计算pytorch标准化(Normalize)所需要数据集的均值和方差

pytorch做标准化利用transforms.Normalize(mean_vals, std_vals)，其中常用数据集的均值方差有：

if 'coco' in args.dataset:
    mean_vals = [0.471, 0.448, 0.408]
    std_vals = [0.234, 0.239, 0.242]
elif 'imagenet' in args.dataset:
    mean_vals = [0.485, 0.456, 0.406]
    std_vals = [0.229, 0.224, 0.225]

计算自己数据集图像像素的均值方差： 

import numpy as np
import cv2
import random

# calculate means and std
train_txt_path = './train_val_list.txt'

CNum = 10000     # 挑选多少图片进行计算

img_h, img_w = 32, 32
imgs = np.zeros([img_w, img_h, 3, 1])
means, stdevs = [], []

with open(train_txt_path, 'r') as f:
    lines = f.readlines()
    random.shuffle(lines)   # shuffle , 随机挑选图片

    for i in tqdm_notebook(range(CNum)):
        img_path = os.path.join('./train', lines[i].rstrip().split()[0])

        img = cv2.imread(img_path)
        img = cv2.resize(img, (img_h, img_w))
        img = img[:, :, :, np.newaxis]
        
        imgs = np.concatenate((imgs, img), axis=3)
#         print(i)

imgs = imgs.astype(np.float32)/255.


for i in tqdm_notebook(range(3)):
    pixels = imgs[:,:,i,:].ravel()  # 拉成一行
    means.append(np.mean(pixels))
    stdevs.append(np.std(pixels))

# cv2 读取的图像格式为BGR，PIL/Skimage读取到的都是RGB不用转
means.reverse() # BGR --> RGB
stdevs.reverse()

print("normMean = {}".format(means))
print("normStd = {}".format(stdevs))
print('transforms.Normalize(normMean = {}, normStd = {})'.format(means, stdevs))