数据归一化方法

Pytorch提供了函数torchvision.transforms.Normalize用于标准化图像数据取值，其计算公式如下

# torchvision.transforms.Normalize(mean, std, inplace=False)
output[channel] = (input[channel] - mean[channel]) / std[channel]

在实践过程中，发现有好几种均值和方差的推荐。

1、ToTensor

Normalize通常配合ToTensor一起使用。ToTensor的功能是将PIL Image或者numpy.ndarray格式的数据转换成tensor格式，同时将取值范围[0, 255]转换成[0.0, 1.0]。

2、transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))

最常用操作，将数据缩放到[-1, 1]之间。

3、transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))

上述均值和标准差来源于ImageNet数据集，如果使用PyTorch提供的预训练模型，推荐该设置

4、自定义均值和标准差

import torch
from tqdm import tqdm
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from HCP2BraTS_dataloader import datasetload
import albumentations as A
from albumentations.pytorch import ToTensorV2


def main():
    all_txt = '/data5/pengzhang/TPRO-main/datasets/HCP2BraTS/all_T2.txt'

    transform = A.Compose([
            A.Resize(height=256, width=256),
            ToTensorV2(transpose_mask=True),
        ])

    all_dataset = datasetload(all_txt, transform)
    data_loader = DataLoader(all_dataset, batch_size=24, shuffle=False)

    nb_samples = 0.
    channel_mean = torch.zeros(1)
    channel_std = torch.zeros(1)
    for img_name, image, cls_labels, gt_label in tqdm(iter(data_loader)):
        # scale image to be between 0 and 1
        N, C, H, W = image.shape[:4]
        data = image.view(N, C, -1).float()

        channel_mean += data.mean(2).sum(0)
        channel_std += data.std(2).sum(0)
        nb_samples += N

    channel_mean /= nb_samples
    channel_std /= nb_samples
    print(channel_mean, channel_std)


if __name__ == '__main__':
    main()

注意：不同图像转换操作会带来不一样标准差结果