search

intersun / CoDIR

Code for EMNLP 2020 paper CoDIR
MIT License
41 stars 8 forks source link

Code implementation is different from the paper #5

Open Xinpeng-Wang opened 2 years ago

Xinpeng-Wang commented 2 years ago

Hi, Siqi Nice work! In the part of selecting teacher features for comparison, the code shows that the teacher features are selected from every two layers, which is different from the paper ( concat all layers and project to lower dimension ). And the result I got with contrastive loss shows only little improvement, which is different from the paper.

Could you give some hint about your training and the hyperparam setting?

jinxinglu commented 2 years ago

HI, Xinpeng, how did you do the implement of the negative examples in the pre-training? I can't find that from the paper and the code

thanks