Acceleration.

Thanks for sharing your re-implementation, the code is clean and easy to follow. But I find it rather slow when I am doing 'training'. Is it possible to sacrifice some consistency and accelerate the process through parallelism?