o Õ2úhà ã@sBdZdZdd„Zdd„Zdaddd „Zd d „ZGd d „d eƒZdS)zÃCheckpoint compatibility functions with SaveableObject. Compatibility methods to ensure that checkpoints are saved with the same metadata attributes before/after the SaveableObject deprecation. Ú_LEGACY_SAVEABLE_NAMEcs‡fdd„}|S)aDDecorator to set the local name to use in the Checkpoint. Needed for migrating certain Trackables (see next paragraph) from the legacy `_gather_saveables_for_checkpoint` to the new `_serialize_to_tensors` function. This decorator should be used if the SaveableObject generates tensors with different names from the name that is passed to the factory. Example migration: *Before* ``` class MyTrackable(Trackable): def _gather_saveables_for_checkpoint(self): return {"key": _MySaveable} class _MySaveable(SaveableObject): def __init__(self, name): specs = [ SaveSpec(tensor1, "", name + "-1") SaveSpec(tensor2, "", name + "-2") ] super().__init__(None, specs, name) ``` *After* ``` @legacy_saveable_name("key") class MyTrackable(Trackable): def _serialize_to_tensors(self): return {"key-1": tensor1, "key-2": tensor2} ``` Args: name: String name of the SaveableObject factory (the key returned in the `_gather_saveables_for_checkpoint` function) Returns: A decorator. cst|tˆƒ|S©N)Úsetattrr©Ú cls_or_obj©Úname©úg/var/www/html/chatgem/venv/lib/python3.10/site-packages/tensorflow/python/checkpoint/saveable_compat.pyÚ decoratorEs z'legacy_saveable_name..decoratorr)rr rrr Úlegacy_saveable_names -r cCs t|tdƒSr)Úgetattrrrrrr Úget_saveable_nameKs r FTcCs|adS)acForces checkpoint to use the new implementation. The new checkpoint implementation is changing the saved metadata slightly, and therefore may break forward compatibility in newly saved checkpoints. This means: - Previous versions of TensorFlow may not be able to load new checkpoints. - Backwards compatibility is unchanged: Old checkpoints can still be loaded. TensorFlow guarantees 3 weeks of forward compatibility, so this flag will be removed in the future weeks, after which checkpoint conversion will happen by default. **What happens when this flag is enabled?** The checkpoint will be saved with different metadata, meaning that previous versions of TensorFlow (<=2.10) will not be able to load this checkpoint. Args: value: Boolean value, whether or not to force checkpoint conversion to the new implementation. N©Ú_FORCE_CHECKPOINT_CONVERSION)Úvaluerrr Úforce_checkpoint_conversionRsrcCstSrrrrrr Ú#force_checkpoint_conversion_enablednsrc@s eZdZdS)ÚCheckpointConversionErrorN)Ú__name__Ú __module__Ú __qualname__rrrr rrsrN)T) Ú__doc__rr r rrrÚ Exceptionrrrrr Ús3