Python での頻繁なファイルの書き込みを高速化する方法

# -*-coding:utf-8-*-
import numpy as np
from glob import glob
import math
import os
import torch
from tqdm import tqdm
import multiprocessing
label_path = '/home/ying/data/shiyongjie/distortion_datasets/new_distortion_dataset/train/label.txt'
file_path = '/home/ying/data/shiyongjie/distortion_datasets/new_distortion_dataset/train/distortion_image'
save_path = '/home/ying/data/shiyongjie/distortion_datasets/new_distortion_dataset/train/flow_field'
r_d_max = 128
image_index = 0
txt_file = open(label_path)
file_list = txt_file.readlines()
txt_file.close()
file_label = {}
for i in file_list:
    i = i.split()
    file_label[i[0]] = i[1]
r_d_max = 128
eps = 1e-32
H = 256
W = 256
def generate_flow_field(image_list):
    for image_file_path in ((image_list)):
        pixel_flow = np.zeros(shape=tuple([256, 256, 2]))  # 按照pytorch中的grid来写
        image_file_name = os.path.basename(image_file_path)
        # print(image_file_name)
        k = float(file_label[image_file_name])*(-1)*1e-7
        # print(k)
        r_u_max = r_d_max/(1+k*r_d_max**2)  # 计算出畸变校正之后的对角线的理论长度
        scale = r_u_max/128  # 将这个长度压缩到256的尺寸，会有一个scale，实际上这里写128*sqrt(2)可能会更加直观
        for i_u in range(256):
            for j_u in range(256):
                x_u = float(i_u - 128)
                y_u = float(128 - j_u)
                theta = math.atan2(y_u, x_u)
                r = math.sqrt(x_u ** 2 + y_u ** 2)
                r = r * scale  # 实际上得到的r，即没有resize到256×256的图像尺寸size，并且带入公式中
                r_d = (1.0 - math.sqrt(1 - 4.0 * k * r ** 2)) / (2 * k * r + eps)  # 对应在原图（畸变图）中的r
                x_d = int(round(r_d * math.cos(theta)))
                y_d = int(round(r_d * math.sin(theta)))
                i_d = int(x_d + W / 2.0)
                j_d = int(H / 2.0 - y_d)
                if i_d < W and i_d >= 0 and j_d < H and j_d >= 0:  # 只有求的的畸变点在原图中的时候才进行赋值
                    value1 = (i_d - 128.0)/128.0
                    value2 = (j_d - 128.0)/128.0
                    pixel_flow[j_u, i_u, 0] = value1  # mesh中存储的是对应的r的比值，在进行畸变校正的时候，给定一张这样的图，进行找像素即可
                    pixel_flow[j_u, i_u, 1] = value2
# 保存成array格式
        saved_image_file_path = os.path.join(save_path, image_file_name.split(&#39;.&#39;)[0] + &#39;.npy&#39;)
        pixel_flow = pixel_flow.astype(&#39;f2&#39;)  # 将数据的格式转换成float16类型， 节省空间
        # print(saved_image_file_path)
        # print(pixel_flow)
        np.save(saved_image_file_path, pixel_flow)
    return
if __name__ == &#39;__main__&#39;:
    file_list = glob(file_path + &#39;/*.JPEG&#39;)
    m = 32
    n = int(math.ceil(len(file_list) / float(m)))  # 向上取整
    result = []
    pool = multiprocessing.Pool(processes=m)  # 32进程
    for i in range(0, len(file_list), n):
        result.append(pool.apply_async(generate_flow_field, (file_list[i: i+n],)))
    pool.close()
    pool.join()

if __name__ == '__main__':
    file_list = glob(file_path + '/*.JPEG')  # 将文件夹下所有的JPEG文件列成一个list
    m = 32  # 假设CPU有32个核心
    n = int(math.ceil(len(file_list) / float(m)))  # 每一个核心需要处理的list的数目
    result = []
    pool = multiprocessing.Pool(processes=m)  # 开32线程的线程池
    for i in range(0, len(file_list), n):
        result.append(pool.apply_async(generate_flow_field, (file_list[i: i+n],)))  # 对每一个list都用上面我们定义的函数进行处理
    pool.close()  # 处理结束之后，关闭线程池
    pool.join()

pool = multiprocessing.Pool(processes=m)  # 开32线程的线程池

result.append(pool.apply_async(generate_flow_field, (file_list[i: i+n],)))  # 对每一个list都用上面我们定义的函数进行处理