Tools - NumPy

NumPy is the fundamental library for scientific computing with Python. NumPy is centered around a powerful N-dimensional array object, and it also contains useful linear algebra, Fourier transform, and random number functions.

Creating arrays

Now let's import numpy. Most people import it as np:

In [2]:
import numpy as np

np.zeros

The zeros function creates an array containing any number of zeros:

In [3]:
np.zeros(5)
Out[3]:
array([ 0.,  0.,  0.,  0.,  0.])

It's just as easy to create a 2D array (ie. a matrix) by providing a tuple with the desired number of rows and columns. For example, here's a 3x4 matrix:

In [4]:
np.zeros((3,4))
Out[4]:
array([[ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.]])

Some vocabulary

  • In NumPy, each dimension is called an axis.
  • The number of axes is called the rank.
    • For example, the above 3x4 matrix is an array of rank 2 (it is 2-dimensional).
    • The first axis has length 3, the second has length 4.
  • An array's list of axis lengths is called the shape of the array.
    • For example, the above matrix's shape is (3, 4).
    • The rank is equal to the shape's length.
  • The size of an array is the total number of elements, which is the product of all axis lengths (eg. 3*4=12)
In [5]:
a = np.zeros((3,4))
a
Out[5]:
array([[ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.]])
In [6]:
a.shape
Out[6]:
(3, 4)
In [7]:
a.ndim  # equal to len(a.shape)
Out[7]:
2
In [8]:
a.size
Out[8]:
12

N-dimensional arrays

You can also create an N-dimensional array of arbitrary rank. For example, here's a 3D array (rank=3), with shape (2,3,4):

In [9]:
np.zeros((2,3,4))
Out[9]:
array([[[ 0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.]],

       [[ 0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.]]])

Array type

NumPy arrays have the type ndarrays:

In [10]:
type(np.zeros((3,4)))
Out[10]:
numpy.ndarray

np.ones

Many other NumPy functions create ndarrays.

Here's a 3x4 matrix full of ones:

In [11]:
np.ones((3,4))
Out[11]:
array([[ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.]])

np.full

Creates an array of the given shape initialized with the given value. Here's a 3x4 matrix full of π.

In [12]:
np.full((3,4), np.pi)
Out[12]:
array([[ 3.14159265,  3.14159265,  3.14159265,  3.14159265],
       [ 3.14159265,  3.14159265,  3.14159265,  3.14159265],
       [ 3.14159265,  3.14159265,  3.14159265,  3.14159265]])

np.empty

An uninitialized 2x3 array (its content is not predictable, as it is whatever is in memory at that point):

In [13]:
np.empty((2,3))
Out[13]:
array([[  2.68156159e+154,  -2.32035135e+077,   2.22457274e-314],
       [  2.24336635e-314,   2.23817082e-314,   4.17203587e-309]])

np.array

Of course you can initialize an ndarray using a regular python array. Just call the array function:

In [14]:
np.array([[1,2,3,4], [10, 20, 30, 40]])
Out[14]:
array([[ 1,  2,  3,  4],
       [10, 20, 30, 40]])

np.arange

You can create an ndarray using NumPy's arange function, which is similar to python's built-in range function:

In [15]:
np.arange(1, 5)
Out[15]:
array([1, 2, 3, 4])

It also works with floats:

In [16]:
np.arange(1.0, 5.0)
Out[16]:
array([ 1.,  2.,  3.,  4.])

Of course you can provide a step parameter:

In [17]:
np.arange(1, 5, 0.5)
Out[17]:
array([ 1. ,  1.5,  2. ,  2.5,  3. ,  3.5,  4. ,  4.5])

However, when dealing with floats, the exact number of elements in the array is not always predictible. For example, consider this:

In [18]:
print(np.arange(0, 5/3, 1/3)) # depending on floating point errors, the max value is 4/3 or 5/3.
print(np.arange(0, 5/3, 0.333333333))
print(np.arange(0, 5/3, 0.333333334))
[ 0.          0.33333333  0.66666667  1.          1.33333333  1.66666667]
[ 0.          0.33333333  0.66666667  1.          1.33333333  1.66666667]
[ 0.          0.33333333  0.66666667  1.          1.33333334]

np.linspace

For this reason, it is generally preferable to use the linspace function instead of arange when working with floats. The linspace function returns an array containing a specific number of points evenly distributed between two values (note that the maximum value is included, contrary to arange):

In [19]:
print(np.linspace(0, 5/3, 6))
[ 0.          0.33333333  0.66666667  1.          1.33333333  1.66666667]

np.rand and np.randn

A number of functions are available in NumPy's random module to create ndarrays initialized with random values. For example, here is a 3x4 matrix initialized with random floats between 0 and 1 (uniform distribution):

In [20]:
np.random.rand(3,4)
Out[20]:
array([[ 0.37454012,  0.95071431,  0.73199394,  0.59865848],
       [ 0.15601864,  0.15599452,  0.05808361,  0.86617615],
       [ 0.60111501,  0.70807258,  0.02058449,  0.96990985]])

Here's a 3x4 matrix containing random floats sampled from a univariate normal distribution (Gaussian distribution) of mean 0 and variance 1:

In [21]:
np.random.randn(3,4)
Out[21]:
array([[-0.46947439,  0.54256004, -0.46341769, -0.46572975],
       [ 0.24196227, -1.91328024, -1.72491783, -0.56228753],
       [-1.01283112,  0.31424733, -0.90802408, -1.4123037 ]])

To give you a feel of what these distributions look like, let's use matplotlib (see the matplotlib tutorial for more details):

In [22]:
%matplotlib inline
import matplotlib.pyplot as plt
In [23]:
plt.hist(np.random.rand(100000), density=True, bins=100, histtype="step", color="blue", label="rand")
plt.hist(np.random.randn(100000), density=True, bins=100, histtype="step", color="red", label="randn")
plt.axis([-2.5, 2.5, 0, 1.1])
plt.legend(loc = "upper left")
plt.title("Random distributions")
plt.xlabel("Value")
plt.ylabel("Density")
plt.show()

np.fromfunction

You can also initialize an ndarray using a function:

In [24]:
def my_function(z, y, x):
    return x * y + z

np.fromfunction(my_function, (3, 2, 10))
Out[24]:
array([[[  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.],
        [  0.,   1.,   2.,   3.,   4.,   5.,   6.,   7.,   8.,   9.]],

       [[  1.,   1.,   1.,   1.,   1.,   1.,   1.,   1.,   1.,   1.],
        [  1.,   2.,   3.,   4.,   5.,   6.,   7.,   8.,   9.,  10.]],

       [[  2.,   2.,   2.,   2.,   2.,   2.,   2.,   2.,   2.,   2.],
        [  2.,   3.,   4.,   5.,   6.,   7.,   8.,   9.,  10.,  11.]]])

