# 2.1. Data Manipulation¶

It is impossible to get anything done if we cannot manipulate data. Generally, there are two important things we need to do with data: (i) acquire it and (ii) process it once it is inside the computer. There is no point in acquiring data if we do not even know how to store it, so let’s get our hands dirty first by playing with synthetic data. We will start by introducing the NDArray, MXNet’s primary tool for storing and transforming data. If you have worked with NumPy before, you will notice that NDArrays are, by design, similar to NumPy’s multi-dimensional array. However, they confer a few key advantages. First, NDArrays support asynchronous computation on CPU, GPU, and distributed cloud architectures. Second, they provide support for automatic differentiation. These properties make NDArray indispensable for deep learning.

## 2.1.1. Getting Started¶

Throughout this chapter, we are aiming to get you up and running with the basic functionality. Do not worry if you do not understand all of the basic math, like element-wise operations or normal distributions. In the next two chapters we will take another pass at the same material, teaching the material in the context of practical examples. On the other hand, if you want to go deeper into the mathematical content, see Section 14.2.

We begin by importing MXNet and the `ndarray`

module from MXNet. Here,
`nd`

is short for `ndarray`

.

```
import mxnet as mx
from mxnet import nd
```

NDArrays represent (possibly multi-dimensional) arrays of numerical
values. NDArrays with one axis correspond (in math-speak) to *vectors*.
NDArrays with two axes correspond to *matrices*. For arrays with more
than two axes, mathematicians do not have special names—they simply call
them *tensors*.

The simplest object we can create is a vector. To start, we can use
`arange`

to create a row vector with 12 consecutive integers.

```
x = nd.arange(12)
x
```

```
[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.]
<NDArray 12 @cpu(0)>
```

When we print `x`

, we can observe the property
`<NDArray 12 @cpu(0)>`

listed, which indicates that `x`

is a
one-dimensional array of length 12 and that it resides in CPU main
memory. The 0 in `@cpu(0)`

has no special meaning and does not
represent a specific core.

We can get the NDArray instance shape through the `shape`

property.

```
x.shape
```

```
(12,)
```

We can also get the total number of elements in the NDArray instance
through the `size`

property. This is the product of the elements of
the shape. Since we are dealing with a vector here, both are identical.

```
x.size
```

```
12
```

We use the `reshape`

function to change the shape of one (possibly
multi-dimensional) array, to another that contains the same number of
elements. For example, we can transform the shape of our line vector
`x`

to (3, 4), which contains the same values but interprets them as a
matrix containing 3 rows and 4 columns. Note that although the shape has
changed, the elements in `x`

have not. Moreover, the `size`

remains
the same.

```
x = x.reshape((3, 4))
x
```

```
[[ 0. 1. 2. 3.]
[ 4. 5. 6. 7.]
[ 8. 9. 10. 11.]]
<NDArray 3x4 @cpu(0)>
```

Reshaping by manually specifying each of the dimensions can get
annoying. Once we know one of the dimensions, why should we have to
perform the division our selves to determine the other? For example,
above, to get a matrix with 3 rows, we had to specify that it should
have 4 columns (to account for the 12 elements). Fortunately, NDArray
can automatically work out one dimension given the other. We can invoke
this capability by placing `-1`

for the dimension that we would like
NDArray to automatically infer. In our case, instead of
`x.reshape((3, 4))`

, we could have equivalently used
`x.reshape((-1, 4))`

or `x.reshape((3, -1))`

.

```
nd.empty((3, 4))
```

```
[[-3.9202182e+27 4.5563220e-41 3.7803224e-23 3.0824362e-41]
[ 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]
[ 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]]
<NDArray 3x4 @cpu(0)>
```

The `empty`

method just grabs some memory and hands us back a matrix
without setting the values of any of its entries. This is very efficient
but it means that the entries might take any arbitrary values, including
very big ones! Typically, we’ll want our matrices initialized either
with ones, zeros, some known constant or numbers randomly sampled from a
known distribution.

