search

paul-tqh-nguyen / reuters_topic_labelling

Deep learning to automatically label Reuter's articles with the relevant topics.
1 stars 0 forks source link

Implement feature to quit early upon convergence #15

Closed paul-tqh-nguyen closed 4 years ago

paul-tqh-nguyen commented 4 years ago

Our initial experiments have shown that our model trains very slowly.

Thus, we made our hyperparameter search train for a large number of epochs.

However, some hyperparameter selections are very bad and we can know after a few epochs (e.g. 5 or so) that the model has converged at some really bad score.

Thus, we shouldn't bother training for 40 epochs if we know after 8 that the model isn't improving.

This task is to make it so that our hyperparameter search gives up early in these cases.

paul-tqh-nguyen commented 4 years ago

https://github.com/paul-tqh-nguyen/reuters_topic_labelling/commit/c039676af1fc7d885dba7c516d804a6854f5efab implements this feature request.