-
Notifications
You must be signed in to change notification settings - Fork 44
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
feat: Policy driven training control
Signed-off-by: Padmanabha V Seshadri <seshapad@in.ibm.com>
- Loading branch information
Showing
1 changed file
with
5 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,5 @@ | ||
# Note on training control definition examples | ||
- `ctldef_step_v0.3.yaml`: Defines a training controller which computes loss at every step and loss consistently increases for three steps, then the training is stopped. | ||
- `ctldef_epoch_v0.3.yaml`: Defines a epoch level training controller which computes loss at every epoch. The rule applied here is to compare a current epoch loss with previous epoch loss and it turns out to be more, then training is stopped. | ||
- `ctldef_epoch_threshold_v0.3.yaml`: Defines a training controller similar to previous case, but also adds a threshold constraint. | ||
- `ctldef_evaluate_v0.3.yaml`: Defines a training controller which behaves similar to the `EarlyStoppingCallback` from hugging face which can be found [here](https://github.com/huggingface/transformers/blob/v4.37.2/src/transformers/trainer_callback.py#L543). |