datquocnguyen / jointre Goto Github PK

Im working on something similar so curios how you load a saved model to be used as inference.I understand we'd have to load the parameters file from what I have seen params file is not getting created.

Not a big issue, but it's probably worth a quick update to include either documentation in the readme to download the GloVe pretrained 100B word embeddings, or a curl command to the scripts, otherwise following procedure causes an assertion error that requires traceback to the missing file.

Traceback (most recent call last):
  File "jNERE.py", line 77, in <module>
    parser = learner.jNeRE(words, nertags, rels, w2i, c2i, options)
  File ".../jointRE/jNERE/learner.py", line 58, in __init__
    assert (ext_emb_dim == self.wdims)
AssertionError

Im trying to modify this code to include shortest path dependency to this. Dependency parsing and pos tags would be additional features. Do you think we can collaborate if you have the badnwidth

Thanks for your hard word,
1.I have learned the paper,found that
'although using Biaffine increases training time over using Linear by 35%, relatively,',I want to know the speed of train and the machine configuration
2.whether to support batch_size?
I want to use it in my Chinese dataset

Looking forward to your reply

I'm noticing that the training routine can fail due to NaN gradient values occasionally.

Starting epoch 91
Processing sentence number: 100 , Loss: 0.0293 , Time: 12.89
Processing sentence number: 200 , Loss: 0.0100 , Time: 14.91
Traceback (most recent call last):
  File "jNERE.py", line 83, in <module>
    parser.Train(train_data, train_id2nerBILOU, id2arg2rel)
  File ".../jointRE/jNERE/learner.py", line 449, in Train
    self.trainer.update()
  File "_dynet.pyx", line 5728, in _dynet.Trainer.update
  File "_dynet.pyx", line 5733, in _dynet.Trainer.update
RuntimeError: Magnitude of gradient is bad: -nan

Hello,

I think there are two issues with your metrics that prevent comparison with state-of-the-art:

1. In your paper, you write about macro-averaged F1 which indeed seems to be what you use in your code. Almost every work in NER or RE use micro-averaged F1 score.

2. You discard the "Other" entity type. As far as I understand in (Bekoulis 2018), this is only done with Entity Classification (i.e. your Setup 2) and not joint NER and RE.