Open
Description
We have users building active learning in the following pattern:
- Create 3 datasets: train_ds, eval_ds, active_pool_ds
- Train a base model with <train_ds, eval_ds>
- Load a pretained model (e.g. HuggingFace transformer)
- Train it with <train_ds, eval_ds>
- Save the trained model
- Run N trials:
- Create a trial_train_ds by sampling active_pool_ds
- Train a model from base model, with <trial_train_ds, eval_ds>
- Load the base model trained above
- Train it with <trial_train_ds, eval_ds>
- Save the trained model
This use case involves the active learning + pre-tained model fine-tuning. In particular, it will use a feature in Tune which treats Dataset as hyperparameters.