BVh2dZddlZddlZddlmZddlmZddlmZddlm Z ddl m Z dZ Gd d eZd Zd Zd ZdZej(dZdZdZy)zSubscribe function.N)ops)tensor) array_ops) variables) tf_loggingct|}t|tjr|St|tj r|j St|ttfrD|Dcgc]}t|}}|tur t|S|tur t|S||S|tur"tfd|jDStd|d|cc}w)aHHelper method to recursively apply a function to structure of tensors. The structure of the tensors should take the form similar to fetches in `tf.compat.v1.Session` and includes single `Tensor`, `list`, nested `list`, `tuple`, `namedtuple`, or `dict`. Args: tensors: Single `Tensor`, `list`, nested `list, `tuple`, `namedtuple`, or `dict`. apply_fn: Function to apply to each `Tensor` and should return a `Tensor`. Returns: Returns the modified tensors with the same structure. Raises: `TypeError` if undefined type in the tensors structure. c3@K|]\}}|t|fywN)_recursive_apply).0kvapply_fns U/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/framework/subscribe.py z#_recursive_apply..:s"Otq!$Q12Osz_recursive_apply argument z has invalid type ) type isinstance tensor_libTensorrVariablevaluelisttupler dictitems TypeError)tensorsr tensors_typets ` rr r s$g,**+ G '9--. GMMO $$'D%=)6=>8,>G>t ']   7^  !!t Ow}}O OO 0 ;M#&( ))?s9C4c(eZdZdZdgZdZdZdZy)_ControlOutputCachezHHelper class to manage calculating and caching control_outputs in graph.cacheci|_yr )r")selfs r__init__z_ControlOutputCache.__init__Es DJci}|jD]8}|jD]'}||vr t||<||j|):|S)zReturns the map of control_outputs for a given graph. Args: graph: The graph to parse. Returns: A map of the control outputs. )get_operationscontrol_inputssetadd)r$graphcontrol_outputsop control_inputs rcalc_control_outputsz(_ControlOutputCache.calc_control_outputsHseO""$/,,/-  /+.5/- ( &**2.// r&c|j|jvr5|j|j}||j|j<n|j|j}|j|gS)zReturn the control outputs for a given op. Args: op: The op to fetch control outputs for. Returns: Iterable of control output ops. )r,r"r0get)r$r.r-s rget_control_outputsz'_ControlOutputCache.get_control_outputsYs_ xxtzz!11"((;o,djj 288,o   r2 &&r&N)__name__ __module__ __qualname____doc__ __slots__r%r0r3r&rr!r!@sPi)"'r&r!c:g}t|jD]7}|j|t|jj |f9|j |j }|j jdz}tj|5g}|D] }|||z }tj|5tj|} dddddd|D]\}} |j|  |D]s}|j} |j | vr| j|j | j j |j!|j#| u S#1swYxYw#1swYxYw)alHelper method that subscribes a single tensor to a list of side_effects. Args: tensor: `tf.Tensor` side_effects: List of side_effect functions see subscribe for details. control_cache: `_ControlOutputCache` helper to get control_outputs faster. Returns: The modified replacement to the passed in tensor which triggers the side effects. /subscription/N)r consumersappendinputsindexr3r.namer name_scopecontrol_dependenciesridentity _update_inputr)remove_remove_all_control_inputs_add_control_inputs) r side_effects control_cache update_input consumer_opupdate_control_inputrAoutssoutr?new_control_inputss r_subscribe_newrQjs,&**,-Okd;+=+=&>&D&DV&LMNO'::699Eyy~~ 00* ~~j!' D  aid ! !$ ''   v &c' ')*k5eS)** 8k %33 yy&& *cff%**,##$67 8 *''' ''s$*F FFF FFc*t|jjdk(s*Jdj|jj|jjd}g}|jjdz}t j |5|D] }|||z } ddd|Dcgc]*}t|tjr |jn|,}}|jj||S#1swY[xYwcc}w)aHelper method to extend the list of side_effects for a subscribed tensor. Args: tensor: A `tf.Tensor` as returned by subscribe(). side_effects: List of side_effect functions, see subscribe for details. Returns: The given subscribed tensor (for API consistency). Op {} must only have one inputrr;N) lenr.r>formatr@rrArrrrG)rrH source_tensorrMrArNrOout_opss r_subscribe_extendrYs VYY   ! #%E%L%L iinn& #))""1%- $$$'77* ~~j!  a d HL @C 3 1 12cff; '  ))( - sD6/DD c|jjdk7rytjd|j}|t |j dk7ry|jd}t |jjdk(s*Jdj|jj|jjd}||jjk7ryy) afChecks if the given tensor is an identity op returned by `subscribe()`. Args: tensor: A `tf.Tensor` to check. Returns: True if the given tensor matches the criteria for subscription identities: its op type is `Identity`, its name matches the name of its input and conforms to the convention for subscribed nodes. False otherwise. IdentityFz0(?P^.*?)/subscription/Identity[^/]+rS prefix_namerTrT) r.rrematchr@rUgroupsgroupr>rV)rr^r\rWs r_is_subscribed_identityras YY^^z!  ((F;; % ]c%,,.)Q.  