Perhaps most often, we want an array of all zeros. To create an NDArray representing a tensor with all elements set to 0 and a shape of (2, 3, 4) we can invoke:

```
nd.zeros((2, 3, 4))
```

```
[[[0. 0. 0. 0.]
[0. 0. 0. 0.]
[0. 0. 0. 0.]]
[[0. 0. 0. 0.]
[0. 0. 0. 0.]
[0. 0. 0. 0.]]]
<NDArray 2x3x4 @cpu(0)>
```

We can create tensors with each element set to 1 works via

```
nd.ones((2, 3, 4))
```

```
[[[1. 1. 1. 1.]
[1. 1. 1. 1.]
[1. 1. 1. 1.]]
[[1. 1. 1. 1.]
[1. 1. 1. 1.]
[1. 1. 1. 1.]]]
<NDArray 2x3x4 @cpu(0)>
```

We can also specify the value of each element in the desired NDArray by supplying a Python list containing the numerical values.

```
y = nd.array([[2, 1, 4, 3], [1, 2, 3, 4], [4, 3, 2, 1]])
y
```

```
[[2. 1. 4. 3.]
[1. 2. 3. 4.]
[4. 3. 2. 1.]]
<NDArray 3x4 @cpu(0)>
```

In some cases, we will want to randomly sample the values of each element in the NDArray according to some known probability distribution. This is especially common when we intend to use the array as a parameter in a neural network. The following snippet creates an NDArray with a shape of (3,4). Each of its elements is randomly sampled in a normal distribution with zero mean and unit variance.

```
nd.random.normal(0, 1, shape=(3, 4))
```

```
[[ 2.2122064 0.7740038 1.0434405 1.1839255 ]
[ 1.8917114 -1.2347414 -1.771029 -0.45138445]
[ 0.57938355 -1.856082 -1.9768796 -0.20801921]]
<NDArray 3x4 @cpu(0)>
```

## 2.1.2. Operations¶

Oftentimes, we want to apply functions to arrays. Some of the simplest
and most useful functions are the element-wise functions. These operate
by performing a single scalar operation on the corresponding elements of
two arrays. We can create an element-wise function from any function
that maps from the scalars to the scalars. In math notations we would
denote such a function as \(f: \mathbb{R} \rightarrow \mathbb{R}\).
Given any two vectors \(\mathbf{u}\) and \(\mathbf{v}\) *of the
same shape*, and the function f, we can produce a vector
\(\mathbf{c} = F(\mathbf{u},\mathbf{v})\) by setting
\(c_i \gets f(u_i, v_i)\) for all \(i\). Here, we produced the
vector-valued \(F: \mathbb{R}^d \rightarrow \mathbb{R}^d\) by
*lifting* the scalar function to an element-wise vector operation. In
MXNet, the common standard arithmetic operators (+,-,/,*,**) have all
been *lifted* to element-wise operations for identically-shaped tensors
of arbitrary shape. We can call element-wise operations on any two
tensors of the same shape, including matrices.

```
x = nd.array([1, 2, 4, 8])
y = nd.ones_like(x) * 2
print('x =', x)
print('x + y', x + y)
print('x - y', x - y)
print('x * y', x * y)
print('x / y', x / y)
```

```
x =
[1. 2. 4. 8.]
<NDArray 4 @cpu(0)>
x + y
[ 3. 4. 6. 10.]
<NDArray 4 @cpu(0)>
x - y
[-1. 0. 2. 6.]
<NDArray 4 @cpu(0)>
x * y
[ 2. 4. 8. 16.]
<NDArray 4 @cpu(0)>
x / y
[0.5 1. 2. 4. ]
<NDArray 4 @cpu(0)>
```

Many more operations can be applied element-wise, such as exponentiation:

```
x.exp()
```

```
[2.7182817e+00 7.3890562e+00 5.4598148e+01 2.9809580e+03]
<NDArray 4 @cpu(0)>
```

In addition to computations by element, we can also perform matrix
operations, like matrix multiplication using the `dot`

function. Next,
we will perform matrix multiplication of `x`

and the transpose of
`y`

. We define `x`

as a matrix of 3 rows and 4 columns, and `y`

is
transposed into a matrix of 4 rows and 3 columns. The two matrices are
multiplied to obtain a matrix of 3 rows and 3 columns (if you are
confused about what this means, do not worry - we will explain matrix
operations in much more detail in Section 2.2).

```
x = nd.arange(12).reshape((3,4))
y = nd.array([[2, 1, 4, 3], [1, 2, 3, 4], [4, 3, 2, 1]])
nd.dot(x, y.T)
```

```
[[ 18. 20. 10.]
[ 58. 60. 50.]
[ 98. 100. 90.]]
<NDArray 3x3 @cpu(0)>
```

We can also merge multiple NDArrays. For that, we need to tell the system along which dimension to merge. The example below merges two matrices along dimension 0 (along rows) and dimension 1 (along columns) respectively.

```
nd.concat(x, y, dim=0)
nd.concat(x, y, dim=1)
```

```
[[ 0. 1. 2. 3. 2. 1. 4. 3.]
[ 4. 5. 6. 7. 1. 2. 3. 4.]
[ 8. 9. 10. 11. 4. 3. 2. 1.]]
<NDArray 3x8 @cpu(0)>
```

Sometimes, we may want to construct binary NDArrays via logical
statements. Take `x == y`

as an example. If `x`

and `y`

are equal
for some entry, the new NDArray has a value of 1 at the same position;
otherwise it is 0.

```
x == y
```

```
[[0. 1. 0. 1.]
[0. 0. 0. 0.]
[0. 0. 0. 0.]]
<NDArray 3x4 @cpu(0)>
```

Summing all the elements in the NDArray yields an NDArray with only one element.

```
x.sum()
```

```
[66.]
<NDArray 1 @cpu(0)>
```

We can transform the result into a scalar in Python using the
`asscalar`

function. In the following example, the \(\ell_2\) norm
of `x`

yields a single element NDArray. The final result is
transformed into a scalar.

```
x.norm().asscalar()
```

```
22.494442
```

For stylistic convenience, we can write `y.exp()`

, `x.sum()`

,
`x.norm()`

, etc. also as `nd.exp(y)`

, `nd.sum(x)`

, `nd.norm(x)`

.

## 2.1.3. Broadcast Mechanism¶

In the above section, we saw how to perform operations on two NDArrays of the same shape. When their shapes differ, a broadcasting mechanism may be triggered analogous to NumPy: first, copy the elements appropriately so that the two NDArrays have the same shape, and then carry out operations by element.

```
a = nd.arange(3).reshape((3, 1))
b = nd.arange(2).reshape((1, 2))
a, b
```

```
(
[[0.]
[1.]
[2.]]
<NDArray 3x1 @cpu(0)>,
[[0. 1.]]
<NDArray 1x2 @cpu(0)>)
```

Since `a`

and `b`

are (3x1) and (1x2) matrices respectively, their
shapes do not match up if we want to add them. NDArray addresses this by
‘broadcasting’ the entries of both matrices into a larger (3x2) matrix
as follows: for matrix `a`

it replicates the columns, for matrix `b`

it replicates the rows before adding up both element-wise.

```
a + b
```

```
[[0. 1.]
[1. 2.]
[2. 3.]]
<NDArray 3x2 @cpu(0)>
```

## 2.1.4. Indexing and Slicing¶

Just like in any other Python array, elements in an NDArray can be
accessed by its index. In good Python tradition the first element has
index 0 and ranges are specified to include the first but not the last
element. By this logic `1:3`

selects the second and third element.
Let’s try this out by selecting the respective rows in a matrix.

```
x[1:3]
```

```
[[ 4. 5. 6. 7.]
[ 8. 9. 10. 11.]]
<NDArray 2x4 @cpu(0)>
```

