transformer-xl why pplx was so high as 1000 when I was training on wt103 dataset

why pplx was so high as 1000 when I was training on wt103 dataset

Open HanelYuki opened this issue 6 years ago • 3 comments

Jan 29 '19 06:01 HanelYuki

Could you give more information about your experiment setting? I did a quick run just now, and everything was OK.

Jan 29 '19 07:01 zihangdai

Could you give more information about your experiment setting? I did a quick run just now, and everything was OK.

I ran on the basis of the parameters of the previous version, using GPU

Jan 29 '19 07:01 HanelYuki

The result is greatly affected by hyper-parameters

Jan 29 '19 10:01 HanelYuki