Improving Transformers with Dynamically Composable Multi-Head Attention

Re-assign community