Improving Transformers with Dynamically Composable Multi-Head Attention

Xiaol.x - Improving Transformers with Dynamically Composable Multi-Head Attention