AssertionError in train.py about coref

[mandarjoshi90]

It looks like your loss is diverging. I'll need more information to be able to help.

1. Were you able to evaluate the pretrained models with the default config? That will ensure that you have the right setup.
2. Did you try training with the default config for base?
3. Do the jsonlines files look ok?
4. Can you post the loss for the first 1000 steps?

[fairy-of-9]

I'm using Titan XP.

1. command : GPU=0 python evaluate.py bert_base << default config.
   result:
   

2/4. Result of GPU=0 python train.py train_bert_base

3. In my opinion, jsonlines files were made successfully.
   {"doc_key": "bc/cctv/00/cctv_0000_0", "sentences": [["[CLS]", "In", "the", "summer", " ...

[mandarjoshi90]

Hmm, the loss is going down so that's good. But I'm not sure why the evaluation is messing up. Could you send me the complete log?

[fairy-of-9]

I tried GPU=0 python evaluate.py bert_base for the complete log to give you.
evalute.py worked well!

and 'train.py' is working!

I don't understand this situation....
Thank you for your comment!!!