Beyond reading, we can also write elements of a matrix.

```
x[1, 2] = 9
x
```

```
[[ 0. 1. 2. 3.]
[ 4. 5. 9. 7.]
[ 8. 9. 10. 11.]]
<NDArray 3x4 @cpu(0)>
```

If we want to assign multiple elements the same value, we simply index
all of them and then assign them the value. For instance, `[0:2, :]`

accesses the first and second rows. While we discussed indexing for
matrices, this obviously also works for vectors and for tensors of more
than 2 dimensions.

```
x[0:2, :] = 12
x
```

```
[[12. 12. 12. 12.]
[12. 12. 12. 12.]
[ 8. 9. 10. 11.]]
<NDArray 3x4 @cpu(0)>
```

## 2.1.5. Saving Memory¶

In the previous example, every time we ran an operation, we allocated
new memory to host its results. For example, if we write `y = x + y`

,
we will dereference the matrix that `y`

used to point to and instead
point it at the newly allocated memory. In the following example we
demonstrate this with Python’s `id()`

function, which gives us the
exact address of the referenced object in memory. After running
`y = y + x`

, we will find that `id(y)`

points to a different
location. That is because Python first evaluates `y + x`

, allocating
new memory for the result and then subsequently redirects `y`

to point
at this new location in memory.

```
before = id(y)
y = y + x
id(y) == before
```

```
False
```

This might be undesirable for two reasons. First, we do not want to run
around allocating memory unnecessarily all the time. In machine
learning, we might have hundreds of megabytes of parameters and update
all of them multiple times per second. Typically, we will want to
perform these updates *in place*. Second, we might point at the same
parameters from multiple variables. If we do not update in place, this
could cause a memory leak, making it possible for us to inadvertently
reference stale parameters.

Fortunately, performing in-place operations in MXNet is easy. We can
assign the result of an operation to a previously allocated array with
slice notation, e.g., `y[:] = <expression>`

. To illustrate the
behavior, we first clone the shape of a matrix using `zeros_like`

to
allocate a block of 0 entries.

```
z = y.zeros_like()
print('id(z):', id(z))
z[:] = x + y
print('id(z):', id(z))
```

```
id(z): 139650460787320
id(z): 139650460787320
```

While this looks pretty, `x+y`

here will still allocate a temporary
buffer to store the result of `x+y`

before copying it to `z[:]`

. To
make even better use of memory, we can directly invoke the underlying
`ndarray`

operation, in this case `elemwise_add`

, avoiding temporary
buffers. We do this by specifying the `out`

keyword argument, which
every `ndarray`

operator supports:

```
before = id(z)
nd.elemwise_add(x, y, out=z)
id(z) == before
```

```
True
```

If the value of `x`

is not reused in subsequent computations, we can
also use `x[:] = x + y`

or `x += y`

to reduce the memory overhead of
the operation.

```
before = id(x)
x += y
id(x) == before
```

```
True
```

## 2.1.6. Mutual Transformation of NDArray and NumPy¶

Converting MXNet NDArrays to and from NumPy is easy. The converted
arrays do *not* share memory. This minor inconvenience is actually quite
important: when you perform operations on the CPU or one of the GPUs,
you do not want MXNet having to wait whether NumPy might want to be
doing something else with the same chunk of memory. The `array`

and
`asnumpy`

functions do the trick.

```
import numpy as np
a = x.asnumpy()
print(type(a))
b = nd.array(a)
print(type(b))
```

```
<class 'numpy.ndarray'>
<class 'mxnet.ndarray.ndarray.NDArray'>
```

## 2.1.7. Exercises¶

Run the code in this section. Change the conditional statement

`x == y`

in this section to`x < y`

or`x > y`

, and then see what kind of NDArray you can get.Replace the two NDArrays that operate by element in the broadcast mechanism with other shapes, e.g. three dimensional tensors. Is the result the same as expected?

Assume that we have three matrices

`a`

,`b`

and`c`

. Rewrite`c = nd.dot(a, b.T) + c`

in the most memory efficient manner.