The results can not be reproduced

[Vincent-Hoo]

How is the hyper-parameters like?

[chxy95]

How is the hyper-parameters like?

I didn't set the random seed, so the performance may be changeable.
It is the truth that in a training process, especially at the middle and later stages of the process,
the performance of DML is always better than Indenpent model.
There is no special parameter need to be adjusted in the code.
You can observe the curve during training if you still have puzzles.