Closed timonbimon closed 4 years ago
Hi, I'd love to support PyTorch 1.1 but unfortunately my schedule has been quite full for some time now so I won't be able to do it anytime soon :(
Ok thanks for the reply and thanks for contributing this in the first place :)
Hi, I just realised that I've put PyTorch 1.1 as a version to test against in travis. Could you check if it indeed works for PyTorch 1.1?
Hi, I just realised that I've put PyTorch 1.1 as a version to test against in travis. Could you check if it indeed works for PyTorch 1.1?
Hi! Thanks for your awesome work. It indeed works for PyTorch 1.1. When I used default initial value of CRF, the performance didn't improve, and then I derived transition matrix from training set, making it be the initial value of CRF, the validation loss almost reduced by half. I am curious about this weird results. Do you have any idea ? Thanks again!
Great! I can close this issue now.
Wow that's impressive. May I know how you exactly derived the transition matrix from the training set?
I just maintained a matrix (number of tags x number of tags), and went through the tags in training set. It's to count the frequency of transition from tag[i] to tag[j]. Finally, each row of the matrix would be normalized by sum of each them.
Kemal Kurniawan notifications@github.com 於 2019年8月28日 週三 上午4:46寫道:
Great! I can close this issue now.
Wow that's impressive. May I know how you exactly derived the transition matrix from the training set?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/kmkurn/pytorch-crf/issues/39?email_source=notifications&email_token=AFMJUAMUIVPTU7W27VF3LOTQGWHBHA5CNFSM4H6ZYVH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5JB75Q#issuecomment-525475830, or mute the thread https://github.com/notifications/unsubscribe-auth/AFMJUAJJYZVLUJIPIE7NVTLQGWHBHANCNFSM4H6ZYVHQ .
I see. I'm not sure why that works nor why that shouldn't work. But thanks for sharing. Hopefully that helps others too.
Hey :) Are there any plans to support PyTorch 1.1 any time soon?