NumPy first creates three ndarrays (one per dimension), each of shape (2, 10). Each array has values equal to the coordinate along a specific axis. For example, all elements in the z array are equal to their z-coordinate:

[[[ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]]

 [[ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]
  [ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]]

 [[ 2.  2.  2.  2.  2.  2.  2.  2.  2.  2.]
  [ 2.  2.  2.  2.  2.  2.  2.  2.  2.  2.]]]

So the terms x, y and z in the expression x * y + z above are in fact ndarrays (we will discuss arithmetic operations on arrays below). The point is that the function my_function is only called once, instead of once per element. This makes initialization very efficient.

Array data

dtype

NumPy's ndarrays are also efficient in part because all their elements must have the same type (usually numbers). You can check what the data type is by looking at the dtype attribute:

In [25]:
c = np.arange(1, 5)
print(c.dtype, c)
int64 [1 2 3 4]
In [26]:
c = np.arange(1.0, 5.0)
print(c.dtype, c)
float64 [ 1.  2.  3.  4.]

Instead of letting NumPy guess what data type to use, you can set it explicitly when creating an array by setting the dtype parameter:

In [27]:
d = np.arange(1, 5, dtype=np.complex64)
print(d.dtype, d)
complex64 [ 1.+0.j  2.+0.j  3.+0.j  4.+0.j]

Available data types include int8, int16, int32, int64, uint8|16|32|64, float16|32|64 and complex64|128. Check out the documentation for the full list.

itemsize

The itemsize attribute returns the size (in bytes) of each item:

In [28]:
e = np.arange(1, 5, dtype=np.complex64)
e.itemsize
Out[28]:
8

data buffer

An array's data is actually stored in memory as a flat (one dimensional) byte buffer. It is available via the data attribute (you will rarely need it, though).

In [29]:
f = np.array([[1,2],[1000, 2000]], dtype=np.int32)
f.data
Out[29]:
<read-write buffer for 0x10f8a18a0, size 16, offset 0 at 0x10f9dbbb0>

In python 2, f.data is a buffer. In python 3, it is a memoryview.

In [30]:
if (hasattr(f.data, "tobytes")):
    data_bytes = f.data.tobytes() # python 3
else:
    data_bytes = memoryview(f.data).tobytes() # python 2

data_bytes
Out[30]:
'\x01\x00\x00\x00\x02\x00\x00\x00\xe8\x03\x00\x00\xd0\x07\x00\x00'

Several ndarrays can share the same data buffer, meaning that modifying one will also modify the others. We will see an example in a minute.

Reshaping an array

In place

Changing the shape of an ndarray is as simple as setting its shape attribute. However, the array's size must remain the same.

In [31]:
g = np.arange(24)
print(g)
print("Rank:", g.ndim)
[ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23]
Rank: 1
In [32]:
g.shape = (6, 4)
print(g)
print("Rank:", g.ndim)
[[ 0  1  2  3]
 [ 4  5  6  7]
 [ 8  9 10 11]
 [12 13 14 15]
 [16 17 18 19]
 [20 21 22 23]]
Rank: 2
In [33]:
g.shape = (2, 3, 4)
print(g)
print("Rank:", g.ndim)
[[[ 0  1  2  3]
  [ 4  5  6  7]
  [ 8  9 10 11]]

 [[12 13 14 15]
  [16 17 18 19]
  [20 21 22 23]]]
Rank: 3

reshape

The reshape function returns a new ndarray object pointing at the same data. This means that modifying one array will also modify the other.

In [34]:
g2 = g.reshape(4,6)
print(g2)
print("Rank:", g2.ndim)
[[ 0  1  2  3  4  5]
 [ 6  7  8  9 10 11]
 [12 13 14 15 16 17]
 [18 19 20 21 22 23]]
Rank: 2

Set item at row 1, col 2 to 999 (more about indexing below).

In [35]:
g2[1, 2] = 999
g2
Out[35]:
array([[  0,   1,   2,   3,   4,   5],
       [  6,   7, 999,   9,  10,  11],
       [ 12,  13,  14,  15,  16,  17],
       [ 18,  19,  20,  21,  22,  23]])

The corresponding element in g has been modified.

In [36]:
g
Out[36]:
array([[[  0,   1,   2,   3],
        [  4,   5,   6,   7],
        [999,   9,  10,  11]],

       [[ 12,  13,  14,  15],
        [ 16,  17,  18,  19],
        [ 20,  21,  22,  23]]])

ravel

Finally, the ravel function returns a new one-dimensional ndarray that also points to the same data:

In [37]:
g.ravel()
Out[37]:
array([  0,   1,   2,   3,   4,   5,   6,   7, 999,   9,  10,  11,  12,
        13,  14,  15,  16,  17,  18,  19,  20,  21,  22,  23])

Arithmetic operations

All the usual arithmetic operators (+, -, *, /, //, **, etc.) can be used with ndarrays. They apply elementwise:

In [38]:
a = np.array([14, 23, 32, 41])
b = np.array([5,  4,  3,  2])
print("a + b  =", a + b)
print("a - b  =", a - b)
print("a * b  =", a * b)
print("a / b  =", a / b)
print("a // b  =", a // b)
print("a % b  =", a % b)
print("a ** b =", a ** b)
a + b  = [19 27 35 43]
a - b  = [ 9 19 29 39]
a * b  = [70 92 96 82]
a / b  = [  2.8          5.75        10.66666667  20.5       ]
a // b  = [ 2  5 10 20]
a % b  = [4 3 2 1]
a ** b = [537824 279841  32768   1681]

Note that the multiplication is not a matrix multiplication. We will discuss matrix operations below.

The arrays must have the same shape. If they do not, NumPy will apply the broadcasting rules.

Broadcasting

In general, when NumPy expects arrays of the same shape but finds that this is not the case, it applies the so-called broadcasting rules:

First rule

If the arrays do not have the same rank, then a 1 will be prepended to the smaller ranking arrays until their ranks match.

