Question about the pseudo-parallel data in the DualRL training stage
sunny371 opened this issue · comments
Thanks for your great work!
I have a question for the pseudo-parallel data used in the annealing pseudo teacher-forcing stage: From your paper, I notice the pseudo-parallel data is generated on-the-fly using the latest model. However, I find the codes for this in the dual_training.py, as shown below,
Lines 84 to 86 in 7983ec0
Lines 308 to 317 in 7983ec0
just load paired data from args.tsf_train_data which stores the pseudo-parallel data generated by the template-based approach for pretraining, instead of generating with the latest model.
Did I understand this correctly? Have I missed anything?
Thank you.
Uh oh! I made a mistake here when I cleaned the code. I have deleted the code which uses the pseudo-data generated on-the-fly but preserved the code which uses the original pseudo-parallel data.
This figure shows the original code (before cleaning):
Wait for a minute, I will update the code! Thanks!
Thank you very much for your reply @luofuli .
But It seems there is still a problem. The mid_ids_in_bs is the transferred output from source style A to the target style B. Shouldn't this be used as the input_ids together with src[''ids''] as targets to train nmts_train[B] instead of nmts_train[A], since the generated text should be put in the source side?
I have updated the code! If you don't have any questions, you can close the issue! Thanks for your feedback!