CIFAR-10 normalize params

[wufeim]

CIFAR-10 normalize params

[chenyaofo]

Sorry for the late reply.

Actually, mean: [0.4914, 0.4822, 0.4465], std: [0.2023, 0.1994, 0.2010] are computed based on CIFAR-10 dataset. For more details that how to calculate mean and std, please refer to https://discuss.pytorch.org/t/computing-the-mean-and-std-of-dataset/34949/2.

However, sometimes, we can not calculate mean and std because of some reasons (for example, the dataset is too large to calculate). In this case, we will use transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)).

This is my understanding.

[chenyaofo]

In my opinion, these two normalized setting will yield competitive performance (I am not very sure with this).

[wufeim]

Thanks very much for the comment. For your information, I think the performance would drop by around 1% when using "the other" mean and std, which is usually significant.