This work is very interesting. Would you mind introducing the detail of training process and releasing the training code?