Why does pytorch's transforms.Normalize() not do the described action from the documentation?

Question

According to the documentation normalize is supposed to do (tensor - mean)/std, but it doesn't. Why?

Docs:

Normalize a tensor image with mean and standard deviation. Given mean: (mean[1],...,mean[n]) and std: (std[1],..,std[n]) for n channels, this transform will normalize each channel of the input torch.*Tensor i.e., output[channel] = (input[channel] - mean[channel]) / std[channel]

a = T.Tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]])
m = a.mean()
std = a.std()
print((m, std))
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)))
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)).mean())
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)).std())
a = (a - m)/std
m = a.mean()
std = a.std()
print((m, std))

Output:

(tensor(5.), tensor(2.7386))
tensor([[[[1.0150, 1.3802, 1.7453],
          [2.1105, 2.4756, 2.8408],
          [3.2059, 3.5711, 3.9362]]]])
tensor(2.4756)
tensor(1.0000)
(tensor(0.), tensor(1.0000))

The std is correct, but the mean is something random. What gives?

Answer 1

Mean value of your tensor is stored in variable m, not mean.

After replace m with mean on Line 2

a = T.Tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]])
mean = a.mean()
std = a.std()
print((mean, std))
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)))
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)).mean())
print(transforms.Normalize(mean, std)(T.unsqueeze(a, 0)).std())
a = (a - mean)/std
mean = a.mean()
std = a.std()
print((mean, std))

Output:

(tensor(5.), tensor(2.7386))
tensor([[[[-1.4606, -1.0954, -0.7303],
          [-0.3651,  0.0000,  0.3651],
          [ 0.7303,  1.0954,  1.4606]]]])
tensor(0.)
tensor(1.0000)
(tensor(0.), tensor(1.0000))