Disable additional shuffling while training

Before this change dataset was shuffled once and
then on every step random sample was returned.
Problem is that this way we do not guarantee that
all samples from dataset will be used.