Adding TimeStopping to tfa.callback #757
Conversation
|
Check out this pull request on You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB. |
|
@seanpmorgan when you have a chance may you review this PR? Thank you so much for your time! |
| Args: | ||
| seconds: maximum amount of time before stopping. | ||
| Defaults to 86400 (1 day). | ||
| verbose: verbosity mode. Defaults to 0. |
There was a problem hiding this comment.
Is there a reason for using an int here instead of a boolean? Doesn't seem like varying verbosity values will make any difference, so better to prevent that confusion
There was a problem hiding this comment.
I use int here for consistency with EarlyStopping Callback.
https://github.com/tensorflow/tensorflow/blob/r2.0/tensorflow/python/keras/callbacks.py#L1134-L1251
And I would assume they also use int instead of just boolean is so that it is consistent with other places that has verbose as a parameter. This is my reasoning (for consistency), but if you feel like changing to boolean is better let me know! Thanks for the feedback!
There was a problem hiding this comment.
Thanks. Still feels a little intuitive, but I agree better to stay with the convention.
ops, got it will fix! |
fixed! |
This PR adds TimeStopping to tfa.callback. Similar to EarlyStopping but stops the training when a specified amount of time has passed since the start of the training.
This may be useful for people who want to train a model for a certain amount of time. Personally, sometimes I would want to train models for a certain amount of time (and there are codes after training ends so CTRL+C is not ideal) and I often find myself using pen and pencil to calculate how many epochs I would want to run. I hope people may find this useful as well.
Relevant Issues:
#750
keras-team/keras#1625
keras-team/keras-contrib#87
Thank you so much for your time!