Training accuracies as weights for ensemble in supervised_sentiment?

[hoonkai]

Target objective:

Creating an ensemble using weights

Question

Looking at https://github.com/NervanaSystems/nlp-architect/blob/master/examples/supervised_sentiment/example_ensemble.py#L96-L97 how come the ensemble weights are training accuracies rather than validation accuracies? If the difference between acc and val_acc is great, wouldn't the ensemble overfit the training set?

[github-actions]

Issue marked as stale, closing.