公认的解决方案对于大量数据来说将会非常慢。获得最多赞数的解决方案读起来有点困难,而且处理数字数据也很慢。如果每个新列都可以独立于其他列计算,那么我将直接分配它们,而不使用apply。
假字符数据的例子
在DataFrame中创建100,000个字符串
df = pd.DataFrame(np.random.choice(['he jumped', 'she ran', 'they hiked'],
size=100000, replace=True),
columns=['words'])
df.head()
words
0 she ran
1 she ran
2 they hiked
3 they hiked
4 they hiked
假设我们想提取一些文本特征,就像在最初的问题中所做的那样。例如,让我们提取第一个字符,计算字母“e”的出现次数,并将短语大写。
df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
df.head()
words first count_e cap
0 she ran s 1 She ran
1 she ran s 1 She ran
2 they hiked t 2 They hiked
3 they hiked t 2 They hiked
4 they hiked t 2 They hiked
计时
%%timeit
df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
127 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
def extract_text_features(x):
return x[0], x.count('e'), x.capitalize()
%timeit df['first'], df['count_e'], df['cap'] = zip(*df['words'].apply(extract_text_features))
101 ms ± 2.96 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
令人惊讶的是,通过遍历每个值可以获得更好的性能
%%timeit
a,b,c = [], [], []
for s in df['words']:
a.append(s[0]), b.append(s.count('e')), c.append(s.capitalize())
df['first'] = a
df['count_e'] = b
df['cap'] = c
79.1 ms ± 294 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
另一个假数字数据的例子
创建100万个随机数并从上面测试幂函数。
df = pd.DataFrame(np.random.rand(1000000), columns=['num'])
def powers(x):
return x, x**2, x**3, x**4, x**5, x**6
%%timeit
df['p1'], df['p2'], df['p3'], df['p4'], df['p5'], df['p6'] = \
zip(*df['num'].map(powers))
1.35 s ± 83.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
为每一列赋值速度快25倍,可读性强:
%%timeit
df['p1'] = df['num'] ** 1
df['p2'] = df['num'] ** 2
df['p3'] = df['num'] ** 3
df['p4'] = df['num'] ** 4
df['p5'] = df['num'] ** 5
df['p6'] = df['num'] ** 6
51.6 ms ± 1.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
我在这里也做了类似的回答,并详细说明了为什么申请通常不是正确的选择。