Wee Tee, Soh comments

Results 20 comments of


                                            Wee Tee, Soh

Running TextGCN processing crashes due to high RAM usage

I have updated the script such that building document-word edges is done on the fly, but this only slightly reduced RAM usage. Graph preprocessing is quite intensive as its quadratic...

Pre-trained final model

Yes if anyone could share a pre-trained model on a large corpus that would be great. I could only pre-train on a small CNN corpus as a proof-of-concept. Will leave...

Missing ./data/BERT_tokenizer.pkl file

I have added instructions as suggested to download & unzip the data

FewRel: meta_labels[-1].item() is always 4

Thanks :)

Can't infer on my own sentences after finetuning model

Yes, the notations have been changed :)

Training and prediction on new dataset

Hi all, if you need to fine-tune on your own dataset other than Semeval, you will have to fork this and modify src/tasks/preprocessing_funcs.py to ingest your own data and reformat...

BERT_REL with SciBert

Hi all, yes @vabatista is right, the constructor of BertModel has been modified here from the original, in order to implement this model. If you want to use this for...

bad accuracy on test data

Hi there, for semeval task the model is evaluated on test data. However for pre-training, there is no test data as the pre-training is self-supervised. You were probably looking at...

Development Set

Hi, what do you mean by development set?

FewRel fine-tuning?

Yes, FewRel fine-tuning have not yet been implemented, I have put as a to-do task