Standard deviation for transforms.Normalize about pytorch-cifar

[kuangliu]

See get_mean_and_std(dataset) in utils.py. It's OK to use a different std for training.

[wenwei202]

Should be [0.2470, 0.2435, 0.2616]. See here.

[lishaofeng]

I think the divide operation in get_mean_and_std(dataset) function should be 255. while not the len(dataset)

[renyiryry]

I think the divide operation in get_mean_and_std(dataset) function should be 255. while not the len(dataset)

It should be len(dataset) because the code is summing up the std for all images. Dividing by len(dataset) gives the averaged std.

[renyiryry]

Should be [0.2470, 0.2435, 0.2616]. See here.

Not really. The way you computed gives the std of the tensor of size [50000, 32, 32]. However, the desired quantity is the "averaged" std across all the images. Also, when computing the std for a single image, degree of freedom should be 1. (See https://github.com/kuangliu/pytorch-cifar/blob/master/utils.py)