reproducity problem

[yangheng95]

reproducity problem

[yangheng95]

Hello, sorry to see your question so late. I've just trained LCF-Bert on the restaurant dataset, with the built-in train.py and default parameters in the repository, the performance quickly achieving 0.85+. Can you post your parameter settings here? Besides, LCF-Bert model defaults to deploy a single Bert layer to reduce memory consumption. To unleash the performance of LCF-Bert, see comments of line 34-35 in the lcf_bert.py file.

[yangheng95]

Hello, I have uploaded a training log for the possible reproduction. Please let me know if there is any question.

[ZhengZixiang]

Thanks for providing training log file! Now I can reproduce the accuracy. I feel sorry for bringing the trouble.

The main reason is my training file different from yours at the beginning, in which I set the learning rate scheduler and adopted early-stop strategy. I forgot to change my code to your style. My original setting hurt the model performance to achieve the better accuracy.

I apologize again for opening this issue. Thank you for your answer and patience!