the variable "loss_att" is overrided

[daaiwusheng]

in the class AttMSEloss, in the forward function.

loss_att = ((( (x_org*y_mask[:,1,...].unsqueeze(dim=1)) - att ) ** 2)).mean() this result is overrided by the next sentence, loss_att = ((( x_org - att ) ** 2)).mean()

is this a bug?