You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Either I am misunderstanding the documentation or the implementation of epsilon scheduling is not consistent with the description.
The documentation of the epsilon scheduler class currently states
That value [the regularization strength] is either the final, targeted regularization, or one that is larger, obtained by geometric decay of an initial value that is larger than the intended target.
def at(self, iteration: Optional[int] = 1) -> float:
"""Return (intermediate) regularizer value at a given iteration."""
if iteration is None:
return self.target
# check the decay is smaller than 1.0.
decay = jnp.minimum(self._decay, 1.0)
# the multiple is either 1.0 or a larger init value that is decayed.
multiple = jnp.maximum(self._init * (decay ** iteration), 1.0)
return multiple * self.target
Could it be that the last two lines should be replaced by
Describe the bug
Either I am misunderstanding the documentation or the implementation of epsilon scheduling is not consistent with the description.
The documentation of the epsilon scheduler class currently states
That value [the regularization strength] is either the final, targeted regularization, or one that is larger, obtained by geometric decay of an initial value that is larger than the intended target.
To Reproduce
The implementation is
Could it be that the last two lines should be replaced by
I am using ott-jax = 0.4.4.
The text was updated successfully, but these errors were encountered: