:06:48,950 --> 00:06:56,600
My input parameter and it's gonna multiply those input times the weights plus the biases and then the

76
00:06:56,690 --> 00:06:58,480
output has to vary right.

77
00:06:58,490 --> 00:07:04,670
The output has to be different for the output layer versus the first layer or the second layer.

78
00:07:05,360 --> 00:07:06,790
So I'm gonna do this.

79
00:07:06,890 --> 00:07:17,180
Well I'll tell you what if the name is equal to W equals two out for output.

80
00:07:17,180 --> 00:07:20,950
In that case my output will be equal to soft Max.

81
00:07:21,170 --> 00:07:25,880
So T F Dot and Dot soft Max parentheses layer underscore in.

82
00:07:26,210 --> 00:07:34,610
However if the name is different from the word out then we're simply going to say the output is going

83
00:07:34,610 --> 00:07:37,930
to be equal to T F dot and then die.

84
00:07:37,930 --> 00:07:38,440
Really.

85
00:07:39,110 --> 00:07:45,740
Which is what we have for all the hidden layers and then layer underscore in my function will then return

86
00:07:46,190 --> 00:07:47,370
my output.

87
00:07:47,480 --> 00:07:48,700
So that should be fairly generic.

88
00:07:48,700 --> 00:07:49,260
Right.

89
00:07:49,280 --> 00:07:55,430
This will allow us to insert a cell below and set up all of our layers and a couple of lines of code

90
00:07:56,000 --> 00:07:57,430
layer underscore one.

91
00:07:57,860 --> 00:08:04,590
It's gonna be equal to set up underscore layer open parentheses and then what's gonna be the input the

92
00:08:04,610 --> 00:08:06,300
input the first hidden layer.

93
00:08:06,330 --> 00:08:14,210
It's gonna be our place holder capital X. The Wait underscored dimensions are gonna be equal to square

94
00:08:14,210 --> 00:08:15,280
brackets.

95
00:08:15,590 --> 00:08:23,240
The total number of inputs and then and then the other dimension for this one is gonna be and one hidden

96
00:08:23,240 --> 00:08:26,060
one so five hundred and twelve.

97
00:08:26,140 --> 00:08:33,340
The bias on a score dimension is gonna be equal to square brackets and hidden one.

98
00:08:33,420 --> 00:08:37,880
So the number of neurons again five hundred and twelve.

99
00:08:38,060 --> 00:08:43,760
And then we're gonna call this name as equal to layer on the score one must split this over a couple

100
00:08:43,760 --> 00:08:45,950
of lines two lines.

101
00:08:46,010 --> 00:08:46,580
Here we go.

102
00:08:47,240 --> 00:08:49,110
That's the setup for layer 1.

103
00:08:49,130 --> 00:08:49,920
Done.

104
00:08:49,940 --> 00:08:53,530
I can do the same thing for layer 2 and for our output layer.

105
00:08:54,200 --> 00:09:00,860
So there are 2 and output and I'm actually going to go back to my setup layer function here.

106
00:09:01,370 --> 00:09:07,560
And I'm going to call this thing layer on a score out that way we avoid confusion.

107
00:09:07,610 --> 00:09:12,720
So I'm just gonna rename these variables to not clash with this one.

108
00:09:13,040 --> 00:09:17,360
I want to keep the output of my model with the name output.

109
00:09:17,580 --> 00:09:21,710
Now let me update the weights and the biased dimensions.

110
00:09:21,750 --> 00:09:29,780
Our second hidden layer is of course going to have 512 inputs but it's only gonna have 64 neurons so

111
00:09:29,870 --> 00:09:39,950
an underscore hidden to the output layer is gonna have 64 inputs and underscore hidden 2 and 10 outputs

112
00:09:40,430 --> 00:09:44,150
and R underscore classes for both of these layers.

113
00:09:44,210 --> 00:09:48,230
The number of neurons also has to be reflected in the biased dimension.

114
00:09:48,230 --> 00:09:55,190
So for the second hidden layer it was an underscore hidden too and for the output layer it was again

115
00:09:55,580 --> 00:10:02,270
number of classes the names for layer one and layer 2 where you're going to keep with layer on the square

116
00:10:02,270 --> 00:10:08,420
one and layer on the square 2 and then for the output layer in order to trigger this if statement.

117
00:10:08,420 --> 00:10:10,990
I'm just gonna name it lowercase out.

118
00:10:11,090 --> 00:10:13,390
So this is gonna be the same as this.

