It's not quite so simple: AFAICS they have not open-sourced the training code, only one for sampling from the pre-trained model.