DHA: Learning Decoupled-Head Attention from Transformer Checkpoints via Adaptive Heads Fusion

Re-assign community