The training success of the Transformer is due to several key design choices: in addition to the efficient parallelization architecture that eliminates the sequential dependency of RNNs and dramatically shortens training time, there is also a carefully tuned optimization strategy -- particularly the Adam optimizer with warmup -- that ensures stable and fast convergence; and an effective combination of regularization techniques, with residual dropout and label smoothing working together to prevent overfitting in deep networks and large models.