Python csv.DictReader内存不足

时间:2019-01-04 16:01:30

标签: python python-3.x csv file-read

我想基于时间戳对csv文件中的值进行排序并将其打印到另一个文件,但是对于具有多行的文件,python的内存不足(读取文件时)。 我可以做些什么使它更有效吗?还是应该使用csv.DictReader之后的其他方法?

import csv, sys
import datetime
from pathlib import Path

localPath = "C:/MyPath"


    # data variables 
dataDir = localPath + "data/" dataExtension = ".dat" 

    pathlistData = Path(dataDir).glob('**/*'+ dataExtension)

    # Generated filename as date, Format: YYYY-DDDTHH
    generatedDataDir = localPath + "result/"
    #generatedExtension = ".dat"
    errorlog = 'errorlog.csv'

    fieldnames = ['TimeStamp', 'A', 'B', 'C', 'C', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L','M', 'N', 'O', 'P', 'Q', 'R'] 

    for dataPath in pathlistData:
        #stores our data in a dictionary
        dataDictionary = {}

        dataFileName = str(dataPath).replace('\\', '/')
        newFilePathString = dataFileName.replace(dataDir,generatedDataDir)

        with open(dataPath, 'r') as readFile:
            print(str("Reading data from " + dataFileName))
            keysAsDate = []#[datetime.datetime.strptime(ts, "%Y-%m-%d") for ts in timestamps]
            reader = csv.DictReader(readFile, fieldnames=fieldnames) 

            for row in reader:

                try:
                    timestamp = row['TimeStamp']
                    #create a key based on the timestamp
                    timestampKey = datetime.datetime.strptime(timestamp[0:16], "%Y-%jT%H:%M:%S")
                    #save this key as a date, used later for sorting
                    keysAsDate.append(timestampKey)
                    #save the row data in a dictionary
                    dataDictionary[timestampKey] = row

                except csv.Error as e:
                    sys.exit('file %s, line %d: %s' % (errorlog, reader.line_num, e))

            #sort the keys
            keysAsDate.sort()
        readFile.close()

        with open(newFilePathString, 'w') as writeFile:
            writer = csv.DictWriter(writeFile, fieldnames=fieldnames, lineterminator='\n')
            print(str("Writing data to " + newFilePathString))
            #loop over the sorted keys
            for idx in range(0, len(keysAsDate)):

                #get the row from our data dictionary 
                writeRow = dataDictionary[keysAsDate[idx]]
                #print(dataDictionary[keysAsDate[key]])
                writer.writerow(writeRow)
                if idx%30000 == 0:
                    print("Writing to new file: " + str(int(idx/len(keysAsDate) * 100)) + "%")


        print(str("Finished writing to file: " + newFilePathString))

        writeFile.close()

更新:我使用了大熊猫,将大文件分成了较小的块,可以分别对其进行排序。 如果我一个接一个地附加文件,这目前还不能解决值放得过大的问题。

for dataPath in pathlistData:

dataFileName = str(dataPath).replace('\\', '/')
#newFilePathString = dataFileName.replace(dataDir,generatedDataDir)


print(str("Reading data from " + dataFileName))
#divide our large data frame into smaller data frame chunks
#so we can sort the content in memory
for df_chunk in pd.read_csv(dataFileName, header = None, chunksize = chunk_size, names = fieldnames):
    dataDictionary = {}
    dataDictionary.clear()

    for idx in range(0, chunk_size):
        #print(df_chunk[idx:idx+1])
        row = df_chunk[idx:idx+1]
        dataDictionary = df_chunk.sort_values(['TimeStamp'], ascending=True)
    firstTimeStampInChunk = dataDictionary[0:1]['TimeStamp']
    #print("first: " + firstTimeStampInChunk)
    lastTimeStampInChunk = dataDictionary[chunk_size-1:chunk_size]['TimeStamp']
    #print("last: " + lastTimeStampInChunk)

    timestampStr = str(firstTimeStampInChunk)[chunk_shift:timestamp_size+chunk_shift] + str(lastTimeStampInChunk)[chunk_shift:timestamp_size+chunk_shift]
    tempFilePathString = str(timestampStr + dataExtension).replace(':', '_').replace('\\', '/')
    dataDictionary.to_csv('temp/'+tempFilePathString, header = None, index=False)

# data variables
tempDataDir = localPath + "temp/"
tempPathlistData = Path(tempDataDir).glob('**/*'+ dataExtension)

tempPathList = list(tempPathlistData)

我解决随机值问题的算法理论(无代码)是:

第1步-分成较小的块,其中“ chunk_size =要处理的最大行数除以2”

第2步-按顺序循环浏览文件,一次合并两个文件,然后将它们排序在一起,然后再次拆分,以使文件不超过chunk_size。

第3步-向后循环,一次合并两个文件并对它们进行排序,然后再次拆分,以使文件不大于chunk_size。

第4步-现在,所有放错了地方的低值都应该传播到最低的部分,而所有放错了地方的高值都应该传播到最高的部分。依次附加文件!

缺点;时间上的复杂性一点都不可取,如果我没记错的话,基本上是O(N ^ 2)

2 个答案:

答案 0 :(得分:1)

尝试使用熊猫csv阅读器,这非常有效。 (https://pandas.pydata.org/pandas-docs/stable/generated/pandas.read_csv.html)。您可以使用https://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.to_dict.html

在熊猫和字典之间轻松转换

答案 1 :(得分:0)

您解释了内存中排序对您不起作用,因为文件大小超出了内存大小。至少有两种方法可以解决此问题。两者都依靠做更多的文件I / O。

  1. 将长记录压缩到单个内存有效的文件偏移量。在读取每个记录(或对行长度求和)时调用tell(),并在内存中仅保留时间戳和文件偏移量。按时间戳对偏移量进行排序。在遍历已排序的元组时,反复调用seek(),对记录进行随机读取,然后将其附加到输出文件中。
  2. 更好的方法是让/usr/bin/sort进行外部合并排序。 Windows用户可以从https://git-scm.com/download/获取coreutils GNU排序。使用subprocess模块来调用它。