Question regarding CIFAR10 standard deviation

[ssgosh]

I computed the mean and standard deviation of cifar dataset (train). While the means match, the standard deviations do not:

mean = [0.4913996458053589, 0.48215845227241516, 0.44653093814849854]
std = [0.2470322549343109, 0.24348513782024384, 0.26158788800239563]

Relevant line from main.py:

https://github.com/kuangliu/pytorch-cifar/blob/49b7aa97b0c12fe0d4054e670403a16b6b834ddd/main.py#L34

I am wondering what is the cause of this discrepancy. Relevant code that I used to compute these stats:

    datasets = [ torchvision.datasets.CIFAR10('./data', download=True, train=True, transform=torchvision.transforms.ToTensor()) ]
    channels = 3
    total = torch.zeros(channels)
    total_sq = torch.zeros(channels)
    num = 0
    for ds in datasets:
        dl = DataLoader(ds, batch_size=1000)
        for images, targets in dl:
            total += torch.sum(images, (0, 2, 3))
            total_sq += torch.sum(images * images, (0, 2, 3))
            num += images.shape[0] * images.shape[2] * images.shape[3]

    mean = total / num
    mean_sq = total_sq / num
    std = torch.sqrt(mean_sq - mean**2)

    print(f'mean = {mean.tolist()}, std = {std.tolist()}')