为什么我的Python脚本比它的R相当慢？

Question

备注：这个问题涵盖了为什么这个脚本太慢了。但是，如果你更喜欢改善某种东西的人，你可以看看my post on CodeReview which aims to improve the performance。

我正在研究一个打开纯文本文件（.lst）的项目。

的文件名（fileName）的名字很重要，因为我会从中提取node（例如abessijn）和component（如WR-P-E-A）转换成数据帧。示例：

abessijn.WR-P-E-A.lst 
A-bom.WR-P-E-A.lst 
acroniem.WR-P-E-C.lst 
acroniem.WR-P-E-G.lst 
adapter.WR-P-E-A.lst 
adapter.WR-P-E-C.lst 
adapter.WR-P-E-G.lst 

每个文件由一行或多行组成。每行包含一个句子（在<sentence>标签内）。实施例（abessijn.WR-P-E-A.lst）

/home/nobackup/SONAR/COMPACT/WR-P-E-A/WR-P-E-A0000364.data.ids.xml: <sentence>Vooral mijn abessijn ruikt heerlijk kruidig .. :))</sentence> 
/home/nobackup/SONAR/COMPACT/WR-P-E-A/WR-P-E-A0000364.data.ids.xml: <sentence>Mijn abessijn denkt daar heel anders over .. :)) Maar mijn kinderen richt ik ook niet af , zit niet in mijn bloed .</sentence> 

从每个线I提取句子，做一些小的修改它，并调用它sentence。接下来是一个名为leftContext的元素，它将node（例如abessijn）之间的第一部分与它来自的句子进行比较。最后，从leftContext我得到的前词，这是nodesentence之前的词，或leftContext中的最右边的词（带有一些限制，例如用连字符形成的化合物的选项）。示例：

ID | filename    | node | component | precedingWord  | leftContext        | sentence 
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- 
1 adapter.WR-P-P-F.lst adapter WR-P-P-F aanpassingseenheid Een aanpassingseenheid (      Een aanpassingseenheid (adapter) , 
2 adapter.WR-P-P-F.lst adapter WR-P-P-F toestel    Het toestel (        Het toestel (adapter) draagt zorg voor de overbrenging van gegevens 
3 adapter.WR-P-P-F.lst adapter WR-P-P-F de     de aansluiting tussen de sensor en de   de aansluiting tussen de sensor en de adapter , 
4 airbag.WS-U-E-A.lst airbag WS-U-E-A den     ja voor den         ja voor den airbag op te pompen eh :p 
5 airbag.WS-U-E-A.lst airbag WS-U-E-A ne     Dobby , als ze valt heeft ze dan wel al ne Dobby , als ze valt heeft ze dan wel al ne airbag hee 

该数据帧作为dataset.csv导出。

之后，我的项目的意图就在眼前：我创建了一个频率表，其中考虑了node和precedingWord。从变量I定义neuter和non_neuter，e.g（在Python）

neuter = ["het", "Het"] 
non_neuter = ["de","De"] 

和休止类别unspecified。当precedingWord是列表中的项目时，将其分配给变量。频率表输出示例：

node | neuter | nonNeuter | unspecified 
------------------------------------------------- 
A-bom  0   4    2 
acroniem 3   0    2 
act   3   2    1 

频率列表导出为频率.csv。

---

我从R开始，考虑到稍后我会对频率进行一些统计分析。我的当前R脚本（也可作为paste）：

# --- 
# STEP 0: Preparations 
    start_time <- Sys.time() 
    ## 1. Set working directory in R 
    setwd("") 

    ## 2. Load required library/libraries 
    library(dplyr) 
    library(mclm) 
    library(stringi) 

    ## 3. Create directory where we'll save our dataset(s) 
    dir.create("../R/dataset", showWarnings = FALSE) 


# --- 
# STEP 1: Loop through files, get data from the filename 

    ## 1. Create first dataframe, based on filename of all files 
    files <- list.files(pattern="*.lst", full.names=T, recursive=FALSE) 
    d <- data.frame(fileName = unname(sapply(files, basename)), stringsAsFactors = FALSE) 

    ## 2. Create additional columns (word & component) based on filename 
    d$node <- sub("\\..+", "", d$fileName, perl=TRUE) 
    d$node <- tolower(d$node) 
    d$component <- gsub("^[^\\.]+\\.|\\.lst$", "", d$fileName, perl=TRUE) 


# --- 
# STEP 2: Loop through files again, but now also through its contents 
# In other words: get the sentences 

    ## 1. Create second set which is an rbind of multiple frames 
    ## One two-column data.frame per file 
    ## First column is fileName, second column is data from each file 
    e <- do.call(rbind, lapply(files, function(x) { 
     data.frame(fileName = x, sentence = readLines(x, encoding="UTF-8"), stringsAsFactors = FALSE) 
    })) 

