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
the init function of the resulting GradientTransformation cannot be jit compiled. The reason is that by default inject_hyperparams treats all arguments as dynamic and one of the argument has to be static to avoid a TracerError. A workaround is to specify the static argument:
This PR adds a test to alias_test to ensure that all optimizers can be wrapped
in inject_hyperarams. This is to check whether issue #412 affects other
optimizers too and not just adafactor.
Currently, adafactor needs static_args to be passed. This will be solved as
part of issue #412.
PiperOrigin-RevId: 474808793
This PR adds a test to alias_test to ensure that all optimizers can be wrapped
in inject_hyperarams. This is to check whether issue #412 affects other
optimizers too and not just adafactor.
Currently, adafactor needs static_args to be passed. This will be solved as
part of issue #412.
PiperOrigin-RevId: 474808793
This PR adds a test to alias_test to ensure that all optimizers can be wrapped
in inject_hyperarams. This is to check whether issue #412 affects other
optimizers too and not just adafactor.
Currently, adafactor needs static_args to be passed. This will be solved as
part of issue #412.
PiperOrigin-RevId: 475810505
When wrapping
optax.adafactor
withoptax.inject_hyperparams
without specifyingstatic_args
the
init
function of the resultingGradientTransformation
cannot be jit compiled. The reason is that by defaultinject_hyperparams
treats all arguments as dynamic and one of the argument has to be static to avoid aTracerError
. A workaround is to specify the static argument:However, this is not ideal since it requires the user to know which arguments should be static and which ones can be dynamic.
We should:
The text was updated successfully, but these errors were encountered: