对于有熊猫的循环 - 我什么时候应该关心？

Question

我熟悉"矢量化"的概念,以及熊猫如何使用矢量化技术来加速计算.矢量化函数在整个系列或DataFrame上广播操作,以实现比传统迭代数据更大的加速.

但是,我很惊讶地看到很多代码(包括Stack Overflow的答案)提供了解决问题的方法,这些问题涉及使用for循环和列表推导来循环数据.阅读完文档后,对API有了不错的理解,我认为循环是"坏的",并且应该"永远"迭代数组,系列或DataFrame.那么,为什么我会不时地看到用户提出循环解决方案呢？

因此,要总结......我的问题是:
是否for循环真正的"坏"？如果不是,在什么情况下它们会比使用更传统的"矢量化"方法更好？¹

_{1 - 虽然这个问题确实听起来有点宽泛,但事实是,当for循环通常比传统的迭代数据更好时,存在非常具体的情况.这篇文章旨在为后人捕捉这一点.}

Answer 1

TLDR; 不,for循环不是"不好",至少,并非总是如此.一些矢量化操作比迭代慢,可能更准确,而不是说迭代比一些矢量化操作更快.了解何时以及为什么是从代码中获得最佳性能的关键.简而言之,这些是值得考虑替代矢量化熊猫函数的情况:

1. 当您的数据很小时(...取决于您正在做的事情),
2. 处理object/混合dtypes时
3. 使用str/ regex访问器功能时

让我们分别研究这些情况.

迭代v/s小数据矢量化

Pandas 在其API设计中遵循"约定配置"方法.这意味着已经安装了相同的API以满足广泛的数据和用例.

当调用pandas函数时,函数必须在内部处理以下内容(以及其他内容),以确保工作

1. 索引/轴对齐
2. 处理混合数据类型
3. 处理丢失的数据

几乎每个函数都必须以不同的程度处理这些函数,这会产生开销.对于数字函数(例如,Series.add)的开销较小,而对于字符串函数(例如,Series.str.replace)则更为明显.

for loops, on the other hand, are faster then you think. What's even better is list comprehensions (which create lists through for loops) are even faster as they are optimized iterative mechanisms for list creation.

List comprehensions follow the pattern

[f(x) for x in seq]

Where seq is a pandas series or DataFrame column. Or, when operating over multiple columns,

[f(x, y) for x, y in zip(seq1, seq2)]

Where seq1 and seq2 are columns.

Numeric Comparison
Consider a simple boolean indexing operation. The list comprehension method has been timed against Series.ne (!=) and query. Here are the functions:

# Boolean indexing with Numeric value comparison.
df[df.A != df.B]                            # vectorized !=
df.query('A != B')                          # query (numexpr)
df[[x != y for x, y in zip(df.A, df.B)]]    # list comp

For simplicity, I have used the perfplot package to run all the timeit tests in this post. The timings for the operations above are below:

The list comprehension outperforms query for moderately sized N, and even outperforms the vectorized not equals comparison for tiny N. Unfortunately, the list comprehension scales linearly, so it does not offer much performance gain for larger N.

Note
It is worth mentioning that much of the benefit of list comprehension come from not having to worry about the index alignment, but this means that if your code is dependent on indexing alignment, this will break. In some cases, vectorised operations over the underlying NumPy arrays can be considered as bringing in the "best of both worlds", allowing for vectorisation without all the unneeded overhead of the pandas functions. This means that you can rewrite the operation above as

df[df.A.values != df.B.values]

其优于大熊猫和列表理解等价物:

NumPy矢量化超出了本文的范围,但如果性能很重要,那绝对值得考虑.

价值计数
再举一个例子 - 这一次,使用另一个比for循环更快的 vanilla python构造- collections.Counter.常见的要求是计算值计数并将结果作为字典返回.这与做value_counts,np.unique以及Counter:

# Value Counts comparison.
ser.value_counts(sort=False).to_dict()           # value_counts
dict(zip(*np.unique(ser, return_counts=True)))   # np.unique
Counter(ser)                                     # Counter

结果更加明显,Counter对于更大范围的小N(~3500),两种矢量化方法都胜出.

Note
More trivia (courtesy @user2357112). The Counter is implemented with a C accelerator, so while it still has to work with python objects instead of the underlying C datatypes, it is still faster than a for loop. Python power!

Of course, the take away from here is that the performance depends on your data and use case. The point of these examples is to convince you not to rule out these solutions as legitimate options. If these still don't give you the performance you need, there is always cython and numba. Let's add this test into the mix.

from numba import njit, prange

@njit(parallel=True)
def get_mask(x, y):
    result = [False] * len(x)
    for i in prange(len(x)):
        result[i] = x[i] != y[i]

    return np.array(result)

df[get_mask(df.A.values, df.B.values)] # numba

Numba offers JIT compilation of loopy python code to very powerful vectorized code. Understanding how to make numba work involves a learning curve.

