BVh*dZddlZddlmZddlmZddlmZdZedGdd ejZ ed Gd d Z d Z y)zOptions for saving SavedModels.N) sharding_util)compat) tf_exportTz'saved_model.experimental.VariablePolicyc8eZdZdZdZdZdZdZdZe dZ y)VariablePolicyaUEnum defining options for variable handling when saving. NONE No policy applied: Distributed variables are saved as one variable, with no device attached. SAVE_VARIABLE_DEVICES When saving variables, also save their device assignment. This is useful if one wants to hardcode devices in saved models, but it also makes them non-portable if soft device placement is disabled (more details in `tf.config.set_soft_device_placement`). This is currently not fully supported by `saved_model.load`, and is mainly intended to be used when one will be reading the saved model at a lower API level. In the example below, the graph saved by the call to `saved_model.save` will have the variable devices correctly specified: ```python exported = tf.train.Checkpoint() with tf.device('/GPU:0'): exported.x_gpu = tf.Variable(1.0) with tf.device('/CPU:0'): exported.x_cpu = tf.Variable(1.0) tf.saved_model.save(exported, export_dir, options = tf.saved_model.SaveOptions( experimental_variable_policy= tf.saved_model.experimental.VariablePolicy.SAVE_VARIABLE_DEVICES)) ``` Distributed variables are still saved as one variable under this policy. EXPAND_DISTRIBUTED_VARIABLES Distributed variables will be saved with information about their components, allowing for their restoration on load. Also, the saved graph will contain references to those variables. This is useful when one wants to use the model for training in environments where the original distribution strategy is not available. Nsave_variable_devicesexpand_distributed_variablesc(|tjk7S)z0Checks whether variable devices should be saved.)rNONEselfs Z/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/saved_model/save_options.py_save_variable_devicesz%VariablePolicy._save_variable_devicesGs >&& &&c(|tjk(S)z8Checks whether distributed variables should be expanded.)rEXPAND_DISTRIBUTED_VARIABLESr s r_expand_distributed_variablesz,VariablePolicy._expand_distributed_variablesKs >>> >>rc|tjSt|tr|St|j }tD]}||j k(s|cSt d|d)z4Tries to convert `obj` to a VariablePolicy instance.z'Received invalid VariablePolicy value: .)rr isinstancestrlowervalue ValueError)objkeypolicys rfrom_objzVariablePolicy.from_objOsk {   #~& j c(.. C     >se1E FFr) __name__ __module__ __qualname____doc__r SAVE_VARIABLE_DEVICESrrr staticmethodrrrrrs;"H $1!?'? G Grrzsaved_model.SaveOptionsc2eZdZdZdZ ddZy) SaveOptionszOptions for saving to SavedModel. This function may be used in the `options` argument in functions that save a SavedModel (`tf.saved_model.save`, `tf.keras.models.save_model`). ) namespace_whitelistsave_debug_infofunction_aliasesexperimental_debug_stripperexperimental_io_deviceexperimental_variable_policyexperimental_custom_gradientsexperimental_image_formatexperimental_skip_saverexperimental_sharding_callback extra_tagsNc |t||_||_|r|n t|_||_||_||_tj||_ | |_ |rtr td||_| 2t| t j"stdt%| d| |_| |_y)aRCreates an object that stores options for SavedModel saving. Args: namespace_whitelist: List of strings containing op namespaces to whitelist when saving a model. Saving an object that uses namespaced ops must explicitly add all namespaces to the whitelist. The namespaced ops must be registered into the framework when loading the SavedModel. If no whitelist is provided, all namespaced ops will be allowed. save_debug_info: Boolean indicating whether debug information is saved. If True, then a debug/saved_model_debug_info.pb file will be written with the contents of a GraphDebugInfo binary protocol buffer containing stack trace information for all ops and functions that are saved. function_aliases: Python dict. Mapping from string to object returned by @tf.function. A single tf.function can generate