This is a refactoring PR that replaces trainer_extra_kwargs, originally passed through model_args, to mixin_args which is now passed through training_args. This has two benefits:
These args are now log-able to wandb (they will show up under the run's config).
We can run an hp search over mixin related arguments.
Before, 2. was not doable since the hp_space method was only for updating the trainers self.args. Now mixin_args are a subset of these and can be tuned as needed. As well, the code should be a bit cleaner and easier to use (hopefully).
lucasosouza
commented
3 years ago
This is a refactoring PR that replaces
trainer_extra_kwargs, originally passed throughmodel_args, tomixin_argswhich is now passed throughtraining_args. This has two benefits:Before, 2. was not doable since the hp_space method was only for updating the trainers
self.args. Nowmixin_argsare a subset of these and can be tuned as needed. As well, the code should be a bit cleaner and easier to use (hopefully).