1. 图像深度学习任务中,一般首先要对输入图像进行归一化,这就需要训练集图像各颜色通道的均值与标准差(mean,std)。关于为什么要进行归一化,可以参考这两位博主的解释:
https://blog.csdn.net/u011267996/article/details/80503907
https://blog.csdn.net/zhou4411781/article/details/101205535
2. 我拿到的任务的数据集图像数量很多40w+,但是都不大,处理这类IO密集型的任务Python的多线程还是可以考虑的。上代码(在linux上跑的,windows请自测):
from itertools import repeat
import os
from multiprocessing.pool import ThreadPool
from pathlib import Path
from PIL import Image
import numpy as np
from tqdm import tqdm
NUM_THREADS = os.cpu_count()
def calc_channel_sum(img_path): # 计算均值的辅助函数,统计单张图像颜色通道和,以及像素数量
img = np.array(Image.open(img_path).convert('RGB')) / 255.0 # 准换为RGB的array形式
h, w, _ = img.shape
pixel_num = h*w
channel_sum = img.sum(axis=(0, 1)) # 各颜色通道像素求和
return channel_sum, pixel_num
def calc_channel_var(img_path, mean): # 计算标准差的辅助函数
img = np.array(Image.open(img_path).convert('RGB')) / 255.0
channel_var = np.sum((img - mean) ** 2, axis=(0, 1))
return channel_var
if __name__ == '__main__':
train_path = Path('../data/train')
img_f = list(train_path.rglob('*.jpg'))
n = len(img_f)
result = ThreadPool(NUM_THREADS).imap(calc_channel_sum, img_f) # 多线程计算
channel_sum = np.zeros(3)
cnt = 0
pbar = tqdm(enumerate(result), total=n)
for i, x in pbar:
channel_sum += x[0]
cnt += x[1]
mean = channel_sum / cnt
print("R_mean is %f, G_mean is %f, B_mean is %f" % (mean[0], mean[1], mean[2]))
result = ThreadPool(NUM_THREADS).imap(lambda x: calc_channel_var(*x), zip(img_f, repeat(mean)))
channel_sum = np.zeros(3)
pbar = tqdm(enumerate(result), total=n)
for i, x in pbar:
channel_sum += x
var = np.sqrt(channel_sum / cnt)
print("R_var is %f, G_var is %f, B_var is %f" % (var[0], var[1], var[2]))
3. 关于python多进程、多线程的编程可以参靠以下链接: