Efficient softmax approximation for GPUs