当我尝试使用NumPy和Pandas两种方式时，使用Pandas有很多优点:

快 减少CPU占用 与NumPy genfromttxt相比，RAM占用了1/3

这是我的测试代码:

$ for f in test_pandas.py test_numpy_csv.py ; do  /usr/bin/time python $f; done
2.94user 0.41system 0:03.05elapsed 109%CPU (0avgtext+0avgdata 502068maxresident)k
0inputs+24outputs (0major+107147minor)pagefaults 0swaps

23.29user 0.72system 0:23.72elapsed 101%CPU (0avgtext+0avgdata 1680888maxresident)k
0inputs+0outputs (0major+416145minor)pagefaults 0swaps

test_numpy_csv.py

from numpy import genfromtxt
train = genfromtxt('/home/hvn/me/notebook/train.csv', delimiter=',')

test_pandas.py

from pandas import read_csv
df = read_csv('/home/hvn/me/notebook/train.csv')

数据文件:

du -h ~/me/notebook/train.csv
 59M    /home/hvn/me/notebook/train.csv

在NumPy和熊猫版本:

$ pip freeze | egrep -i 'pandas|numpy'
numpy==1.13.3
pandas==0.20.2

我试了一下:

from numpy import genfromtxt
genfromtxt(fname = dest_file, dtype = (<whatever options>))

对比:

import csv
import numpy as np
with open(dest_file,'r') as dest_f:
    data_iter = csv.reader(dest_f,
                           delimiter = delimiter,
                           quotechar = '"')
    data = [data for data in data_iter]
data_array = np.asarray(data, dtype = <whatever options>)

对460万行约70列进行了分析，发现NumPy路径花费了2分16秒，csv-list理解方法花费了13秒。

我会推荐csv-list理解方法，因为它很可能依赖于预编译的库，而不是像NumPy那样依赖解释器。我怀疑pandas方法也有类似的解释器开销。

使用numpy.loadtxt

一个非常简单的方法。但它要求所有元素都是float (int等)

import numpy as np 
data = np.loadtxt('c:\\1.csv',delimiter=',',skiprows=0)  

In [329]: %time my_data = genfromtxt('one.csv', delimiter=',')
CPU times: user 19.8 s, sys: 4.58 s, total: 24.4 s
Wall time: 24.4 s

In [330]: %time df = pd.read_csv("one.csv", skiprows=20)
CPU times: user 1.06 s, sys: 312 ms, total: 1.38 s
Wall time: 1.38 s

这件作品很有魅力……

import csv
with open("data.csv", 'r') as f:
    data = list(csv.reader(f, delimiter=";"))

import numpy as np
data = np.array(data, dtype=np.float)

使用numpy.genfromtxt()，将分隔符kwarg设置为逗号:

from numpy import genfromtxt
my_data = genfromtxt('my_file.csv', delimiter=',')