How to normalize a tensor to 0 mean and 1 variance?

22

you can do it for any tensor or batch of tensors in 3 steps:

import torch
a = torch.randn(2,3,4, 4) # batch=2, channel=3,H=4,W=4 and I want to normalize across per channel
mu = torch.mean(a,dim=(2,3),keepdim=True)
sd = torch.std(a,dim=(2,3),keepdim=True)
normalized_res = (a - mu)/sd
1 Like
23

From where you get mean and std? Why did you choose these particular values?

24

I’ve used the ImageNet stats, since they were used to pretrain the torchvision.models for classification as described here.