Operations with Mixed/object dtypes

String-based Comparison
Revisiting the filtering example from the first section, what if the columns being compared are strings? Consider the same 3 functions above, but with the input DataFrame cast to string.

# Boolean indexing with string value comparison.
df[df.A != df.B]                            # vectorized !=
df.query('A != B')                          # query (numexpr)
df[[x != y for x, y in zip(df.A, df.B)]]    # list comp

So, what changed? The thing to note here is that string operations are inherently difficult to vectorize. Pandas treats strings as objects, and all operations on objects fall back to a slow, loopy implementation.

Now, because this loopy implementation is surrounded by all the overhead mentioned above, there is a constant magnitude difference between these solutions, even though they scale the same.

When it comes to operations on mutable/complex objects, there is no comparison. List comprehension outperforms all operations involving dicts and lists.

Accessing Dictionary Value(s) by Key
Here are timings for two operations that extract a value from a column of dictionaries: map and the list comprehension. The setup is in the Appendix, under the heading "Code Snippets".

# Dictionary value extraction.
ser.map(operator.itemgetter('value'))     # map
pd.Series([x.get('value') for x in ser])  # list comprehension

Positional List Indexing
Timings for 3 operations that extract the 0th element from a list of columns (handling exceptions), map, str.get accessor method, and the list comprehension:

# List positional indexing. 
def get_0th(lst):
    try:
        return lst[0]
    # Handle empty lists and NaNs gracefully.
    except (IndexError, TypeError):
        return np.nan

ser.map(get_0th)                                          # map
ser.str[0]                                                # str accessor
pd.Series([x[0] if len(x) > 0 else np.nan for x in ser])  # list comp
pd.Series([get_0th(x) for x in ser])                      # list comp safe

Note
If the index matters, you would want to do:

pd.Series([...], index=ser.index)

When reconstructing the series.

List Flattening
A final example is flattening lists. This is another common problem, and demonstrates just how powerful pure python is here.

# Nested list flattening.
pd.DataFrame(ser.tolist()).stack().reset_index(drop=True)  # stack
pd.Series(list(chain.from_iterable(ser.tolist())))         # itertools.chain
pd.Series([y for x in ser for y in x])                     # nested list comp

Both itertools.chain.from_iterable and the nested list comprehension are pure python constructs, and scale much better than the stack solution.

These timings are a strong indication of the fact that pandas is not equipped to work with mixed dtypes, and that you should probably refrain from using it to do so. Wherever possible, data should be present as scalar values (ints/floats/strings) in separate columns.

Lastly, the applicability of these solutions depend widely on your data. So, the best thing to do would be to test these operations on your data before deciding what to go with. Notice how I have not timed apply on these solutions, because it would skew the graph (yes, it's that slow).

Regex Operations, and .str Accessor Methods

Pandas can apply regex operations such as str.contains, str.extract, and str.extractall, as well as other "vectorized" string operations (such as str.split, str.find,str.translate`, and so on) on string columns. These functions are slower than list comprehensions, and are meant to be more convenience functions than anything else.

It is usually much faster to pre-compile a regex pattern and iterate over your data with re.compile (also see Is it worth using Python's re.compile?). The list comp equivalent to str.contains looks something like this:

p = re.compile(...)
ser2 = pd.Series([x for x in ser if p.search(x)])

Or,

ser2 = ser[[bool(p.search(x)) for x in ser]]

If you need to handle NaNs, you can do something like

ser[[bool(p.search(x)) if pd.notnull(x) else False for x in ser]]

The list comp equivalent to str.extract (without groups) will look something like:

df['col2'] = [p.search(x).group(0) for x in df['col']]

If you need to handle no-matches and NaNs, you can use a custom function (still faster!):

def matcher(x):
    m = p.search(str(x))
    if m:
        return m.group(0)
    return np.nan

df['col2'] = [matcher(x) for x in df['col']]

The matcher function is very extensible. It can be fitted to return a list for each capture group, as needed. Just extract query the group or groups attribute of the matcher object.

For str.extractall, change p.search to p.findall.

String Extraction
Consider a simple filtering operation. The idea is to extract 4 digits if it is preceded by an upper case letter.

# Extracting strings.
p = re.compile(r'(?<=[A-Z])(\d{4})')
def matcher(x):
    m = p.search(x)
    if m:
        return m.group(0)
    return np.nan

ser.str.extract(r'(?<=[A-Z])(\d{4})', expand=False)   #  str.extract
pd.Series([matcher(x) for x in ser])                  #  list comprehension

More Examples
Full disclosure - I am the author (in part or whole) of these posts listed below.

• Fast punctuation removal with pandas

• String concatenation of two pandas columns

• Remove unwanted parts from strings in a column

• Replace all but the last occurrence of a character in a dataframe

Conclusion

As shown from the examples above, iteration shines when working with small rows of DataFrames, mixed datatypes, and regular expressions.

The speedup you get depends on your data and your problem, so your mileage may vary. The best thing to do is to carefully run tests and see if the payout is worth the effort.

