search

jiesutd / LatticeLSTM

Chinese NER using Lattice LSTM. Code for ACL 2018 paper.
1.8k stars 453 forks source link

About the speed of training on MSRA dataset... #23

Closed DaveYin closed 6 years ago

DaveYin commented 6 years ago

Excuse me, I have some trouble training your model on MSRA dataset with a GTX 1080Ti card. I've found the speed of training is quite slow. So, may I know your solution to this problem? (Note: The video memory almost runs out, but there is still much unused computing power left.)

jiesutd commented 6 years ago

Hi, as this code is not implemented in batch. So it is not as faster as the batched implementation ( such as NCRF++). It is a little complex to using batched implementation, as the matches words is various for different input sentence. One possible solution is to calculate all the possible word link and use mask to filter unmatched word link. Currently, I do not have time to implement the batched version.