Chapter12_Advanced_Numpy

import numpy as npimport pandas as pdimport matplotlib.pyplot as plt

//anaconda/lib/python2.7/site-packages/matplotlib/font_manager.py:273: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.  warnings.warn('Matplotlib is building the font cache using fc-list. This may take a moment.')

%matplotlib inline

Setting Array Values by Broadcasting

The same broadcsting rule governing arithmetic operations also applies to setting values via array indexing.

from pandas import DataFrame,Series

arr = np.zeros((4,3))

arr[:] = 5

arr

array([[ 5.,  5.,  5.],       [ 5.,  5.,  5.],       [ 5.,  5.,  5.],       [ 5.,  5.,  5.]])

However, if we had a one-dimentional array of values we wanted to set into the columns of the array,we can do that as long as the shape is compatible:

Advanced ufunc Usage

While many NumPy users will only make use of the fast element-wise operations pro- vided by the universal functions, there are a number of additional features that occa- sionally can help you write more concise code without loops.

ufunc Instance Methods

Each of NumPy’s binary ufuncs has special methods for performing certain kinds of special vectorized operations.

Method Description reduce(x) Aggregate values by successive applications of the operation accumulate(x) Aggregate values, preserving all partial aggregates reduceat(x, bins) “Local” reduce or “group by”. Reduce contiguous slices of data to produce aggregated array. outer(x, y) Apply operation to all pairs of elements in x and y. Result array has shape x.shape + y.shape

reduce

takes a single array and aggregates its values, optionally along an axis, by performing a sequence of binary operations.

For example, an alternate way to sum elements in an array is to use np.add.reduce:

arr = np.arange(10)

np.add.reduce(arr)

45

np.sum(arr)

45

a less trivial example

use np.logical_and to check whether the values in each row of an array are sorted:

arr = np.random.randn(5, 5)

arr[::2].sort(1)

np.sort

arr

array([[ -1.58645456e+00,  -1.36502511e+00,  -5.41627950e-04,          4.86430078e-01,   1.17172633e+00],       [ -7.44597469e-01,  -4.24979151e-01,  -2.24296763e+00,          3.15071902e-01,   9.84530055e-01],       [ -1.54861818e+00,  -3.99856553e-01,  -1.67025551e-01,          8.91978379e-02,   1.31957741e+00],       [  4.00996840e-01,  -2.38865213e-03,   1.70717270e-01,         -9.66816316e-01,   3.94653542e-01],       [ -1.46761823e+00,  -8.56722218e-01,  -5.79742413e-01,          2.53720291e-01,   2.97676083e+00]])

arr[:,:-1]

array([[ -1.58645456e+00,  -1.36502511e+00,  -5.41627950e-04,          4.86430078e-01],       [ -7.44597469e-01,  -4.24979151e-01,  -2.24296763e+00,          3.15071902e-01],       [ -1.54861818e+00,  -3.99856553e-01,  -1.67025551e-01,          8.91978379e-02],       [  4.00996840e-01,  -2.38865213e-03,   1.70717270e-01,         -9.66816316e-01],       [ -1.46761823e+00,  -8.56722218e-01,  -5.79742413e-01,          2.53720291e-01]])

arr[:,1:]

array([[ -1.36502511e+00,  -5.41627950e-04,   4.86430078e-01,          1.17172633e+00],       [ -4.24979151e-01,  -2.24296763e+00,   3.15071902e-01,          9.84530055e-01],       [ -3.99856553e-01,  -1.67025551e-01,   8.91978379e-02,          1.31957741e+00],       [ -2.38865213e-03,   1.70717270e-01,  -9.66816316e-01,          3.94653542e-01],       [ -8.56722218e-01,  -5.79742413e-01,   2.53720291e-01,          2.97676083e+00]])

arr[:,:-1] < arr[:,1:]

array([[ True,  True,  True,  True],       [ True, False,  True,  True],       [ True,  True,  True,  True],       [False,  True, False,  True],       [ True,  True,  True,  True]], dtype=bool)

np.logical_and.reduce(arr[:,:-1] < arr[:,1:], axis=1)

array([ True, False,  True, False,  True], dtype=bool)

arr[::2]

array([[ -1.58645456e+00,  -1.36502511e+00,  -5.41627950e-04,          4.86430078e-01,   1.17172633e+00],       [ -1.54861818e+00,  -3.99856553e-01,  -1.67025551e-01,          8.91978379e-02,   1.31957741e+00],       [ -1.46761823e+00,  -8.56722218e-01,  -5.79742413e-01,          2.53720291e-01,   2.97676083e+00]])

accumulate

---

accumulate is related to reduce like cumsum is related to sum. It produces an array of the
same size with the intermediate “accumulated” values:

arr = np.arange(15).reshape((3,5))

np.add.accumulate(arr, axis=1)

array([[ 0,  1,  3,  6, 10],       [ 5, 11, 18, 26, 35],       [10, 21, 33, 46, 60]])

outer

outer performs a pairwise cross-product between two arrays:

arr = np.arange(3).repeat([1,2,3])

arr

array([0, 1, 1, 2, 2, 2])

np.multiply.outer(arr, np.arange(5))

array([[0, 0, 0, 0, 0],       [0, 1, 2, 3, 4],       [0, 1, 2, 3, 4],       [0, 2, 4, 6, 8],       [0, 2, 4, 6, 8],       [0, 2, 4, 6, 8]])

np.multiply.outer

outer(A, B)

Apply the ufunc op to all pairs (a, b) with a in A and b in B.

result = np.subtract.outer(np.random.randn(3,4),np.random.randn(5))

result

array([[[-1.54216517, -0.52986349, -1.34379847, -2.32636464, -2.53081338],        [-0.32168391,  0.69061777, -0.12331721, -1.10588338, -1.31033212],        [ 1.00529622,  2.0175979 ,  1.20366292,  0.22109675,  0.01664801],        [ 1.02270583,  2.0350075 ,  1.22107253,  0.23850636,  0.03405762]],       [[ 2.18387989,  3.19618157,  2.38224659,  1.39968042,  1.19523168],        [-0.36959033,  0.64271135, -0.17122363, -1.1537898 , -1.35823854],        [ 0.49572124,  1.50802291,  0.69408793, -0.28847823, -0.49292697],        [-3.38464023, -2.37233855, -3.18627353, -4.1688397 , -4.37328844]],       [[-0.28755193,  0.72474974, -0.08918523, -1.0717514 , -1.27620014],        [-0.99607282,  0.01622886, -0.79770612, -1.78027229, -1.98472103],        [ 0.57069522,  1.5829969 ,  0.76906192, -0.21350425, -0.41795299],        [ 0.70100281,  1.71330448,  0.8993695 , -0.08319666, -0.2876454 ]]])

reduceat

The last method, reduceat, performs a “local reduce”, in essence an array groupby operation in which slices of the array are aggregated together. While it’s less flexible than the GroupBy capabilities in pandas, it can be very fast and powerful in the right circumstances. It accepts a sequence of “bin edges” which indicate how to split and aggregate the values:

np.add.reduceat

arr = np.arange(8)

arr[::2]

array([0, 2, 4, 6])

np.add.reduceat(arr, [1,5,6,2])

array([10,  5,  6, 27])

arr = np.multiply.outer(np.arange(4), np.arange(5))

arr

array([[ 0,  0,  0,  0,  0],       [ 0,  1,  2,  3,  4],       [ 0,  2,  4,  6,  8],       [ 0,  3,  6,  9, 12]])

对于reduce at 可以选择列作为聚合的对象

np.add.reduceat(arr, [0, 2, 4], axis=1)

array([[ 0,  0,  0],       [ 1,  5,  4],       [ 2, 10,  8],       [ 3, 15, 12]])

User Function

np.frompyfunc

Docstring:
frompyfunc(func, nin, nout)

Takes an arbitrary Python function and returns a Numpy ufunc.

Can be used, for example, to add broadcasting to a built-in Python
function (see Examples section).

Parameters name func : Python function object An arbitrary Python function. nin : int The number of input arguments. nout : int The number of objects returned by func.

Returns

out : ufunc
Returns a Numpy universal function (ufunc) object.

Notes

The returned ufunc always returns PyObject arrays.

Examples

Use frompyfunc to add broadcasting to the Python function oct:

>>> oct_array = np.frompyfunc(oct, 1, 1)>>> oct_array(np.array((10, 30, 100)))array([012, 036, 0144], dtype=object)>>> np.array((oct(10), oct(30), oct(100))) # for comparisonarray(['012', '036', '0144'],      dtype='|S4')Type:      builtin_function_or_method

def user_function(a, b):    return a**b

ufunc = np.frompyfunc(user_function, 2, 1)

ufunc(np.arange(10), np.arange(10))

array([1, 1, 4, 27, 256, 3125, 46656, 823543, 16777216, 387420489], dtype=object)

These functions provide a way to create ufunc-like functions, but they are very slow because they require a Python function call to compute each element, which is a lot slower than NumPy’s C-based ufunc loops.

Structured and Record Arrays

You may have noticed up until now that ndarray is a homogeneous data container; that is, it represents a block of memory in which each element takes up the same number of bytes, determined by the dtype. On the surface, this would appear to not allow you to represent heterogeneous or tabular-like data. A structured array is an ndarray in which each element can be thought of as representing a struct in C (hence the “struc- tured” name) or a row in a SQL table with multiple named fields:

dtype = [('x', np.float),('y',np.float128)]

sarr = np.array([(1.5, 6),(np.pi, -2)],dtype=dtype)

sarr

array([(1.5, 6.0), (3.141592653589793, -2.0)],       dtype=[('x', '<f8'), ('y', '<f16')])

sarr.dtype.names

('x', 'y')

np.dtype.names

<attribute 'names' of 'numpy.dtype' objects>

structured data type

The elements of a numpy array is the same,so if we want to change the type of the elemnets here are three ways

• use columns of tuples like [(typy name,type)]

if you want to change the dimention of the type you can just add an int or tuple into the dtype columns like this

• [(type name,type,int)]

if you want to nest the dtype you can use something like this

• [(type name,[(typr name, type),(type name, type)])]

Sorting

arr.sort() is a in-place function which means it will change the sequence of the original array
- if you sort the view of an array ,the original array will be sorted too
- numpy.sort will create a new array
- arr[::-1] can return a descending array since the sort method will always return the ascending array

arr = np.array([0, 1, 2, 3, 5])

arr.argsort()

array([0, 1, 2, 3, 4])

first_name = np.array(['Bob', 'Jane', 'Steve', 'Bill', 'Barbara'])

last_name = np.array(['Jones', 'Arnold', 'Arnold', 'Jones', 'Walters'])

np.lexsort((first_name, last_name))

array([1, 2, 3, 0, 4])

sort = np.lexsort((first_name, last_name))

zip(last_name[sort], first_name[sort])

[('Arnold', 'Jane'), ('Arnold', 'Steve'), ('Jones', 'Bill'), ('Jones', 'Bob'), ('Walters', 'Barbara')]

numpy.lexisort

> lexisort can sort a hierachical data structure by thisarr.lexisort((secondary order list, primary order list))