Grid to pick hyperparameter for each model
kudkudak opened this issue · comments
Stanislaw Jastrzebski commented
Proper fitting of models to old and new dev split. Right now hyperparameters are shared, which might be bias heavily results.
For each model tune learning rate and l2 regularization.
Stanislaw Jastrzebski commented
For NAACL we will just fix hyperparams to the ones tuned for DNNs. Performance of factorized and prototypical is approximately stable wrt to l2_a and lr hyperparameter. I haven't checked others.
If we submit it somewhere else I would tune these hyperparams, but closing for now
Stanislaw Jastrzebski commented
But it actually seems important for glove.. will try to run proper grid.