    ## 2. Clean fileName 
    e$fileName <- sub("^\\.\\/", "", e$fileName, perl=TRUE) 

    ## 3. Get the sentence and clean 
    e$sentence <- gsub(".*?<sentence>(.*?)</sentence>", "\\1", e$sentence, perl=TRUE) 
    e$sentence <- tolower(e$sentence) 
     # Remove floating space before/after punctuation 
     e$sentence <- gsub("\\s(?:(?=[.,:;?!) ])|(?<=\\())", "\\1", e$sentence, perl=TRUE) 
    # Add space after triple dots ... 
     e$sentence <- gsub("\\.{3}(?=[^\\s])", "... ", e$sentence, perl=TRUE) 

    # Transform HTML entities into characters 
    # It is unfortunate that there's no easier way to do this 
    # E.g. Python provides the HTML package which can unescape (decode) HTML 
    # characters 
     e$sentence <- gsub("&apos;", "'", e$sentence, perl=TRUE) 
     e$sentence <- gsub("&amp;", "&", e$sentence, perl=TRUE) 
     # Avoid R from wrongly interpreting ", so replace by single quotes 
     e$sentence <- gsub("&quot;|\"", "'", e$sentence, perl=TRUE) 

     # Get rid of some characters we can't use such as ³ and ¾ 
     e$sentence <- gsub("[^[:graph:]\\s]", "", e$sentence, perl=TRUE) 


# --- 
# STEP 3: 
# Create final dataframe 

    ## 1. Merge d and e by common column name fileName 
    df <- merge(d, e, by="fileName", all=TRUE) 

    ## 2. Make sure that only those sentences in which df$node is present in df$sentence are taken into account 
    matchFunction <- function(x, y) any(x == y) 
    matchedFrame <- with(df, mapply(matchFunction, node, stri_split_regex(sentence, "[ :?.,]"))) 
    df <- df[matchedFrame, ] 

    ## 3. Create leftContext based on the split of the word and the sentence 
    # Use paste0 to make sure we are looking for the node, not a compound 
    # node can only be preceded by a space, but can be followed by punctuation as well 
    contexts <- strsplit(df$sentence, paste0("(^|)", df$node, "(|[!\",.:;?})\\]])"), perl=TRUE) 
    df$leftContext <- sapply(contexts, `[`, 1) 

    ## 4. Get the word preceding the node 
    df$precedingWord <- gsub("^.*\\b(?<!-)(\\w+(?:-\\w+)*)[^\\w]*$","\\1", df$leftContext, perl=TRUE) 

    ## 5. Improve readability by sorting columns 
    df <- df[c("fileName", "component", "precedingWord", "node", "leftContext", "sentence")] 

    ## 6. Write dataset to dataset dir 
    write.dataset(df,"../R/dataset/r-dataset.csv") 


# --- 
# STEP 4: 
# Create dataset with frequencies 

    ## 1. Define neuter and nonNeuter classes 
    neuter <- c("het") 
    non.neuter<- c("de") 

    ## 2. Mutate df to fit into usable frame 
    freq <- mutate(df, gender = ifelse(!df$precedingWord %in% c(neuter, non.neuter), "unspecified", 
     ifelse(df$precedingWord %in% neuter, "neuter", "non_neuter"))) 

    ## 3. Transform into table, but still usable as data frame (i.e. matrix) 
    ## Also add column name "node" 
    freqTable <- table(freq$node, freq$gender) %>% 
     as.data.frame.matrix %>% 
     mutate(node = row.names(.)) 

    ## 4. Small adjustements 
    freqTable <- freqTable[,c(4,1:3)] 

    ## 5. Write dataset to dataset dir 
    write.dataset(freqTable,"../R/dataset/r-frequencies.csv") 


    diff <- Sys.time() - start_time # calculate difference 
    print(diff) # print in nice format 

然而，由于我使用的是大数据集（16500个文件，全部用多线）它似乎需要相当长。在我的系统上，整个过程花费了大约一个小时四分之一的时间。我认为自己应该有一种语言可以更快地完成这项工作，所以我去了，自学了一些Python，并在这里问了很多问题。

最后我想出了以下脚本（paste）。

import os, pandas as pd, numpy as np, regex as re 

from glob import glob 
from datetime import datetime 
from html import unescape 

start_time = datetime.now() 

# Create empty dataframe with correct column names 
columnNames = ["fileName", "component", "precedingWord", "node", "leftContext", "sentence" ] 
df = pd.DataFrame(data=np.zeros((0,len(columnNames))), columns=columnNames) 

# Create correct path where to fetch files 
subdir = "rawdata" 
path = os.path.abspath(os.path.join(os.getcwd(), os.pardir, subdir)) 

