**debug feature** for debugging numeric instability, like nan or infinity #37904
Labels
comp:apis
Highlevel API related issues
stat:awaiting tensorflower
Status - Awaiting response from tensorflower
type:feature
Feature requests
System information
Describe the feature and the current behaviour/state.
I propose a debug feature for debugging numeric instability. The functionality should be similar to
tf.debugging.enable_check_numerics()
but more granular, but still more high level thentf.debugging.check_numerics
.Implementation Idea:
A
tf.debugging.check_func_numerics
annotation is added. It can be applied to ops warped in a python eager function and to tf.functons.It will add numerics checks to all ops used in this function and, this is IMPORTANT, to all corresponding gradient ops.
The part of a model which courses numeric instability can be identified with this.
Will this change the current api? How?
it will add another debugging function/annotation, so the standard api is not changed.
Who will benefit with this feature?
Everyone with problems with nan and infinity in there models.
Especially for big models where
tf.debugging.check_func_numerics
can not be applied to the hole model.The text was updated successfully, but these errors were encountered: