Numpy 性能

#简单好用的python（速度慢）+ 速度较快的numpy（基于C语言）
#在numpy中，创建二维数组的默认方式为“C-type”，以行序在内存中存储数据
#如果是“Fortran”方式创建，是以列序在内存中存储
import numpy as np
import pandas as pd
import time
row_major = np.zeros((10,10),order = 'C') #C-type,按行序合并（axis=0），较快
col_major = np.zeros((10,10),order = 'F') #Fortran，按列序合并（axis=1），较快

a = np.zeros((200,200),order ='C')
b = np.zeros((200,200),order ='F')
N = 9999

def f1(a):
    for _ in range(N):
        np.concatenate((a,a),axis = 0)

def f2(b):
    for _ in range(N):
        np.concatenate((b,b),axis = 0)

t0 = time.time()

f1(a)

t1 = time.time()

f2(b)

t2 = time.time()

print((t1-t0)/N)
print((t2-t1)/N)

#################
np.vstack((a,a))#与concatenate相比，较慢

np.concatenate((a,a),axis = 0) #相对于vstack而言，较快
#################

indices = np.random.randint(0,100,size = 10)
a[indices,:] #以行序存储数据，在行上创建数据要更快
a[:,indices] #（1）concatenate 优于 vstack （2）axis = 0，order = 'C'更优
             #axis = 1, order = 'F'更优（3）行序，a[indices,:]更优
             #列序，a[:,indices]更优

#view 优于copy，view是取源数据的索引部分，copy是将数据复制出来，存到内存的另一个地方
a = np.arange(1,7).reshape((3,2))
a_view = a[:2]
a_copy = a[:2].copy()

a_copy[1,1] = 0
print(a)#源数据不变

a_view[1,1] = 0
print(a)#源数据改变

a = np.zeros((1000,1000))
b = np.zeros((1000,1000))
N = 9999

def f1(a):
    for _ in range(N):
        a *= 2 #等同于a[:] *= 2
def f2(b):
    for _ in range(N):
        b = 2*b
t0 = time.time()
f1(a)
t1 = time.time()
f2(b)
t2 = time.time()

print('%f' %((t1-t0)/N)) #速度更快
print('%f' %((t2-t1)/N))

#在矩阵展平的时候，np.ravel()返回的是一个view索引，np.flatten()
#返回的是一个copy，ravel的速度更快

def f1(a):
    for _ in range(N):
        a.flatten()
def f2(b):
    for _ in range(N):
        b.ravel()
#以下都是使用view的方式
a_view1 = a[1:2,3:6] #切片slice
a_view2 = a[:100] #同上
a_view3 = a[::2] #跳步
a_view4 = a.ravel()#矩阵展平
#以下都是使用copy的方式
a_copy1 = a[[1,4,6],[2,4,6]] #用index选择
a_copy2 = a[[True,True],[False,True]] #用mask
a_copy3 = a[[1,2],:]
a_copy4 = a[a[1,:] != 0,:] #fancy indexing
a_copy5 = a[np.isnan(a),:] #fancy indexing

#用np.take()选择数据比用index要快
a = np.random.rand(1000000,10)
N = 99
indices = np.random.randint(0,1000000,size = 10000)

def f1(a):
    for _ in range(N):
        _ = np.take(a,indices,axis = 0) #快
def f2(b):
    for _ in range(N):
        _ = b[indices]

#用np.compress()代替mask
mask = a[:,0] < 0.5
def f1(a):
    for _ in range(N)
        _ = np.compress(mask,a,axis = 0)#快

def f2(b):
    for _ in range(N):
        _ = b[mask]

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

import numpy as np

np.add(a,1,out = a) #如果已经有了一个placeholder（比如a),没有必要再复制一个
#速度最快
a += 1 #速度次之
a = np.add(a,1) #速度再次之
a = a+1 #速度最慢

a = np.zeros(3,dtype = [('foo',np.int32),('bar',np.float16)])
b = pd.DataFrame(np.zeros((3,2),dtype=np.int32),columns = ['foo','bar'])
b['bar'] = b['bar'].astype(np.float16)

def f1(a):
    for _ in range(N):
        a['bar'] *= a['foo']#numpy比较快

def f2(b):
    for _ in range(N):
        b['bar'] *= b['foo']#pandas比较慢