# "Cache" regex 
# See http://stackoverflow.com/q/452104/1150683 
p_filename = re.compile(r"[./\\]") 

p_sentence = re.compile(r"<sentence>(.*?)</sentence>") 
p_typography = re.compile(r" (?:(?=[.,:;?!) ])|(?<=\())") 
p_non_graph = re.compile(r"[^\x21-\x7E\s]") 
p_quote = re.compile(r"\"") 
p_ellipsis = re.compile(r"\.{3}(?=[^ ])") 

p_last_word = re.compile(r"^.*\b(?<!-)(\w+(?:-\w+)*)[^\w]*$", re.U) 

# Loop files in folder 
for file in glob(path+"\\*.lst"): 
    with open(file, encoding="utf-8") as f: 
     [n, c] = p_filename.split(file.lower())[-3:-1] 
     fn = ".".join([n, c]) 
     for line in f: 
      s = p_sentence.search(unescape(line)).group(1) 
      s = s.lower() 
      s = p_typography.sub("", s) 
      s = p_non_graph.sub("", s) 
      s = p_quote.sub("'", s) 
      s = p_ellipsis.sub("... ", s) 

      if n in re.split(r"[ :?.,]", s): 
       lc = re.split(r"(^|)" + n + "(|[!\",.:;?})\]])", s)[0] 

       pw = p_last_word.sub("\\1", lc) 

       df = df.append([dict(fileName=fn, component=c, 
            precedingWord=pw, node=n, 
            leftContext=lc, sentence=s)]) 
      continue 

# Reset indices 
df.reset_index(drop=True, inplace=True) 

# Export dataset 
df.to_csv("dataset/py-dataset.csv", sep="\t", encoding="utf-8") 

# Let's make a frequency list 
# Create new dataframe 

# Define neuter and non_neuter 
neuter = ["het"] 
non_neuter = ["de"] 

# Create crosstab 
df.loc[df.precedingWord.isin(neuter), "gender"] = "neuter" 
df.loc[df.precedingWord.isin(non_neuter), "gender"] = "non_neuter" 
df.loc[df.precedingWord.isin(neuter + non_neuter)==0, "gender"] = "rest" 

freqDf = pd.crosstab(df.node, df.gender) 

freqDf.to_csv("dataset/py-frequencies.csv", sep="\t", encoding="utf-8") 

# How long has the script been running? 
time_difference = datetime.now() - start_time 
print("Time difference of", time_difference) 

在确定两个脚本的输出是相同的之后，我想我会让他们参加测试。

我使用四核处理器和8 GB RAM在Windows 10 64位上运行。对于R我使用的是RGUI 64位3.2.2，Python在版本3.4.3（Anaconda）上运行，并在Spyder中执行。请注意，我以32位运行Python，因为我希望将来使用nltk module，并且阻止用户使用64位。

我发现R约在55分钟内完成。但是Python已经运行了两个小时，并且我可以在变量浏览器中看到它只在business.wr-p-p-g.lst（文件按字母顺序排序）。 waaaaayyyy慢！

所以我做了一个测试用例，看看两个脚本如何用一个小得多的数据集执行。我花了大约100个文件（而不是16,500个）并运行脚本。再一次，R要快得多。 R大概在2秒内完成，17中的Python！

看到Python的目标是让所有事情都变得更加顺利，我感到困惑。我读Python很快（而且R很慢），所以我哪里出错了？问题是什么？ Python在阅读文件和行，或者在执行正则表达式时速度较慢？或者R只是更好地处理数据帧，不能被熊猫打败？ 或是我的代码简单地严重优化，应该Python确实是胜利者？

因此，我的问题是：为什么Python在这种情况下比R慢，以及 - 如果可能的话 - 我们如何改进Python来发光？

每个愿意试一试脚本的人都可以下载我使用的测试数据here。请在下载文件时提醒我。

Answer 1

你要做的就是调用一个循环DataFrame.append方法最效率极其低下的东西，即

df = pandas.DataFrame(...) 
for file in files: 
    ... 
    for line in file: 
     ... 
     df = df.append(...) 

NumPy的数据结构设计时考虑了函数式编程，因此这种操作并不意味着在使用迭代的必要方式，因为调用不会在原地改变你的数据帧，但会创建一个新的，导致时间和内存复杂度的巨大增加。如果您确实想要使用数据帧，请在list中累积行并将其传递给DataFrame构造函数，例如，

pre_df = [] 
for file in files: 
    ... 
    for line in file: 
     ... 
     pre_df.append(processed_line) 

df = pandas.DataFrame(pre_df, ...) 

这是因为它会引入最少的修改你的代码最简单的方法。但是更好的（并且计算上很漂亮）的方法是想办法如何生成你的数据集。这可以通过将工作流分解为离散函数（从函数式编程风格的意义上来说）并使用懒惰生成器表达式和/或高阶函数来组成它们来轻松实现。然后，您可以使用生成的生成器来构建数据框，例如

df = pandas.DataFrame.from_records(processed_lines_generator, columns=column_names, ...) 

至于在一次运行中读取多个文件，您可能需要阅读this。

P.S.

如果你有性能问题，你应该在试图优化它之前对你的代码进行剖析。