119
00:10:13,460 --> 00:10:20,150
So my output layer is gonna have these soft Max activation function and that leaves us with one last

120
00:10:20,150 --> 00:10:21,250
thing to do.

121
00:10:21,350 --> 00:10:28,310
The outputs of the first layer are gonna be the inputs for the second layer and the inputs for the third

122
00:10:28,310 --> 00:10:31,370
layer are gonna be the outputs of the second layer.

123
00:10:31,940 --> 00:10:37,580
So this place will look at this capital X it's only gonna be input for the first layer the input for

124
00:10:37,580 --> 00:10:46,340
the second layer is actually gonna be layer number 1 and the inputs for the output layer is gonna be

125
00:10:46,700 --> 00:10:49,610
layer number 2.

126
00:10:49,610 --> 00:10:57,560
Now what I'm gonna do is I take these lines of code and I want to move them to the very bottom of our

127
00:10:57,560 --> 00:10:58,670
notebook.

128
00:10:58,670 --> 00:10:59,690
So I'm going to keep hold of them.

129
00:10:59,720 --> 00:11:05,960
I'm not going to delete them but I'm going to move them all the way down and I want to set a markdown

130
00:11:05,990 --> 00:11:06,590
cell here.

131
00:11:07,070 --> 00:11:14,310
That's kind of read code for first part of module.

132
00:11:14,460 --> 00:11:16,100
I think what I'll do is I'll come at the Mount.

133
00:11:16,130 --> 00:11:21,540
So we're keeping them around as a reference but we're not using them anymore.

134
00:11:21,560 --> 00:11:23,400
We're not executing them anymore.

135
00:11:23,570 --> 00:11:29,810
Now that we've done all of that we can come back up here to where we've got our mark down set up tensor

136
00:11:29,810 --> 00:11:36,080
flow graph and we're to run all below and what we get is the crash.

137
00:11:36,080 --> 00:11:36,430
Why.

138
00:11:36,860 --> 00:11:40,860
Because we've still got some old code floating around this line here.

139
00:11:41,030 --> 00:11:41,480
Well now.

140
00:11:41,480 --> 00:11:49,610
Error out because there is no more tensor called B so I'll move this line of code down to the very bottom

141
00:11:49,880 --> 00:11:58,130
as well and I'm going to comment it out before I rerun everything I'm going to execute this cell here

142
00:11:58,670 --> 00:12:02,090
which closes our session and resets our graph.

143
00:12:02,090 --> 00:12:06,500
And then when I come back here and try again so I'll run all below.

144
00:12:06,500 --> 00:12:08,990
Now our model starts training again.

145
00:12:08,990 --> 00:12:10,910
We can come back into tensor board.

146
00:12:11,480 --> 00:12:16,880
So now that we've successfully created some named scopes for our layers we can also create some name

147
00:12:16,880 --> 00:12:19,890
scopes for the other calculations that we're doing.

148
00:12:20,060 --> 00:12:25,910
In particular I'm thinking of our optimization because our optimizer is actually connected to all of

149
00:12:25,910 --> 00:12:32,120
our layers and I'm thinking of our lost function and our accuracy calculation.

150
00:12:32,180 --> 00:12:33,750
There's a lot going on here.

151
00:12:33,860 --> 00:12:38,680
There's a lot of nodes and a lot of edges and we can consolidate those.

152
00:12:38,780 --> 00:12:41,060
Let's tackle one at a time.

153
00:12:41,090 --> 00:12:43,890
I'm going to hit up the accuracy calculation first.

154
00:12:43,970 --> 00:12:51,440
So for the accuracy calculation I'm going to provide a name scope that reads with T F dot named scope

155
00:12:51,800 --> 00:13:02,300
open parentheses single quotes accuracy underscore calc colon and tab this over so that it's inside

156
00:13:02,450 --> 00:13:04,580
the body of this scope.

157
00:13:04,580 --> 00:13:11,120
I'm also going to come down here where we've got our two summaries and I'm also going to provide a name

158
00:13:11,120 --> 00:13:13,470
scope for these two with TFT.

159
00:13:13,520 --> 00:13:19,440
Name scope open parentheses and I'm gonna call this performance.

160
00:13:19,520 --> 00:13:21,410
This is gonna be the performance named scope.

161
00:13:21,500 --> 00:13:25,790
Next I'm gonna tackle the optimizer the optimizer.

162
00:13:25,790 --> 00:13:31,970
I'm just gonna give the name f named scope open parentheses optimizer.

163
00:13:32,180 --> 00:13:33,490
I think that's pretty straightforward.

164
00:13:33,570 --> 00:13:36,150
And also I kind of like imagination.

