Typing the lamb optimizer. #931
Conversation
gabrieldemarmiesse
requested review from
facaiy,
Squadrick,
WindQAQ and
a team
as code owners
There was a problem hiding this comment.
Failing on linux tests for:
TypeError: type of argument "learning_rate" must be one of (Tensor, float, float16, float32); got tensorflow.python.keras.optimizer_v2.learning_rate_schedule.InverseTimeDecay instead
Seems like this can be handled by modifying the types that are accepted... but I'm more confused as to why this fails only for linux builds. typeguard is an OS agnostic python library.
|
That's definitly strange. I read the docs a second time, and it says:
Do you think it's related? |
| exclude_from_weight_decay=None, | ||
| exclude_from_layer_adaptation=None, | ||
| name='LAMB', | ||
| learning_rate: Union[FloatTensorLike, Callable] = 0.001, |
There was a problem hiding this comment.
Looks good. Did we ever figure out why this was only failing for linux?
There was a problem hiding this comment.
Nope haha, my overall position on this is that we're very early in the history of tooling atound typing. We can try typeguard now, but if we see that it doesn't work well, we can drop it until the field is more mature.
Linked to #894