Training model

[luoyu5023]

Hi, thank you for source code and great work.
When I was training the model, I found that the results of training/submit_exp_020.sh did not seem to correspond to training/exp_020/log.txt.
There are very few steps in an epoch. It seems that there are too few input sequences, which also leads to overfitting in my training process. Is there something wrong with my training code? I look forward to your reply. Thank you.