many ConcreteFunctions. If a downstream tool wants to refer to all concrete functions generated by a single tf.function you can use the `function_aliases` argument to store a map from the alias name to all concrete function names. E.g. >>> class Adder(tf.Module): ... @tf.function ... def double(self, x): ... return x + x >>> model = Adder() >>> model.double.get_concrete_function( ... tf.TensorSpec(shape=[], dtype=tf.float32, name="float_input")) >>> model.double.get_concrete_function( ... tf.TensorSpec(shape=[], dtype=tf.string, name="string_input")) >>> options = tf.saved_model.SaveOptions( ... function_aliases={'double': model.double}) >>> tf.saved_model.save(model, '/tmp/adder', options=options) experimental_debug_stripper: bool. If set to True, this strips the debug nodes from the graph, from both the nodes and the function defs. Note that this currently only strips the `Assert` nodes from the graph and converts them into `NoOp`s instead. experimental_io_device: string. Applies in a distributed setting. Tensorflow device to use to access the filesystem. If `None` (default) then for each variable the filesystem is accessed from the CPU:0 device of the host where that variable is assigned. If specified, the filesystem is instead accessed from that device for all variables. This is for example useful if you want to save to a local directory, such as "/tmp" when running in a distributed setting. In that case pass a device for the host where the "/tmp" directory is accessible. experimental_variable_policy: The policy to apply to variables when saving. This is either a `saved_model.experimental.VariablePolicy` enum instance or one of its value strings (case is not important). See that enum documentation for details. A value of `None` corresponds to the default policy. experimental_custom_gradients: Boolean. When True, will save traced gradient functions for the functions decorated by `tf.custom_gradient`. Defaults to `True`. experimental_image_format: New (highly) experimental format that is capable of saving models larger than the 2GB protobuf limit. Enabling this option will likely break compatibility with downstream consumers. This option is currently disabled in OSS. experimental_skip_saver: If True, will prevent SavedModel from creating its native checkpointing ops - this is for models that do not use SavedModel's native checkpointing functionality to avoid the costs associated with creating and serializing those ops. experimental_sharding_callback: `tf.train.experimental.ShardingCallback`. A pre-made or custom callback that determines how checkpoints are sharded on disk. Pre-made callback options are `tf.train.experimental.ShardByDevicePolicy` and `tf.train.experimental.MaxShardSizePolicy`. You may also write a custom callback, see `tf.train.experimental.ShardingCallback`. extra_tags: Extra tags to be saved with the MetaGraph in the SavedModel. z:The option `experimental_image_format` is disabled in OSS.NzuThe experimental_sharding_callback checkpoint optionmust be of type ShardingCallback. The option providedwas of type r)_validate_namespace_whitelistr(r)dictr*r.r+r,rrr-r0is_ossrr/rrShardingCallbacktyper1r2) r r(r)r*r+r,r-r.r/r0r1r2s r__init__zSaveOptions.__init__tsX = D+D0@,dfD)FD&'BD$"8D(6(?(?$)D%$;D !V  F &?D"%1  (-*H*H  >?@ C  +ID' DOr) NFNFNNTFFNN)rr r!r" __slots__r9r%rrr'r']s7 ) "'!#'$( %#%)k!rr'c |yt|tstd|dt|dg}|D]Q}t|tst d|dt|d|j tj|S|S)z'Validates namespace whitelist argument.Nz6`namespace_whitelist` must be a list of strings. Got: z with type rz-Whitelisted namespace must be a string. Got: z of type ) rlist TypeErrorr8rrappendras_str)r( processed namespaces rr4r4s  ' . @  { # $ %Q (  )&/i i %  9 Kyi 1 4 V]]9-. / r) r"enum%tensorflow.python.checkpoint.shardingrtensorflow.python.utilr tensorflow.python.util.tf_exportrr6Enumrr'r4r%rrrGsn& ?)6  45>GTYY>G6>GB $%A!A!&A!Hr