CIFAR-100 results

[adambielski]

I think there's a mistake in README. It looks like the train and validation accuracy are swapped in the curves, and the reported loss, top-1 and top-5 accuracies are for train and not val datasets. If not, then you got a new state-of-the-art on CIFAR-100 ;)

[cvtower]

convert.py +line 25&&26 replace x = X_train[i] y = y_train[i] with x = X_test[i] y = y_test[i].