1
00:00:00,360 --> 00:00:06,300
OK now we've seen a few different ways to manipulate a raise but most of the ways we've worked with

2
00:00:06,300 --> 00:00:08,550
have been arithmetic or aggregation.

3
00:00:08,550 --> 00:00:12,700
So kind of performing calculations on Empire right now.

4
00:00:12,810 --> 00:00:17,460
What if you need to change the shape of your name higher array.

5
00:00:17,610 --> 00:00:20,480
Now why would you need to change the shape of an umpire right.

6
00:00:20,700 --> 00:00:26,910
Well remember a lot of machine learning is lining up your inputs into a machine learning algorithm and

7
00:00:26,910 --> 00:00:31,080
making sure that they come out with the right output.

8
00:00:31,080 --> 00:00:37,950
One of those ways of lining up inputs is making sure your data or your list of numbers or your array

9
00:00:37,950 --> 00:00:40,150
of numbers is in the right shape.

10
00:00:40,170 --> 00:00:43,500
So let's have a look at how to do that in num pi.

11
00:00:43,500 --> 00:00:49,230
So we'll go reshaping and transposing this is this section.

12
00:00:49,230 --> 00:00:51,350
Nice little heading there.

13
00:00:51,380 --> 00:00:52,130
So let's have a look.

14
00:00:52,260 --> 00:00:53,290
A 2 array.

15
00:00:53,310 --> 00:00:54,250
Beautiful.

16
00:00:54,300 --> 00:00:56,330
Let's figure out what shape it is.

17
00:00:56,340 --> 00:00:57,360
Two by three.

18
00:00:57,440 --> 00:00:58,680
Yeah that makes sense.

19
00:00:58,680 --> 00:01:04,630
Now remember before in a previous video we tried to multiply A2 by A3.

20
00:01:04,680 --> 00:01:08,820
Let's visualize a three so we can see it again a three.

21
00:01:08,820 --> 00:01:09,570
There we go.

22
00:01:09,930 --> 00:01:12,910
Let's check out the shape a 3D up shape.

23
00:01:13,750 --> 00:01:21,850
Okay now what's going to happen if we try to multiply these operations could not be broadcast together

24
00:01:22,180 --> 00:01:26,010
with shapes 2 3 or 2 3 3.

25
00:01:26,500 --> 00:01:33,220
Well the beautiful thing about an umpire rains is that they can be reshaped or transposed.

26
00:01:33,220 --> 00:01:35,710
Let's have a look at reshape first.

27
00:01:35,800 --> 00:01:41,580
So what does reshaped to reshape is a method you can call on any array.

28
00:01:41,650 --> 00:01:49,540
So if we type in here dot reshape shift tab reshape dock string returns an array containing the same

29
00:01:49,540 --> 00:01:53,820
data with a new shape beautiful.

30
00:01:53,860 --> 00:01:54,280
There we go.

31
00:01:54,280 --> 00:01:56,320
Get some more information there.

32
00:01:56,380 --> 00:01:58,980
Let's see what it requires shape.

33
00:01:59,330 --> 00:02:00,400
Mm hmm.

34
00:02:00,640 --> 00:02:05,650
So one of the rules of broadcasting if we look up broadcasting one more time

35
00:02:08,880 --> 00:02:12,500
one of the rules of broadcasting is here.

36
00:02:12,630 --> 00:02:17,970
General broadcasting rules when operating on two arrays num pi compares their shapes.

37
00:02:18,000 --> 00:02:21,850
Element wise it starts with the trailing dimensions and works its way forward.

38
00:02:21,990 --> 00:02:27,810
Two dimensions are compatible when they are equal or one of them is one.

39
00:02:27,810 --> 00:02:35,010
So what this means is that if the shapes of two arrays are equal the multiplication can happen but if

40
00:02:35,010 --> 00:02:38,280
they're not equal one of them has to be one.

41
00:02:38,310 --> 00:02:40,170
So that's what we're going to do here.

42
00:02:40,170 --> 00:02:50,310
We're going to reshape the A2 array from its shape to three incompatible here.

43
00:02:50,940 --> 00:02:55,560
And we're going to change one of these dimensions we're going to add a dimension actually to reshape

44
00:02:55,590 --> 00:02:57,880
2 3 1.

45
00:02:57,960 --> 00:03:00,640
Let's see what happens.

46
00:03:00,910 --> 00:03:01,390
OK.

47
00:03:01,400 --> 00:03:03,470
So it's shuffled our array around a bean.

48
00:03:03,650 --> 00:03:07,960
Let's print the shape here so see our A2.

49
00:03:07,960 --> 00:03:11,590
We've used the reshape method passed in a different shape.

50
00:03:11,590 --> 00:03:17,530
So now we've added another dimension here and now its shape is 2 3 1.

51
00:03:17,540 --> 00:03:23,040
We'll check a three shape okay two three three.

52
00:03:23,150 --> 00:03:28,680
So if we refer back to the rules of broadcasting are the dimensions equal.

53
00:03:28,680 --> 00:03:29,670
No.

54
00:03:29,910 --> 00:03:31,210
Is one of them one.

55
00:03:31,500 --> 00:03:32,670
Yes.