The "vectorized" functions shine in their simplicity and readability, so if performance is not critical, you should definitely prefer those.

Another side note, certain string operations deal with constraints that favour the use of NumPy. Here are two examples where careful NumPy vectorization outperforms python:

• Create new column with incremental values in a faster and efficient way - Answer by Divakar

• Fast punctuation removal with pandas - Answer by Paul Panzer

Additionally, sometimes just operating on the underlying arrays via .values as opposed to on the Series or DataFrames can offer a healthy enough speedup for most usual scenarios (see the Note in the Numeric Comparison section above). So, for example df[df.A.values != df.B.values] would show instant performance boosts over df[df.A != df.B]. Using .values may not be appropriate in every situation, but it is a useful hack to know.

As mentioned above, it's up to you to decide whether these solutions are worth the trouble of implementing.

Appendix: Code Snippets

import perfplot  
import operator 
import pandas as pd
import numpy as np
import re

from collections import Counter
from itertools import chain

# Boolean indexing with Numeric value comparison.
perfplot.show(
    setup=lambda n: pd.DataFrame(np.random.choice(1000, (n, 2)), columns=['A','B']),
    kernels=[
        lambda df: df[df.A != df.B],
        lambda df: df.query('A != B'),
        lambda df: df[[x != y for x, y in zip(df.A, df.B)]],
        lambda df: df[get_mask(df.A.values, df.B.values)]
    ],
    labels=['vectorized !=', 'query (numexpr)', 'list comp', 'numba'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N'
)

# Value Counts comparison.
perfplot.show(
    setup=lambda n: pd.Series(np.random.choice(1000, n)),
    kernels=[
        lambda ser: ser.value_counts(sort=False).to_dict(),
        lambda ser: dict(zip(*np.unique(ser, return_counts=True))),
        lambda ser: Counter(ser),
    ],
    labels=['value_counts', 'np.unique', 'Counter'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',
    equality_check=lambda x, y: dict(x) == dict(y)
)

# Boolean indexing with string value comparison.
perfplot.show(
    setup=lambda n: pd.DataFrame(np.random.choice(1000, (n, 2)), columns=['A','B'], dtype=str),
    kernels=[
        lambda df: df[df.A != df.B],
        lambda df: df.query('A != B'),
        lambda df: df[[x != y for x, y in zip(df.A, df.B)]],
    ],
    labels=['vectorized !=', 'query (numexpr)', 'list comp'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',
    equality_check=None
)

# Dictionary value extraction.
ser1 = pd.Series([{'key': 'abc', 'value': 123}, {'key': 'xyz', 'value': 456}])
perfplot.show(
    setup=lambda n: pd.concat([ser1] * n, ignore_index=True),
    kernels=[
        lambda ser: ser.map(operator.itemgetter('value')),
        lambda ser: pd.Series([x.get('value') for x in ser]),
    ],
    labels=['map', 'list comprehension'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',
    equality_check=None
)

# List positional indexing. 
ser2 = pd.Series([['a', 'b', 'c'], [1, 2], []])        
perfplot.show(
    setup=lambda n: pd.concat([ser2] * n, ignore_index=True),
    kernels=[
        lambda ser: ser.map(get_0th),
        lambda ser: ser.str[0],
        lambda ser: pd.Series([x[0] if len(x) > 0 else np.nan for x in ser]),
        lambda ser: pd.Series([get_0th(x) for x in ser]),
    ],
    labels=['map', 'str accessor', 'list comprehension', 'list comp safe'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',
    equality_check=None
)

# Nested list flattening.
ser3 = pd.Series([['a', 'b', 'c'], ['d', 'e'], ['f', 'g']])
perfplot.show(
    setup=lambda n: pd.concat([ser2] * n, ignore_index=True),
    kernels=[
        lambda ser: pd.DataFrame(ser.tolist()).stack().reset_index(drop=True),
        lambda ser: pd.Series(list(chain.from_iterable(ser.tolist()))),
        lambda ser: pd.Series([y for x in ser for y in x]),
    ],
    labels=['stack', 'itertools.chain', 'nested list comp'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',    
    equality_check=None

)

# Extracting strings.
ser4 = pd.Series(['foo xyz', 'test A1234', 'D3345 xtz'])
perfplot.show(
    setup=lambda n: pd.concat([ser4] * n, ignore_index=True),
    kernels=[
        lambda ser: ser.str.extract(r'(?<=[A-Z])(\d{4})', expand=False),
        lambda ser: pd.Series([matcher(x) for x in ser])
    ],
    labels=['str.extract', 'list comprehension'],
    n_range=[2**k for k in range(0, 15)],
    xlabel='N',
    equality_check=None
)

Answer 2

简而言之

• for循环+iterrows非常慢。开销在大约 1k 行上并不显着，但在 10k+ 行上则很明显。
• for 循环 +比oritertuples快得多。iterrowsapply
• 矢量化通常比itertuples

基准