In [39]:
h = np.arange(5).reshape(1, 1, 5)
h
Out[39]:
array([[[0, 1, 2, 3, 4]]])

Now let's try to add a 1D array of shape (5,) to this 3D array of shape (1,1,5). Applying the first rule of broadcasting!

In [40]:
h + [10, 20, 30, 40, 50]  # same as: h + [[[10, 20, 30, 40, 50]]]
Out[40]:
array([[[10, 21, 32, 43, 54]]])

Second rule

Arrays with a 1 along a particular dimension act as if they had the size of the array with the largest shape along that dimension. The value of the array element is repeated along that dimension.

In [41]:
k = np.arange(6).reshape(2, 3)
k
Out[41]:
array([[0, 1, 2],
       [3, 4, 5]])

Let's try to add a 2D array of shape (2,1) to this 2D ndarray of shape (2, 3). NumPy will apply the second rule of broadcasting:

In [42]:
k + [[100], [200]]  # same as: k + [[100, 100, 100], [200, 200, 200]]
Out[42]:
array([[100, 101, 102],
       [203, 204, 205]])

Combining rules 1 & 2, we can do this:

In [43]:
k + [100, 200, 300]  # after rule 1: [[100, 200, 300]], and after rule 2: [[100, 200, 300], [100, 200, 300]]
Out[43]:
array([[100, 201, 302],
       [103, 204, 305]])

And also, very simply:

In [44]:
k + 1000  # same as: k + [[1000, 1000, 1000], [1000, 1000, 1000]]
Out[44]:
array([[1000, 1001, 1002],
       [1003, 1004, 1005]])

Third rule

After rules 1 & 2, the sizes of all arrays must match.

In [45]:
try:
    k + [33, 44]
except ValueError as e:
    print(e)
operands could not be broadcast together with shapes (2,3) (2,) 

Broadcasting rules are used in many NumPy operations, not just arithmetic operations, as we will see below. For more details about broadcasting, check out the documentation.

Upcasting

When trying to combine arrays with different dtypes, NumPy will upcast to a type capable of handling all possible values (regardless of what the actual values are).

In [46]:
k1 = np.arange(0, 5, dtype=np.uint8)
print(k1.dtype, k1)
uint8 [0 1 2 3 4]
In [47]:
k2 = k1 + np.array([5, 6, 7, 8, 9], dtype=np.int8)
print(k2.dtype, k2)
int16 [ 5  7  9 11 13]

Note that int16 is required to represent all possible int8 and uint8 values (from -128 to 255), even though in this case a uint8 would have sufficed.

In [48]:
k3 = k1 + 1.5
print(k3.dtype, k3)
float64 [ 1.5  2.5  3.5  4.5  5.5]

Conditional operators

The conditional operators also apply elementwise:

In [49]:
m = np.array([20, -5, 30, 40])
m < [15, 16, 35, 36]
Out[49]:
array([False,  True,  True, False], dtype=bool)

And using broadcasting:

In [50]:
m < 25  # equivalent to m < [25, 25, 25, 25]
Out[50]:
array([ True,  True, False, False], dtype=bool)

This is most useful in conjunction with boolean indexing (discussed below).

In [51]:
m[m < 25]
Out[51]:
array([20, -5])

Mathematical and statistical functions

Many mathematical and statistical functions are available for ndarrays.

ndarray methods

Some functions are simply ndarray methods, for example:

In [52]:
a = np.array([[-2.5, 3.1, 7], [10, 11, 12]])
print(a)
print("mean =", a.mean())
[[ -2.5   3.1   7. ]
 [ 10.   11.   12. ]]
mean = 6.76666666667

Note that this computes the mean of all elements in the ndarray, regardless of its shape.

Here are a few more useful ndarray methods:

In [53]:
for func in (a.min, a.max, a.sum, a.prod, a.std, a.var):
    print(func.__name__, "=", func())
min = -2.5
max = 12.0
sum = 40.6
prod = -71610.0
std = 5.08483584352
var = 25.8555555556

These functions accept an optional argument axis which lets you ask for the operation to be performed on elements along the given axis. For example:

In [54]:
c=np.arange(24).reshape(2,3,4)
c
Out[54]:
array([[[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11]],

       [[12, 13, 14, 15],
        [16, 17, 18, 19],
        [20, 21, 22, 23]]])
In [55]:
c.sum(axis=0)  # sum across matrices
Out[55]:
array([[12, 14, 16, 18],
       [20, 22, 24, 26],
       [28, 30, 32, 34]])
In [56]:
c.sum(axis=1)  # sum across rows
Out[56]:
array([[12, 15, 18, 21],
       [48, 51, 54, 57]])

You can also sum over multiple axes:

In [57]:
c.sum(axis=(0,2))  # sum across matrices and columns
Out[57]:
array([ 60,  92, 124])
In [58]:
0+1+2+3 + 12+13+14+15, 4+5+6+7 + 16+17+18+19, 8+9+10+11 + 20+21+22+23
Out[58]:
(60, 92, 124)

Universal functions

NumPy also provides fast elementwise functions called universal functions, or ufunc. They are vectorized wrappers of simple functions. For example square returns a new ndarray which is a copy of the original ndarray except that each element is squared:

In [59]:
a = np.array([[-2.5, 3.1, 7], [10, 11, 12]])
np.square(a)
Out[59]:
array([[   6.25,    9.61,   49.  ],
       [ 100.  ,  121.  ,  144.  ]])

Here are a few more useful unary ufuncs:

In [60]:
print("Original ndarray")
print(a)
for func in (np.abs, np.sqrt, np.exp, np.log, np.sign, np.ceil, np.modf, np.isnan, np.cos):
    print("\n", func.__name__)
    print(func(a))
Original ndarray
[[ -2.5   3.1   7. ]
 [ 10.   11.   12. ]]

 absolute
[[  2.5   3.1   7. ]
 [ 10.   11.   12. ]]

 sqrt
[[        nan  1.76068169  2.64575131]
 [ 3.16227766  3.31662479  3.46410162]]

 exp
[[  8.20849986e-02   2.21979513e+01   1.09663316e+03]
 [  2.20264658e+04   5.98741417e+04   1.62754791e+05]]

 log
[[        nan  1.13140211  1.94591015]
 [ 2.30258509  2.39789527  2.48490665]]

 sign
[[-1.  1.  1.]
 [ 1.  1.  1.]]

