@fanq15 There is a relevant discussion about your question here: Normalization in the mnist example
The mean/std used on Imagenet should be generally applicable unless you’re dealing with b/w images or something special. Then you can just use a subset of your training set (randomly sampled) to compute a new mean/std.