data normalization - possible bug

[rosenfeldamir]

I noticed that the data normalization is

transforms.Normalize((0.5, 0.5, 0.5),(0.5, 0.5, 0.5))

(for example in twoafc_dataset.py).
The imagenet normalization coefficients are
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]
This raises the questions:

1. possibly you confused mean with variance? (the Normalize function accepts std as the second argument, not variance)
2. any reason behind the design choice not to use the imagenet normalization?

[richzhang]

This "normalize" converts images from [0,+1] to [-1,+1]. We do the actual imagenet normalization here: https://github.com/richzhang/PerceptualSimilarity/blob/master/models/networks_basic.py#L112 .

[zakajd]

Hi
I don't understand where did you get the normalization values.
Currently you are using mean = [ -.030,-.088,-.188] and std = [.458,.448,.450] which doesn't correspond to ImageNet statistics.

So the question is: Where did you get this values and what do they mean?

[richzhang]

Thanks for the question. The imagenet statistics assume images are normalized between [0,1]. This code uses images from [-1,+1], so the appropriate transformation is done (notice that the std is just doubled)

[zakajd]

Thank you for the fast answer =)