Hi Do you have any training performance of this implementation that you can share? And do you achieve similar performance to the original paper?