 ceil
[[ -2.   4.   7.]
 [ 10.  11.  12.]]

 modf
(array([[-0.5,  0.1,  0. ],
       [ 0. ,  0. ,  0. ]]), array([[ -2.,   3.,   7.],
       [ 10.,  11.,  12.]]))

 isnan
[[False False False]
 [False False False]]

 cos
[[-0.80114362 -0.99913515  0.75390225]
 [-0.83907153  0.0044257   0.84385396]]
-c:5: RuntimeWarning: invalid value encountered in sqrt
-c:5: RuntimeWarning: invalid value encountered in log

Binary ufuncs

There are also many binary ufuncs, that apply elementwise on two ndarrays. Broadcasting rules are applied if the arrays do not have the same shape:

In [61]:
a = np.array([1, -2, 3, 4])
b = np.array([2, 8, -1, 7])
np.add(a, b)  # equivalent to a + b
Out[61]:
array([ 3,  6,  2, 11])
In [62]:
np.greater(a, b)  # equivalent to a > b
Out[62]:
array([False, False,  True, False], dtype=bool)
In [63]:
np.maximum(a, b)
Out[63]:
array([2, 8, 3, 7])
In [64]:
np.copysign(a, b)
Out[64]:
array([ 1.,  2., -3.,  4.])

Array indexing

One-dimensional arrays

One-dimensional NumPy arrays can be accessed more or less like regular python arrays:

In [65]:
a = np.array([1, 5, 3, 19, 13, 7, 3])
a[3]
Out[65]:
19
In [66]:
a[2:5]
Out[66]:
array([ 3, 19, 13])
In [67]:
a[2:-1]
Out[67]:
array([ 3, 19, 13,  7])
In [68]:
a[:2]
Out[68]:
array([1, 5])
In [69]:
a[2::2]
Out[69]:
array([ 3, 13,  3])
In [70]:
a[::-1]
Out[70]:
array([ 3,  7, 13, 19,  3,  5,  1])

Of course, you can modify elements:

In [71]:
a[3]=999
a
Out[71]:
array([  1,   5,   3, 999,  13,   7,   3])

You can also modify an ndarray slice:

In [72]:
a[2:5] = [997, 998, 999]
a
Out[72]:
array([  1,   5, 997, 998, 999,   7,   3])

Differences with regular python arrays

Contrary to regular python arrays, if you assign a single value to an ndarray slice, it is copied across the whole slice, thanks to broadcasting rules discussed above.

In [73]:
a[2:5] = -1
a
Out[73]:
array([ 1,  5, -1, -1, -1,  7,  3])

Also, you cannot grow or shrink ndarrays this way:

In [74]:
try:
    a[2:5] = [1,2,3,4,5,6]  # too long
except ValueError as e:
    print(e)
cannot copy sequence with size 6 to array axis with dimension 3

You cannot delete elements either:

In [75]:
try:
    del a[2:5]
except ValueError as e:
    print(e)
cannot delete array elements

Last but not least, ndarray slices are actually views on the same data buffer. This means that if you create a slice and modify it, you are actually going to modify the original ndarray as well!

In [76]:
a_slice = a[2:6]
a_slice[1] = 1000
a  # the original array was modified!
Out[76]:
array([   1,    5,   -1, 1000,   -1,    7,    3])
In [77]:
a[3] = 2000
a_slice  # similarly, modifying the original array modifies the slice!
Out[77]:
array([  -1, 2000,   -1,    7])

If you want a copy of the data, you need to use the copy method:

In [78]:
another_slice = a[2:6].copy()
another_slice[1] = 3000
a  # the original array is untouched
Out[78]:
array([   1,    5,   -1, 2000,   -1,    7,    3])
In [79]:
a[3] = 4000
another_slice  # similary, modifying the original array does not affect the slice copy
Out[79]:
array([  -1, 3000,   -1,    7])

Multi-dimensional arrays

Multi-dimensional arrays can be accessed in a similar way by providing an index or slice for each axis, separated by commas:

In [80]:
b = np.arange(48).reshape(4, 12)
b
Out[80]:
array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11],
       [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23],
       [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
       [36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]])
In [81]:
b[1, 2]  # row 1, col 2
Out[81]:
14
In [82]:
b[1, :]  # row 1, all columns
Out[82]:
array([12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23])
In [83]:
b[:, 1]  # all rows, column 1
Out[83]:
array([ 1, 13, 25, 37])

Caution: note the subtle difference between these two expressions:

In [84]:
b[1, :]
Out[84]:
array([12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23])
In [85]:
b[1:2, :]
Out[85]:
array([[12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]])

The first expression returns row 1 as a 1D array of shape (12,), while the second returns that same row as a 2D array of shape (1, 12).

Fancy indexing

You may also specify a list of indices that you are interested in. This is referred to as fancy indexing.

In [86]:
b[(0,2), 2:5]  # rows 0 and 2, columns 2 to 4 (5-1)
Out[86]:
array([[ 2,  3,  4],
       [26, 27, 28]])
In [87]:
b[:, (-1, 2, -1)]  # all rows, columns -1 (last), 2 and -1 (again, and in this order)
Out[87]:
array([[11,  2, 11],
       [23, 14, 23],
       [35, 26, 35],
       [47, 38, 47]])

If you provide multiple index arrays, you get a 1D ndarray containing the values of the elements at the specified coordinates.

In [88]:
b[(-1, 2, -1, 2), (5, 9, 1, 9)]  # returns a 1D array with b[-1, 5], b[2, 9], b[-1, 1] and b[2, 9] (again)
Out[88]:
array([41, 33, 37, 33])

Higher dimensions

Everything works just as well with higher dimensional arrays, but it's useful to look at a few examples:

In [89]:
c = b.reshape(4,2,6)
c
Out[89]:
array([[[ 0,  1,  2,  3,  4,  5],
        [ 6,  7,  8,  9, 10, 11]],

       [[12, 13, 14, 15, 16, 17],
        [18, 19, 20, 21, 22, 23]],

       [[24, 25, 26, 27, 28, 29],
        [30, 31, 32, 33, 34, 35]],

       [[36, 37, 38, 39, 40, 41],
        [42, 43, 44, 45, 46, 47]]])
In [90]:
c[2, 1, 4]  # matrix 2, row 1, col 4
Out[90]:
34
In [91]:
c[2, :, 3]  # matrix 2, all rows, col 3
Out[91]:
array([27, 33])

If you omit coordinates for some axes, then all elements in these axes are returned:

In [92]:
c[2, 1]  # Return matrix 2, row 1, all columns.  This is equivalent to c[2, 1, :]
Out[92]:
array([30, 31, 32, 33, 34, 35])

Ellipsis (...)

