Open zhaijianyang opened 1 year ago
Default parameters in the original code result in underfit model. You should train much more time than specified in the scripts. See our reproducibility study
You can also check out our hugging face based implementation https://github.com/asash/bert4rec_repro It usually converges faster than original.
RecBole reported that they fixed some bugs in their ber4rec implementation after our repro paper, so maybe it is now not bad (thought I didn't try their new version)
ok thank you
The BERT4Rec in recbole is not so bad. I ran the code and it is much better than the one in the paper. Why?