consistent_loss did not sum among the M synthetic noisy data ?

[yuanwei0908]

[LiJunnan1992]

The consistency_loss gradient is accumulated across M synthetic data, which is equivalent to summing the loss.