You may also write an ellipsis (...) to ask that all non-specified axes be entirely included.

In [93]:
c[2, ...]  #  matrix 2, all rows, all columns.  This is equivalent to c[2, :, :]
Out[93]:
array([[24, 25, 26, 27, 28, 29],
       [30, 31, 32, 33, 34, 35]])
In [94]:
c[2, 1, ...]  # matrix 2, row 1, all columns.  This is equivalent to c[2, 1, :]
Out[94]:
array([30, 31, 32, 33, 34, 35])
In [95]:
c[2, ..., 3]  # matrix 2, all rows, column 3.  This is equivalent to c[2, :, 3]
Out[95]:
array([27, 33])
In [96]:
c[..., 3]  # all matrices, all rows, column 3.  This is equivalent to c[:, :, 3]
Out[96]:
array([[ 3,  9],
       [15, 21],
       [27, 33],
       [39, 45]])

Boolean indexing

You can also provide an ndarray of boolean values on one axis to specify the indices that you want to access.

In [97]:
b = np.arange(48).reshape(4, 12)
b
Out[97]:
array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11],
       [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23],
       [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
       [36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]])
In [98]:
rows_on = np.array([True, False, True, False])
b[rows_on, :]  # Rows 0 and 2, all columns. Equivalent to b[(0, 2), :]
Out[98]:
array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11],
       [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]])
In [99]:
cols_on = np.array([False, True, False] * 4)
b[:, cols_on]  # All rows, columns 1, 4, 7 and 10
Out[99]:
array([[ 1,  4,  7, 10],
       [13, 16, 19, 22],
       [25, 28, 31, 34],
       [37, 40, 43, 46]])

np.ix_

You cannot use boolean indexing this way on multiple axes, but you can work around this by using the ix_ function:

In [100]:
b[np.ix_(rows_on, cols_on)]
Out[100]:
array([[ 1,  4,  7, 10],
       [25, 28, 31, 34]])
In [101]:
np.ix_(rows_on, cols_on)
Out[101]:
(array([[0],
        [2]]), array([[ 1,  4,  7, 10]]))

If you use a boolean array that has the same shape as the ndarray, then you get in return a 1D array containing all the values that have True at their coordinate. This is generally used along with conditional operators:

In [102]:
b[b % 3 == 1]
Out[102]:
array([ 1,  4,  7, 10, 13, 16, 19, 22, 25, 28, 31, 34, 37, 40, 43, 46])

Iterating

Iterating over ndarrays is very similar to iterating over regular python arrays. Note that iterating over multidimensional arrays is done with respect to the first axis.

In [103]:
c = np.arange(24).reshape(2, 3, 4)  # A 3D array (composed of two 3x4 matrices)
c
Out[103]:
array([[[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11]],

       [[12, 13, 14, 15],
        [16, 17, 18, 19],
        [20, 21, 22, 23]]])
In [104]:
for m in c:
    print("Item:")
    print(m)
Item:
[[ 0  1  2  3]
 [ 4  5  6  7]
 [ 8  9 10 11]]
Item:
[[12 13 14 15]
 [16 17 18 19]
 [20 21 22 23]]
In [105]:
for i in range(len(c)):  # Note that len(c) == c.shape[0]
    print("Item:")
    print(c[i])
Item:
[[ 0  1  2  3]
 [ 4  5  6  7]
 [ 8  9 10 11]]
Item:
[[12 13 14 15]
 [16 17 18 19]
 [20 21 22 23]]

If you want to iterate on all elements in the ndarray, simply iterate over the flat attribute:

In [106]:
for i in c.flat:
    print("Item:", i)
Item: 0
Item: 1
Item: 2
Item: 3
Item: 4
Item: 5
Item: 6
Item: 7
Item: 8
Item: 9
Item: 10
Item: 11
Item: 12
Item: 13
Item: 14
Item: 15
Item: 16
Item: 17
Item: 18
Item: 19
Item: 20
Item: 21
Item: 22
Item: 23

Stacking arrays

It is often useful to stack together different arrays. NumPy offers several functions to do just that. Let's start by creating a few arrays.

In [107]:
q1 = np.full((3,4), 1.0)
q1
Out[107]:
array([[ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.]])
In [108]:
q2 = np.full((4,4), 2.0)
q2
Out[108]:
array([[ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.]])
In [109]:
q3 = np.full((3,4), 3.0)
q3
Out[109]:
array([[ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.]])

vstack

Now let's stack them vertically using vstack:

In [110]:
q4 = np.vstack((q1, q2, q3))
q4
Out[110]:
array([[ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.]])
In [111]:
q4.shape
Out[111]:
(10, 4)

This was possible because q1, q2 and q3 all have the same shape (except for the vertical axis, but that's ok since we are stacking on that axis).

hstack

We can also stack arrays horizontally using hstack:

In [112]:
q5 = np.hstack((q1, q3))
q5
Out[112]:
array([[ 1.,  1.,  1.,  1.,  3.,  3.,  3.,  3.],
       [ 1.,  1.,  1.,  1.,  3.,  3.,  3.,  3.],
       [ 1.,  1.,  1.,  1.,  3.,  3.,  3.,  3.]])
In [113]:
q5.shape
Out[113]:
(3, 8)

This is possible because q1 and q3 both have 3 rows. But since q2 has 4 rows, it cannot be stacked horizontally with q1 and q3:

In [114]:
try:
    q5 = np.hstack((q1, q2, q3))
except ValueError as e:
    print(e)
all the input array dimensions except for the concatenation axis must match exactly

concatenate

The concatenate function stacks arrays along any given existing axis.

In [115]:
q7 = np.concatenate((q1, q2, q3), axis=0)  # Equivalent to vstack
q7
Out[115]:
array([[ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 2.,  2.,  2.,  2.],
       [ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.]])
In [116]:
q7.shape
Out[116]:
(10, 4)

As you might guess, hstack is equivalent to calling concatenate with axis=1.

stack

The stack function stacks arrays along a new axis. All arrays have to have the same shape.

