cifar100-channel_wise_mean_and_std.py

# CIFAR-100

import numpy as np
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from torchvision.datasets import CIFAR100


train_transform = transforms.Compose(
    [
        transforms.ToTensor(),
    ])

train_set = CIFAR100(
    root='~/data',
    train=True,
    download=True,
    transform=train_transform)

train_loader = DataLoader(
    train_set,
    batch_size=50_000,
    shuffle=True,
)

data = iter(train_loader).next()
print(
    data[0].mean(dim=[0, 2, 3]).numpy(),
    data[0].std(dim=[0, 2, 3]).numpy()
)    

Output: [0.5070779 0.48654884 0.44091937] [0.26733428 0.25643846 0.27615047]
So, I use these values for Normalize:

transforms.Normalize(
    [0.507077, 0.48654863, 0.4409177], [0.26733428, 0.25643846, 0.27615047]
)