Scaling input

[pencilzhang]

Hi, thanks for your work. Can I ask a question? Why do you scale input to -0.5 to 0.5?

[kunzmi]

Hi,
the reason is the (parametric) Relu in the convolutional layers and the fact that we only have a few layers. If the data is normalized to 0..1, only very few pixels would get negative after the first convolution. With a normalization -0.5..+0.5, on average half of the pixels fall in the negative Relu part, which should help training.
As the project is already quite old, I don't remember how much that actually influenced the result, though.

Cheers,
Michael