In [117]:
q8 = np.stack((q1, q3))
q8
Out[117]:
array([[[ 1.,  1.,  1.,  1.],
        [ 1.,  1.,  1.,  1.],
        [ 1.,  1.,  1.,  1.]],

       [[ 3.,  3.,  3.,  3.],
        [ 3.,  3.,  3.,  3.],
        [ 3.,  3.,  3.,  3.]]])
In [118]:
q8.shape
Out[118]:
(2, 3, 4)

Splitting arrays

Splitting is the opposite of stacking. For example, let's use the vsplit function to split a matrix vertically.

First let's create a 6x4 matrix:

In [119]:
r = np.arange(24).reshape(6,4)
r
Out[119]:
array([[ 0,  1,  2,  3],
       [ 4,  5,  6,  7],
       [ 8,  9, 10, 11],
       [12, 13, 14, 15],
       [16, 17, 18, 19],
       [20, 21, 22, 23]])

Now let's split it in three equal parts, vertically:

In [120]:
r1, r2, r3 = np.vsplit(r, 3)
r1
Out[120]:
array([[0, 1, 2, 3],
       [4, 5, 6, 7]])
In [121]:
r2
Out[121]:
array([[ 8,  9, 10, 11],
       [12, 13, 14, 15]])
In [122]:
r3
Out[122]:
array([[16, 17, 18, 19],
       [20, 21, 22, 23]])

There is also a split function which splits an array along any given axis. Calling vsplit is equivalent to calling split with axis=0. There is also an hsplit function, equivalent to calling split with axis=1:

In [123]:
r4, r5 = np.hsplit(r, 2)
r4
Out[123]:
array([[ 0,  1],
       [ 4,  5],
       [ 8,  9],
       [12, 13],
       [16, 17],
       [20, 21]])
In [124]:
r5
Out[124]:
array([[ 2,  3],
       [ 6,  7],
       [10, 11],
       [14, 15],
       [18, 19],
       [22, 23]])

Transposing arrays

The transpose method creates a new view on an ndarray's data, with axes permuted in the given order.

For example, let's create a 3D array:

In [125]:
t = np.arange(24).reshape(4,2,3)
t
Out[125]:
array([[[ 0,  1,  2],
        [ 3,  4,  5]],

       [[ 6,  7,  8],
        [ 9, 10, 11]],

       [[12, 13, 14],
        [15, 16, 17]],

       [[18, 19, 20],
        [21, 22, 23]]])

Now let's create an ndarray such that the axes 0, 1, 2 (depth, height, width) are re-ordered to 1, 2, 0 (depth→width, height→depth, width→height):

In [126]:
t1 = t.transpose((1,2,0))
t1
Out[126]:
array([[[ 0,  6, 12, 18],
        [ 1,  7, 13, 19],
        [ 2,  8, 14, 20]],

       [[ 3,  9, 15, 21],
        [ 4, 10, 16, 22],
        [ 5, 11, 17, 23]]])
In [127]:
t1.shape
Out[127]:
(2, 3, 4)

By default, transpose reverses the order of the dimensions:

In [128]:
t2 = t.transpose()  # equivalent to t.transpose((2, 1, 0))
t2
Out[128]:
array([[[ 0,  6, 12, 18],
        [ 3,  9, 15, 21]],

       [[ 1,  7, 13, 19],
        [ 4, 10, 16, 22]],

       [[ 2,  8, 14, 20],
        [ 5, 11, 17, 23]]])
In [129]:
t2.shape
Out[129]:
(3, 2, 4)

NumPy provides a convenience function swapaxes to swap two axes. For example, let's create a new view of t with depth and height swapped:

In [130]:
t3 = t.swapaxes(0,1)  # equivalent to t.transpose((1, 0, 2))
t3
Out[130]:
array([[[ 0,  1,  2],
        [ 6,  7,  8],
        [12, 13, 14],
        [18, 19, 20]],

       [[ 3,  4,  5],
        [ 9, 10, 11],
        [15, 16, 17],
        [21, 22, 23]]])
In [131]:
t3.shape
Out[131]:
(2, 4, 3)

Linear algebra

NumPy 2D arrays can be used to represent matrices efficiently in python. We will just quickly go through some of the main matrix operations available. For more details about Linear Algebra, vectors and matrics, go through the Linear Algebra tutorial.

Matrix transpose

The T attribute is equivalent to calling transpose() when the rank is ≥2:

In [132]:
m1 = np.arange(10).reshape(2,5)
m1
Out[132]:
array([[0, 1, 2, 3, 4],
       [5, 6, 7, 8, 9]])
In [133]:
m1.T
Out[133]:
array([[0, 5],
       [1, 6],
       [2, 7],
       [3, 8],
       [4, 9]])

The T attribute has no effect on rank 0 (empty) or rank 1 arrays:

In [134]:
m2 = np.arange(5)
m2
Out[134]:
array([0, 1, 2, 3, 4])
In [135]:
m2.T
Out[135]:
array([0, 1, 2, 3, 4])

We can get the desired transposition by first reshaping the 1D array to a single-row matrix (2D):

In [136]:
m2r = m2.reshape(1,5)
m2r
Out[136]:
array([[0, 1, 2, 3, 4]])
In [137]:
m2r.T
Out[137]:
array([[0],
       [1],
       [2],
       [3],
       [4]])

Matrix multiplication

Let's create two matrices and execute a matrix multiplication using the dot() method.

In [138]:
n1 = np.arange(10).reshape(2, 5)
n1
Out[138]:
array([[0, 1, 2, 3, 4],
       [5, 6, 7, 8, 9]])
In [139]:
n2 = np.arange(15).reshape(5,3)
n2
Out[139]:
array([[ 0,  1,  2],
       [ 3,  4,  5],
       [ 6,  7,  8],
       [ 9, 10, 11],
       [12, 13, 14]])
In [140]:
n1.dot(n2)
Out[140]:
array([[ 90, 100, 110],
       [240, 275, 310]])

Caution: as mentionned previously, n1*n2 is not a matric multiplication, it is an elementwise product (also called a Hadamard product)).

