Manually setting self.train in training_step

[pengbao7598]

I changed my code from normal training to Lightning. However, I cannot get the same results. I carefully checked the random seed and batch index.

If I don't manually call self.train() in the training_step, I get different predictions for the same model and the same batch sequence. However, if I call self.train() in the training_step, I get the same prediction with my previous codes.

I think this is weird. It should produce the same results whether I manually call self.train() or not, right?