Bug description: training for the task of event extraction using more than one dataset results in a model that is actuallly trained on only one of the EE datasets, selected at random.
How to reproduce: create a set of train/valid/test jsonl files containing at least 2 event extraction datasets and start training a model on EE. Trig and Args losses and R/P/F1 metrics will stay at 0 for all but one dataset.
How to fix: a typo was found in the dygie/models/events/EventExtractor constructor resulting in an incorrect key for the self._trigger_scorers ModuleDict. The key was set to namespace when it should have been set to trigger_namespace.
Bug description: training for the task of event extraction using more than one dataset results in a model that is actuallly trained on only one of the EE datasets, selected at random.
How to reproduce: create a set of train/valid/test jsonl files containing at least 2 event extraction datasets and start training a model on EE. Trig and Args losses and R/P/F1 metrics will stay at 0 for all but one dataset.
How to fix: a typo was found in the
dygie/models/events/EventExtractor
constructor resulting in an incorrect key for theself._trigger_scorers
ModuleDict. The key was set tonamespace
when it should have been set totrigger_namespace
.