165
00:13:36,170 --> 00:13:37,890
So there we go.

166
00:13:37,910 --> 00:13:40,180
That's the name scope for optimizer done.

167
00:13:40,220 --> 00:13:44,960
And while we're at it let's add a name scope for our loss function as well.

168
00:13:44,960 --> 00:13:55,730
So with TAF dot name underscore scope parentheses loss on a score calc colon at the end tab everything

169
00:13:55,730 --> 00:14:01,660
over and now let's go to the very top and run all the cells again.

170
00:14:01,910 --> 00:14:03,330
So will it run all the cells below.

171
00:14:04,010 --> 00:14:10,700
They'll go into tensor bode well refresh our graph will refresh our page and we'll take a look at our

172
00:14:10,700 --> 00:14:11,640
latest graph.

173
00:14:11,680 --> 00:14:17,120
Remember changes dropped on here to your most recent run after you've made all the changes and now you'd

174
00:14:17,120 --> 00:14:18,820
see something like this.

175
00:14:18,920 --> 00:14:26,570
You should see that we've got everything a lot more tidy we've got our inputs here x those go into a

176
00:14:26,570 --> 00:14:30,480
layer 1 but they also are connected to our optimizer.

177
00:14:30,530 --> 00:14:35,000
We've cut our first hidden layer feeding through to our second hidden layer which is feeds through to

178
00:14:35,000 --> 00:14:41,600
our output layer and our outputs go into the accuracy calculations as well as the lost calculations

179
00:14:42,180 --> 00:14:48,950
the lost calculations of course also feed through to the optimizer because the optimizer is connected

180
00:14:49,070 --> 00:14:54,860
to all the weights in all the layers and of course it's also connected to the lost calculations.

181
00:14:54,860 --> 00:14:57,530
It's not connected to the accuracy calculations.

182
00:14:57,530 --> 00:14:59,300
It doesn't have anything to do with that.

183
00:14:59,340 --> 00:15:04,030
So looking at the graph in this way when everything's grouped together it's a lot more clear.

184
00:15:04,910 --> 00:15:10,460
But one thing that we can do is right click on this and we can actually remove this thing from the main

185
00:15:10,460 --> 00:15:13,790
graph because it's a little bit distracting it's connected to everything right.

186
00:15:14,330 --> 00:15:18,240
So if we move it to the side then we can see that's still there.

187
00:15:18,680 --> 00:15:25,550
You can still see it connected to the optimizer but we see that this now represents our neural network

188
00:15:25,850 --> 00:15:27,150
in the clearest way yet.

189
00:15:27,290 --> 00:15:33,230
We have our inputs our features our Xs feeding through from the bottom going through all the layers

190
00:15:33,520 --> 00:15:40,420
and at the very top of our graph we have the two outputs that we care about the loss and the accuracy.

191
00:15:40,430 --> 00:15:45,770
Now that almost covers everything I want to talk about with the graph but before I end this lesson I

192
00:15:45,770 --> 00:15:51,400
want to show you one other cool thing that tensor board can do tensor board can actually also show us

193
00:15:51,410 --> 00:15:57,630
some images namely the images of the numbers we have in our dataset to show you how.

194
00:15:57,820 --> 00:16:04,170
We just have to add a single operation here and we add this operation here just above the code we're

195
00:16:04,180 --> 00:16:05,500
running our session.

196
00:16:05,590 --> 00:16:14,890
So add a quick markdown cell here that's gonna read check input images in board and the way we're going

197
00:16:14,890 --> 00:16:18,340
to do this is with a summary.

198
00:16:18,640 --> 00:16:28,420
So there's a tense about summary called Image T F dot summary dot image and then open the parentheses.

199
00:16:28,420 --> 00:16:30,900
And then I want to provide a name for the summary.

200
00:16:30,930 --> 00:16:37,600
Going to call this image underscore input and then I'm going to provide a value.

201
00:16:37,960 --> 00:16:39,510
Now this is the tricky part.

202
00:16:40,120 --> 00:16:43,070
I have to get an image pack out of our data.

203
00:16:43,240 --> 00:16:45,640
So I'm going to leave this for now.

204
00:16:46,240 --> 00:16:52,410
And I'm going to have X underscore image as the variable.

205
00:16:52,410 --> 00:16:55,670
I want to feed into my image summary here.

206
00:16:56,280 --> 00:17:00,460
That's gonna be equal to some data from our placeholder tensor.

207
00:17:00,600 --> 00:17:04,680
All the data and the place holder tensor is like in a single row right.

