wmt19 adapts to multiple domains and the results I reproduce are inconsistent

[NCUIui]

Hello, can you share with me the wmt19 model you distilled and trained?

[FadedCosine]

I am sorry, but we did not train our model on wmt19 dataset. We only tested on IWSLT’14 De-En, IWSLT’15 En-Vi, and multi-domains translation datasets.

[NCUIui]

Thank you, I also want to ask you a question. When I reproduced your code, when run build_knntargets.sh, there was a warning:"fixed_shape none type“, Will this have any impact? As a result, my final result will be different from yours, sincere thanks!