Some differences I've noted from the Torch WRN implementation

[RaananHadar]

Hi,
2 things:

1. Using the 26 layer code, i've reached 93% accuracy on cifar-10. Which is the same as you currently had for the 40 layer version.
2. re: the dataset used in the torch implementation. Please note that it uses a dataset that was already prewhitened and uses a global contrast normalization with a scale value of 55.

As a final note, I would suggest that you normalize the data by 255 so it would be from 0 to 1. Please take note that I have tried training on such a normalized set, It did not improve the performance significantly. The main difference from the torch implementation seems to be in the augmentation.