SRU
Training RNNs as Fast as CNNs
...We also obtain an average of 0.7 BLEU improvement over the Transformer model on the translation by incorporating SRU into the architecture. The experimental code and SRU++ implementation are available on the dev branch which will be merged into master later.