208
00:17:04,680 --> 00:17:07,110
Seven hundred and eighty four values.

209
00:17:07,110 --> 00:17:16,290
In order to display the image which is 28 by 28 pixels we have to reshape the thing so as CTF data reshape

210
00:17:16,930 --> 00:17:22,410
and I'm going to grab my placeholder tensor and then I have to say how I'm going to reshape it.

211
00:17:22,440 --> 00:17:29,430
So when I add some square brackets him and I'm going to reshape it into the following dimensions minus

212
00:17:29,430 --> 00:17:41,760
one comma 28 comma 28 comma 1 oneness the color channel 28 is the width and 28 is the height at minus

213
00:17:41,760 --> 00:17:48,050
1 is that because I've only got three dimensions and this other dimension would be for all the samples.

214
00:17:48,090 --> 00:17:54,300
In other words I would have more than one image that's 28 by twenty eight with one color channel.

215
00:17:54,300 --> 00:18:04,080
Now that I've got this I can come down here and I can supply X on a square image to my summary and I'm

216
00:18:04,080 --> 00:18:10,620
going to supply one more argument here namely the number of images I actually want to show up intensive

217
00:18:10,620 --> 00:18:10,920
mode.

218
00:18:11,220 --> 00:18:18,600
So Max on the score outputs is the name of the parameter and I'm going to say for I don't want to see

219
00:18:18,600 --> 00:18:26,130
all sixty thousand of them right now since reshaping stuff and working out the summary are an operation

220
00:18:26,130 --> 00:18:26,940
and tensor flow.

221
00:18:27,390 --> 00:18:30,020
We can also give them a name scope.

222
00:18:30,210 --> 00:18:39,210
So I'm going to say with space T F dot name scope open parentheses and then I'll say show on a score.

223
00:18:39,210 --> 00:18:48,120
Image as the name scope for this thing and alt tab this over so that it's in the body of this with block.

224
00:18:48,120 --> 00:18:53,910
Now both this reshape calculation and my summary have the same context.

225
00:18:54,390 --> 00:19:00,040
Let's go back up to where we set up tents aboard and run everything below.

226
00:19:00,240 --> 00:19:05,070
So our models already training and we refresh tents aboard.

227
00:19:05,670 --> 00:19:08,750
And what we should see is a new type of summary.

228
00:19:08,790 --> 00:19:15,680
Previously we only had the scalar and the graph but now we've got another type of summary here namely

229
00:19:15,690 --> 00:19:19,770
our images summary because we added an image summary here.

230
00:19:19,770 --> 00:19:27,090
If I change my graph to my latest run then you will also see this additional node here.

231
00:19:27,090 --> 00:19:35,640
We've got our show image name scope for two calculations namely our summary and our reshape calculation.

232
00:19:35,640 --> 00:19:39,630
These in turn are connected to our features tensor are x.

233
00:19:39,690 --> 00:19:43,820
Now let's check out our images tab here in my images tab.

234
00:19:43,860 --> 00:19:50,460
I get an image summary for every run that I've done and what I'm looking at here are four images from

235
00:19:50,610 --> 00:19:54,350
my training run and my validation run in my training run.

236
00:19:54,390 --> 00:19:59,610
I've got the number six in my validation run I've got the number five showing up if I just wanted to

237
00:19:59,610 --> 00:20:05,370
see one of them I could hit this radio button him so we'll see the screen one I've just got my validation

238
00:20:05,370 --> 00:20:09,360
data set for which I'm looking at these four images why four.

239
00:20:09,390 --> 00:20:11,780
Because that's our max output right.

240
00:20:11,850 --> 00:20:13,530
So that's pretty cool right.

241
00:20:13,530 --> 00:20:20,040
Not only did we learn about name scoping and grouping operations in our tensor flow graph we also got

242
00:20:20,040 --> 00:20:26,010
to use a different kind of summary we've been using this scalar type summaries many many times before

243
00:20:26,250 --> 00:20:29,890
and now we got to see the image summary as well in the next lesson.

244
00:20:29,970 --> 00:20:36,240
We don't try out a few different models and this will allow us to look at a new kind of summary as well.

245
00:20:36,240 --> 00:20:42,570
The histogram the histogram is incredibly powerful because it will allow us to see how our weights are

246
00:20:42,570 --> 00:20:49,920
changing over time as we're running the model and training it over the course of several epochs.

247
00:20:49,920 --> 00:20:53,440
So stay tuned for all of that and more in the next lesson.

248
00:20:53,470 --> 00:20:54,230
I'll see you in a bit.