简单拆分PDF文件,将一个PDF文件拆分成指定份数

清华大佬耗费三个月吐血整理的几百G的资源,免费分享!....>>>

from pyPdf import PdfFileWriter, PdfFileReader
import os
import time
import sys

def part_pdf(input_file, output_file, config_count, f_w, now, file_name):
    file1 = file(input_file, 'rb')
    pdf = PdfFileReader(file1)
    pdf_pages_len = len(pdf.pages)
    if config_count <= pdf_pages_len:
        ye = pdf_pages_len / config_count
        lst_ye = pdf_pages_len % config_count
        part_count = 0
        part_count_ye = 0
        for fen in range(config_count):
            part_count += 1
            if part_count == config_count:
                part_ye = ye + lst_ye
            else:
                part_ye = ye

            write_pdf(pdf, part_count_ye, part_count_ye+part_ye, fen, output_file)
            part_count_ye += ye
    else:
        f_w.writelines('time: '+now+' file name: '+file_name+' status: part_num > pdf pages  [error]\n')
        sys.exit(1)

def write_pdf(pdf, part_count_ye, part_count_ye_end, fen, output_file):
    out = PdfFileWriter()
    for pp in range(part_count_ye, part_count_ye_end):
        out.addPage(pdf.getPage(pp))
    ous = file(output_file+'_'+str(fen+1)+'.pdf', 'wb')
    out.write(ous)
    ous.close()

def pdf_main():
    f = open('configure.txt', 'r')
    f_w = open('pp_log.txt', 'a')
    now = time.strftime('%Y-%m-%d %H:%M:%S')
    for i in f:
        i_ = i.strip()
        aa = i_.split('=')[1]
        if i_.find('part_num=') != -1 and aa.isdigit():
            config_count = int(aa)
        else:
            f_w.writelines('time: '+now+' status: part_num in configure.txt is error  [error]\n')
            sys.exit(1)

    files = os.listdir('input_dir/')
    for each in files:
        input_file = 'input_dir/'+each
        file_name = input_file[input_file.index('/'):input_file.index('.')]
        output_file = 'output_dir/'+file_name
        part_pdf(input_file, output_file, config_count, f_w, now, file_name)
        f_w.writelines('time: '+now+' file name: '+file_name+' status: success\n')

pdf_main()