Matrix inverse and pseudo-inverse

Many of the linear algebra functions are available in the numpy.linalg module, in particular the inv function to compute a square matrix's inverse:

In [141]:
import numpy.linalg as linalg

m3 = np.array([[1,2,3],[5,7,11],[21,29,31]])
m3
Out[141]:
array([[ 1,  2,  3],
       [ 5,  7, 11],
       [21, 29, 31]])
In [142]:
linalg.inv(m3)
Out[142]:
array([[-2.31818182,  0.56818182,  0.02272727],
       [ 1.72727273, -0.72727273,  0.09090909],
       [-0.04545455,  0.29545455, -0.06818182]])

You can also compute the pseudoinverse using pinv:

In [143]:
linalg.pinv(m3)
Out[143]:
array([[-2.31818182,  0.56818182,  0.02272727],
       [ 1.72727273, -0.72727273,  0.09090909],
       [-0.04545455,  0.29545455, -0.06818182]])

Identity matrix

The product of a matrix by its inverse returns the identiy matrix (with small floating point errors):

In [144]:
m3.dot(linalg.inv(m3))
Out[144]:
array([[  1.00000000e+00,  -1.11022302e-16,  -6.93889390e-18],
       [ -1.33226763e-15,   1.00000000e+00,  -5.55111512e-17],
       [  2.88657986e-15,   0.00000000e+00,   1.00000000e+00]])

You can create an identity matrix of size NxN by calling eye:

In [145]:
np.eye(3)
Out[145]:
array([[ 1.,  0.,  0.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  1.]])

QR decomposition

The qr function computes the QR decomposition of a matrix:

In [146]:
q, r = linalg.qr(m3)
q
Out[146]:
array([[-0.04627448,  0.98786672,  0.14824986],
       [-0.23137241,  0.13377362, -0.96362411],
       [-0.97176411, -0.07889213,  0.22237479]])
In [147]:
r
Out[147]:
array([[-21.61018278, -29.89331494, -32.80860727],
       [  0.        ,   0.62427688,   1.9894538 ],
       [  0.        ,   0.        ,  -3.26149699]])
In [148]:
q.dot(r)  # q.r equals m3
Out[148]:
array([[  1.,   2.,   3.],
       [  5.,   7.,  11.],
       [ 21.,  29.,  31.]])

Determinant

The det function computes the matrix determinant:

In [149]:
linalg.det(m3)  # Computes the matrix determinant
Out[149]:
43.999999999999972

Eigenvalues and eigenvectors

The eig function computes the eigenvalues and eigenvectors of a square matrix:

In [150]:
eigenvalues, eigenvectors = linalg.eig(m3)
eigenvalues # λ
Out[150]:
array([ 42.26600592,  -0.35798416,  -2.90802176])
In [151]:
eigenvectors # v
Out[151]:
array([[-0.08381182, -0.76283526, -0.18913107],
       [-0.3075286 ,  0.64133975, -0.6853186 ],
       [-0.94784057, -0.08225377,  0.70325518]])
In [152]:
m3.dot(eigenvectors) - eigenvalues * eigenvectors  # m3.v - λ*v = 0
Out[152]:
array([[  8.88178420e-15,   2.49800181e-15,  -3.33066907e-16],
       [  1.77635684e-14,  -1.66533454e-16,  -3.55271368e-15],
       [  3.55271368e-14,   3.61516372e-15,  -4.44089210e-16]])

Singular Value Decomposition

The svd function takes a matrix and returns its singular value decomposition:

In [153]:
m4 = np.array([[1,0,0,0,2], [0,0,3,0,0], [0,0,0,0,0], [0,2,0,0,0]])
m4
Out[153]:
array([[1, 0, 0, 0, 2],
       [0, 0, 3, 0, 0],
       [0, 0, 0, 0, 0],
       [0, 2, 0, 0, 0]])
In [154]:
U, S_diag, V = linalg.svd(m4)
U
Out[154]:
array([[ 0.,  1.,  0.,  0.],
       [ 1.,  0.,  0.,  0.],
       [ 0.,  0.,  0., -1.],
       [ 0.,  0.,  1.,  0.]])
In [155]:
S_diag
Out[155]:
array([ 3.        ,  2.23606798,  2.        ,  0.        ])

The svd function just returns the values in the diagonal of Σ, but we want the full Σ matrix, so let's create it:

In [156]:
S = np.zeros((4, 5))
S[np.diag_indices(4)] = S_diag
S  # Σ
Out[156]:
array([[ 3.        ,  0.        ,  0.        ,  0.        ,  0.        ],
       [ 0.        ,  2.23606798,  0.        ,  0.        ,  0.        ],
       [ 0.        ,  0.        ,  2.        ,  0.        ,  0.        ],
       [ 0.        ,  0.        ,  0.        ,  0.        ,  0.        ]])
In [157]:
V
Out[157]:
array([[-0.        ,  0.        ,  1.        , -0.        ,  0.        ],
       [ 0.4472136 ,  0.        ,  0.        ,  0.        ,  0.89442719],
       [-0.        ,  1.        ,  0.        , -0.        ,  0.        ],
       [ 0.        ,  0.        ,  0.        ,  1.        ,  0.        ],
       [-0.89442719,  0.        ,  0.        ,  0.        ,  0.4472136 ]])
In [158]:
U.dot(S).dot(V) # U.Σ.V == m4
Out[158]:
array([[ 1.,  0.,  0.,  0.,  2.],
       [ 0.,  0.,  3.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.],
       [ 0.,  2.,  0.,  0.,  0.]])

Diagonal and trace

In [159]:
np.diag(m3)  # the values in the diagonal of m3 (top left to bottom right)
Out[159]:
array([ 1,  7, 31])
In [160]:
np.trace(m3)  # equivalent to np.diag(m3).sum()
Out[160]:
39

Solving a system of linear scalar equations

The solve function solves a system of linear scalar equations, such as:

  • $2x + 6y = 6$
  • $5x + 3y = -9$
In [161]:
coeffs  = np.array([[2, 6], [5, 3]])
depvars = np.array([6, -9])
solution = linalg.solve(coeffs, depvars)
solution
Out[161]:
array([-3.,  2.])

