search

gpleiss / efficient_densenet_pytorch

A memory-efficient implementation of DenseNets
MIT License
1.52k stars 327 forks source link

Is the normalizatin values for CIFAR-10 correct? #68

Closed abduallahmohamed closed 3 years ago

abduallahmohamed commented 4 years ago

Hi, I believe that the CIFAR-10 normalization values used are actually for CIFAR-100.

The current values are (which actually are CIFAR-100 normalization values): 

mean = [0.5071, 0.4867, 0.4408]
stdv = [0.2675, 0.2565, 0.2761]

But they should be: 

mean=[0.49139968  0.48215841  0.44653091]
stdv= [0.24703223  0.24348513  0.26158784]

Ref: https://github.com/Armour/pytorch-nn-practice/blob/master/utils/meanstd.py

gpleiss commented 4 years ago

That seems correct. Shouldn't make a difference in practice w/ batch norm. I'd be happy to accept a PR fixing these numbers