56
00:03:32,760 --> 00:03:33,990
So let's try this.

57
00:03:33,990 --> 00:03:45,400
A two will make a variable a to reshape Eagles a to reshape 2 3 1 Beautiful.

58
00:03:45,720 --> 00:03:57,130
So we'll have a look at this a to reshape there's our reshaped a to a to reshape times a 3 There we

59
00:03:57,130 --> 00:03:57,540
go.

60
00:03:57,580 --> 00:03:58,630
Beautiful.

61
00:03:58,810 --> 00:04:01,390
So that is the benefit of reshape.

62
00:04:01,540 --> 00:04:03,760
Now this is only a small example here.

63
00:04:03,820 --> 00:04:11,710
Before they were incompatible but because we reshaped it we reshaped A2 to be within the broadcasting

64
00:04:11,710 --> 00:04:13,350
rules of num pie.

65
00:04:13,390 --> 00:04:18,700
We could then use the multiplier function across these two arrays.

66
00:04:18,700 --> 00:04:23,040
Now if this concept is kind of strange to begin with that's perfectly fine.

67
00:04:23,080 --> 00:04:23,980
And it will be.

68
00:04:24,100 --> 00:04:29,620
It will take a fair bit of practice to figure out how to use reshape properly but what the important

69
00:04:29,620 --> 00:04:37,120
takeaway here is is that just because your name pie array comes in a certain shape like it begins like

70
00:04:37,120 --> 00:04:45,850
this doesn't mean it has to stay in that shape so this reshaped version of A2 still contains the same

71
00:04:45,850 --> 00:04:48,610
information as the original A2.

72
00:04:48,640 --> 00:04:50,300
It's just in a different shape.

73
00:04:50,410 --> 00:04:55,540
So this is one of the problems you'll come across in many different machine learning problems is making

74
00:04:55,540 --> 00:05:01,330
sure that the data that you're trying to find patterns within is the same shape.

75
00:05:01,390 --> 00:05:08,170
So the machine learning algorithm can run over it like this broadcasting operation as quickly as possible.

76
00:05:08,290 --> 00:05:15,100
And if they aren't in the same shape you'll often get back errors like this which say the data you're

77
00:05:15,100 --> 00:05:17,760
trying to find patterns in isn't the same shape.

78
00:05:17,830 --> 00:05:24,010
So that's when you have to figure out how to use something like reshape to turn them into the same shape

79
00:05:24,370 --> 00:05:26,530
so that they're compatible with each other.

80
00:05:27,560 --> 00:05:35,370
Now there's one more thing which is transpose now what transpose is let's have a look at it first.

81
00:05:35,720 --> 00:05:38,210
Let's put a little comment here so we know what's happening.

82
00:05:38,210 --> 00:05:41,730
Transpose A2.

83
00:05:41,780 --> 00:05:43,620
Actually we might leave it here.

84
00:05:43,660 --> 00:05:44,800
A2 Doc T.

85
00:05:44,810 --> 00:05:48,260
So we can see what's changed here.

86
00:05:48,260 --> 00:05:53,300
Let's get A2 back up A2 there transpose what's changed.

87
00:05:53,300 --> 00:05:57,160
Let's go dot shape 3 2.

88
00:05:57,290 --> 00:05:57,630
Okay.

89
00:05:57,630 --> 00:06:01,070
Now if we go here Hey to shape.

90
00:06:01,410 --> 00:06:09,310
So what a transpose does it switches the accesses.

91
00:06:09,370 --> 00:06:13,410
So now A2 shape was 2 3.

92
00:06:13,510 --> 00:06:18,070
But if we call transpose and now t is 4 transpose it's going to swap.

93
00:06:18,100 --> 00:06:19,530
The axis is around.

94
00:06:19,720 --> 00:06:21,920
That's where we can see that now.

95
00:06:21,940 --> 00:06:28,220
One two three point three is down here in the column and four five six point five.

96
00:06:28,310 --> 00:06:29,950
It is on this column.

97
00:06:30,220 --> 00:06:33,490
So what it's done is essentially just swap the rows and columns.

98
00:06:33,490 --> 00:06:35,890
Now you can do this on a larger array as well.

99
00:06:35,920 --> 00:06:41,970
Let's view A3 A3 not shape 2 3 3.

100
00:06:41,980 --> 00:06:51,080
So if we call a three dot t it's going to swap these axes around so let's say a three dot teed up shape.

101
00:06:51,130 --> 00:06:55,130
So now we've essentially flipped it.

102
00:06:55,150 --> 00:07:00,590
So now this dimension so too is right at the end.

103
00:07:00,780 --> 00:07:04,530
And 3 and 3 have swapped to the front.

104
00:07:04,530 --> 00:07:10,860
The important difference here between transpose and reshape is that transpose just flips the accesses

105
00:07:10,860 --> 00:07:12,960
around and reshape.

106
00:07:12,960 --> 00:07:15,810
You can create your own custom shapes.

107
00:07:15,810 --> 00:07:21,780
So for example we reshaped A2 to be compatible with our broadcasting rules.

108
00:07:21,950 --> 00:07:24,370
So have a practice reshaping some of your arrays.

109
00:07:24,420 --> 00:07:26,070
And I'll see you in the next video.