Closed injala-kumar closed 2 years ago
Hi,
first of all, it seems as if each entity cluster in your dataset only consists of a single mention. Are you sure JEREX is the right approach here? Maybe a model that extracts relations between single mentions and not clusters of mentions is better suited in this case, like SpERT (https://github.com/lavis-nlp/spert -> self-promotion :) ). Maybe the code is missing a corner case handling in case there are no (positive) coreference samples. Unfortunately, I won't be able to look at the problem for two weeks since I'm on vacation as of tomorrow.
Does the model train when you only localize entities (by setting 'model_type' in configuration file to 'mention_localization')? Also, if you want to, you can send me the dataset by email (markus_eberts@yahoo.de). This will certainly help for debugging.
I just added some corner case handling (412d3a94f9e866923da35799fc5c2f91ea7dc18e) before going on vacation. Please check if this already fixes your problem. Also, please try to increase the number of epochs (e.g. by setting training: min_epochs: 100
and training: max_epochs: 100
in configs/docred_joint/train.yaml
) and wait till the training finishes.
Please leave a comment if your issue is still not resolved.
Can you please help me with the following things. I have create a custom dataset with same structure as docred and tried to train the model. First of all it's not seems training by looking at performace metrics. Also after some epochs it's gives error and stop training. you can check the following logs.