M*+ VYY   ! #%E%L%L iinn& #))""1%-M$$)))  r&c||jjs;tjdj |j |j|St |r t||S|jj dz}|j}|Dcgc] }|j j|s|"}}t|dks*Jdj |jj t|dk(r)|djd}t |r t||St|||Scc}w)aHelper method that subscribes a single tensor to a list of side_effects. This method will check if the given tensor has already been subscribed or if it's a tensor returned by a previous call to `subscribe()` and, if so, will reuse the existing identity op, appending the given side effects to the list of existing ones. Args: tensor: The `tf.Tensor` to be subscribed. side_effects: List of side_effect functions, see subscribe for details. control_cache: `_ControlOutputCache` helper to get control_outputs faster. Returns: The modified replacement to the passed in tensor which triggers the side effects or the given tensor, if it was already been subscribed. z?Tensor {} has an un-supported {} type and cannot be subscribed.z/subscription/IdentityrSz8Op {} must only have one subscription op connected to itr)dtypeis_numpy_compatibleloggingdebugrVr@rarYr.r< startswithrUoutputsrQ)rrHrIrAr<r. matching_opscandidate_tensors r _subscriberks&  ) ) MM!"(&fll"CE MV$ V\ 22yy~~ 88* )(KBGG,>,>z,J"K,K \ a O#78>vyy~~8NO !#A..q1/0 / >>  m <<Ls  D97D9c#K|jj}|r|j d|r|jyy#|r|jwwxYww)aPreserve the control flow context for the given tensor. Sets the graph context to the tensor's context so that side effect ops are added under the same context. This is needed when subscribing to tensors defined within a conditional block or a while loop. In these cases we need that the side-effect ops are created within the same control flow context as that of the tensor they are attached to. Args: tensor: tensor whose context should be preserved. Yields: None N)r._get_control_flow_contextEnterExit)rcontexts r_preserve_control_flow_contextrqsP( II / / 1'  MMO  llnw llns-A AA AA ctj|j5t|5t|||cdddcdddS#1swYnxYw dddy#1swYyxYw)aeHelper method that subscribes a single tensor to a list of side_effects. This is a thin wrapper around `_subscribe` and ensures that the side effect ops are added within the same device and control flow context of the subscribed tensor. Args: tensor: The `tf.Tensor` to be subscribed. side_effects: List of side_effect functions, see subscribe for details. control_cache: `_ControlOutputCache` helper to get control_outputs faster. Returns: The modified replacement to the passed in tensor which triggers the side effects or the given tensor, if it was already been subscribed. N)rdevicerqrk)rrHrIs r_scoped_subscribert"sf" zz&-- = ' /=  m <==========s! A" A  A" A A""A+c\tdsgtt|fd}|S)aSubscribe to a tensor. This method will attach side effect graphs to a given set of tensors. Set of tensors follows from session.run and supports single `Tensor`, `list`, nested `list`, `tuple`, `namedtuple`, or `dict`. It returns the tensors in the same passed in structure, but as clones with side effects applied. The supplied side effect graphs are specified as a constructor function which takes the target tensor and constructs a side effect graph and returns a list of ops that should be control dependencies on fetching the tensor. It will append 'subscription' to the name scope of the tensor for every node in the side effect graph. These control dependencies are what trigger the side effects. Subscribe will construct the additions to your graph and return the created identity tensor downstream of the control dependencies. Use these tensors as you would normally in the rest of your tensorflow code. If a given tensor has already been subscribed or a tensor returned by a call to subscribe is passed, the previously created identity tensor will be reused and the side effect graphs will be added to the existing ones. Args: tensors: `Tensor` or set of tensors to subscribe to. Set of tensors format follows from `Session.run` and supports single `Tensor`, `list`, nested `list`, `tuple`, `namedtuple`, or `dict`. side_effects: Function(s) that takes a `Tensor`, construct a subgraph, and return a nonempty list of control dependencies. This can be a single function or list of functions. Returns: Subscribed tensors, which are identity copies of the passed in tensors in the same passed in structure, but the graph has been modified such that these are downstream of the control dependencies for the side effect graphs. Use these functionally equivalent tensors instead of the passed in tensors for further construction or running. __iter__ct|Sr )rt)rr-rHs rzsubscribe..as*1lOLr&)hasattrr!r )rrHresultr-s ` @r subscriber{8s5H z * >L')/  L N& -r&)r7 contextlibr]tensorflow.python.frameworkrrrtensorflow.python.opsrrtensorflow.python.platformrrer objectr!rQrYrarkcontextmanagerrqrtr{r9r&rrsm +<++<")J''&''T, ^<D'=T <=,*r&