How many training steps to get best results?

[xiyan]

Actually, I try to write this code by tensorflow. I have made every parameters in my code are the same as mentioned in paper. I use 4 examples as a batch and train the model for 200,000 times. While I test the ROUGE1 of the test sets, whose grade is much lower than mentioned in paper. How about your code? Would you like to provide concrete training times for reference? Thanks a lot.