About the 20 iterations of updates in the pytorch implementation

[HuangYujie94]

The result is not satisfactory enough in the pytorch implementation without the 20 iterations of updates. So I don't know if it is the problem of the algorithm or just the pytorch implementation

[yangyazi]

Is your Pytorch code for multi-GPU training or single-GPU training? And would it be possible for you to provide me with the code for the single-GPU version? Thank you