Perfomance on English dataset NYT is not good to match the origin paper'result!

[GorgeOus2020]

I use this pytorch version reimplement of CasRel to run on dataset：NYT，and use bert-base-cased and other settings as same as the original paper does, but I got the best F1 of 70，which is lower than its paper's result 89.6, but when I use the origin keras-version CasRel to run on same dataset NYT, got the best F1 of 88.8, which is almost similar as paper's result! So, I wonder the correctness of this reimplement! Have you had any similar problems?

[GorgeOus2020]

And， I just change some non-core code just like matching the English tokens or Parital Match as original paper does， but don‘t change the code about casrel model and forward computation.

[Onion12138]

Thanks, there is a slight difference when I reimplement, for example I only sample one relation each time. This is due to my superficial understanding of the original code. I will read the keras version again and reimplement it again.

发自我的iPad

------------------ Original ------------------ From: GorgeOus2020 @.> Date: Sun,Oct 24,2021 9:31 AM To: Onion12138/CasRelPyTorch @.> Cc: Subscribed @.***> Subject: Re: [Onion12138/CasRelPyTorch] Perfomance on English dataset NYT isnot good to match the origin paper'result! (Issue #5)

And， I just change some non-core code just like matching the English tokens or Parital Match as original paper does， but don‘t change the code about casrel model and forward computation.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[GorgeOus2020]

Thank you for your reply and open source!