Let's check the solution:

In [162]:
coeffs.dot(solution), depvars  # yep, it's the same
Out[162]:
(array([ 6., -9.]), array([ 6, -9]))

Looks good! Another way to check the solution:

In [163]:
np.allclose(coeffs.dot(solution), depvars)
Out[163]:
True

Vectorization

Instead of executing operations on individual array items, one at a time, your code is much more efficient if you try to stick to array operations. This is called vectorization. This way, you can benefit from NumPy's many optimizations.

For example, let's say we want to generate a 768x1024 array based on the formula $sin(xy/40.5)$. A bad option would be to do the math in python using nested loops:

In [164]:
import math
data = np.empty((768, 1024))
for y in range(768):
    for x in range(1024):
        data[y, x] = math.sin(x*y/40.5)  # BAD! Very inefficient.

Sure, this works, but it's terribly inefficient since the loops are taking place in pure python. Let's vectorize this algorithm. First, we will use NumPy's meshgrid function which generates coordinate matrices from coordinate vectors.

In [165]:
x_coords = np.arange(0, 1024)  # [0, 1, 2, ..., 1023]
y_coords = np.arange(0, 768)   # [0, 1, 2, ..., 767]
X, Y = np.meshgrid(x_coords, y_coords)
X
Out[165]:
array([[   0,    1,    2, ..., 1021, 1022, 1023],
       [   0,    1,    2, ..., 1021, 1022, 1023],
       [   0,    1,    2, ..., 1021, 1022, 1023],
       ..., 
       [   0,    1,    2, ..., 1021, 1022, 1023],
       [   0,    1,    2, ..., 1021, 1022, 1023],
       [   0,    1,    2, ..., 1021, 1022, 1023]])
In [166]:
Y
Out[166]:
array([[  0,   0,   0, ...,   0,   0,   0],
       [  1,   1,   1, ...,   1,   1,   1],
       [  2,   2,   2, ...,   2,   2,   2],
       ..., 
       [765, 765, 765, ..., 765, 765, 765],
       [766, 766, 766, ..., 766, 766, 766],
       [767, 767, 767, ..., 767, 767, 767]])

As you can see, both X and Y are 768x1024 arrays, and all values in X correspond to the horizontal coordinate, while all values in Y correspond to the the vertical coordinate.

Now we can simply compute the result using array operations:

In [167]:
data = np.sin(X*Y/40.5)

Now we can plot this data using matplotlib's imshow function (see the matplotlib tutorial).

In [168]:
import matplotlib.pyplot as plt
import matplotlib.cm as cm
fig = plt.figure(1, figsize=(7, 6))
plt.imshow(data, cmap=cm.hot, interpolation="bicubic")
plt.show()

Saving and loading

NumPy makes it easy to save and load ndarrays in binary or text format.

Binary .npy format

Let's create a random array and save it.

In [169]:
a = np.random.rand(2,3)
a
Out[169]:
array([[ 0.41307972,  0.20933385,  0.32025581],
       [ 0.19853514,  0.408001  ,  0.6038287 ]])
In [170]:
np.save("my_array", a)

Done! Since the file name contains no file extension was provided, NumPy automatically added .npy. Let's take a peek at the file content:

In [171]:
with open("my_array.npy", "rb") as f:
    content = f.read()

content
Out[171]:
"\x93NUMPY\x01\x00F\x00{'descr': '<f8', 'fortran_order': False, 'shape': (2, 3), }          \n\xa8\x96\x1d\xeb\xe5o\xda? \x06W\xa1s\xcb\xca?*\xdeB>\x12\x7f\xd4?x<h\x81\x99i\xc9?@\xa4\x027\xb0\x1c\xda?<P\x05\x8f\x90R\xe3?"

To load this file into a NumPy array, simply call load:

In [172]:
a_loaded = np.load("my_array.npy")
a_loaded
Out[172]:
array([[ 0.41307972,  0.20933385,  0.32025581],
       [ 0.19853514,  0.408001  ,  0.6038287 ]])

Text format

Let's try saving the array in text format:

In [173]:
np.savetxt("my_array.csv", a)

Now let's look at the file content:

In [174]:
with open("my_array.csv", "rt") as f:
    print(f.read())
4.130797191668116319e-01 2.093338525574361952e-01 3.202558143634371968e-01
1.985351449843368865e-01 4.080009972772735694e-01 6.038286965726977762e-01

This is a CSV file with tabs as delimiters. You can set a different delimiter:

In [175]:
np.savetxt("my_array.csv", a, delimiter=",")

To load this file, just use loadtxt:

In [176]:
a_loaded = np.loadtxt("my_array.csv", delimiter=",")
a_loaded
Out[176]:
array([[ 0.41307972,  0.20933385,  0.32025581],
       [ 0.19853514,  0.408001  ,  0.6038287 ]])

Zipped .npz format

It is also possible to save multiple arrays in one zipped file:

In [177]:
b = np.arange(24, dtype=np.uint8).reshape(2, 3, 4)
b
Out[177]:
array([[[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11]],

       [[12, 13, 14, 15],
        [16, 17, 18, 19],
        [20, 21, 22, 23]]], dtype=uint8)
In [178]:
np.savez("my_arrays", my_a=a, my_b=b)

Again, let's take a peek at the file content. Note that the .npz file extension was automatically added.

In [179]:
with open("my_arrays.npz", "rb") as f:
    content = f.read()

repr(content)[:180] + "[...]"
Out[179]:
u'"PK\\x03\\x04\\x14\\x00\\x00\\x00\\x00\\x00x\\x94cH\\xb6\\x96\\xe4{h\\x00\\x00\\x00h\\x00\\x00\\x00\\x08\\x00\\x00\\x00my_b.npy\\x93NUMPY\\x01\\x00F\\x00{\'descr\': \'|u1\', \'fortran_order\': False, \'shape\': (2,[...]'

You then load this file like so:

In [180]:
my_arrays = np.load("my_arrays.npz")
my_arrays
Out[180]:
<numpy.lib.npyio.NpzFile at 0x10fa4d4d0>

This is a dict-like object which loads the arrays lazily:

In [181]:
my_arrays.keys()
Out[181]:
['my_b', 'my_a']
In [182]:
my_arrays["my_a"]
Out[182]:
array([[ 0.41307972,  0.20933385,  0.32025581],
       [ 0.19853514,  0.408001  ,  0.6038287 ]])

What next?

Now you know all the fundamentals of NumPy, but there are many more options available. The best way to learn more is to experiment with NumPy, and go through the excellent reference documentation to find more functions and features you may be interested in.