Cannot fine-tune your model with default arguments

[ftakanashi]

Cannot fine-tune your model with default arguments

[yinhanliu]

Can you try it with --reset-optimizer?

[ftakanashi]

Can you try it with --reset-optimizer?

It works with --reset-optimizer. Thx!

[SunbowLiu]

Can you try it with --reset-optimizer?

It works with --reset-optimizer. Thx!

Hi,

I am curious about the fine-tuned performance, have you get any improvement?

[ftakanashi]

Can you try it with --reset-optimizer?

It works with --reset-optimizer. Thx!

Hi,

I am curious about the fine-tuned performance, have you get any improvement?

Hello,

Umm, actually now I think what I had done is more like continuous-training rather than fine-tuning...
Anyway, I used the same WMT17-En-Zh dataset mentioned in the paper and trained the released en-zh model for another extra 10 epochs.
The result is not very satisifying because final BLEU score by the extra-trained model is 25.22(T=4)
while the released one has a 32.63(T=4) as mentioned in the paper.