Shuffle-invariant Network for Action Recognition in Videos