BVh8 dZddlZddlZddlmZddlmZddlmZddlmZddlm Z ddl m Z dd l m Z e d g ejZhd Zgd ZdZdZdej_e j&dee dg d%dZdZdZdZdZdZe j&dee dg dZe j&dee dg dZe j&dee dg d&dZe dg  d%d ejd!ejd"ed#efd$Zy)'z0Helpers to manipulate a tensor graph in python. N) graph_pb2) node_def_pb2)_proto_comparators)dtypes)ops) deprecation) tf_exportGraphDef)v1> QueueAssignVariable AssignAdd AssignSub ScatterAdd ScatterSub VariableV2 ScatterUpdateTruncatedNormal)SwitchEnterExitIdentityMerge NextIterationzThis API was designed for TensorFlow v1. See https://www.tensorflow.org/guide/migrate for instructions on how to migrate your code to TensorFlow v2.c|tvS)z/Returns true if 'op' refers to a Variable node.) _VARIABLE_OPS)ops [/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/framework/graph_util_impl.py_is_variable_opr ;s } aA protobuf containing the graph of operations. @compatibility(TF2) This API is not available in TensorFlow 2.x. You should not need to use `GraphDef`s directly in TF2. To load `GraphDef`s in TF2, use SavedModel. The SavedModel contains the `GraphDef`. Before: ```python with tf.io.gfile.GFile('/tmp/graph.pb', 'rb') as f: graph_def = tf.compat.v1.GraphDef() graph_def.ParseFromString(f.read()) ``` After: ```python tf.saved_model.load('/tmp/saved_model') ``` If you would like to create a `GraphDef` in TF2, use `tf.function` and `get_concrete_function`. >>> @tf.function >>> def f(x): >>> return x >>> >>> graph_def = f.get_concrete_function(1.).graph.as_graph_def() >>> print(graph_def) @end_compatibility )date instructionszgraph_util.must_run_on_cpucHt|tjr |j}nt|tj sJ|}|rt |jry|jdk(r@|jdj}|tjk(s|tjk(ry|jdvr-|jdj}|tjk(ry|jdvr-|jdj}|tjk(ryy) arReturns True if the given node_def must run on CPU, otherwise False. Args: node: The node to be assigned to a device. Could be either an ops.Operation or NodeDef. pin_variables_on_cpu: If True, this function will return False if node_def represents a variable-related op. Returns: True if the given node must run on CPU, otherwise False. TConstdtype) DynamicStitchParallelDynamicStitchT)CastSrcTF) isinstancer Operationnode_defrNodeDefr rattrtyperstringint32)nodepin_variables_on_cpur.r&s rmust_run_on_cpur6fs"cmm$}}H dL00 11 1Hohkk: [[G MM' " ' 'E  &,,!6  [[>> MM#  # #E    [[H MM& ! & &E    r!cV|jdr|ddS|jddS)N^:r) startswithsplit)ns r _node_namer>s+\\# QR5L 773<?r!cT|jd}|jdr|ddS|S)zCDecodes colocated node name and returns it without loc:@ prepended.zutf-8zloc:@N)decoder;)colocated_node_namecolocated_node_decodeds r_get_colocated_node_namerDs4.55g>&&w/ !!" %% r!ci}i}i}d}|jD]}t|j}|||<|jDcgc] }t|c}||<d|jvrE|jdj j D]}||jt|!|||<|dz }|||fScc}w)znameinputr0listsappendrD) graph_defname_to_input_name name_to_nodename_to_seq_numseqr4r=xrBs r_extract_graph_summaryrRs,/ #nn  d499ALO48JJ?qZ]?q499!%8!4!9!9!;!;; 1$$ $%8 9 ;;OA1HC   \? ::@sB<c.|D]}||vrJd|zy)z+Assert that nodes are present in the graph.z%s is not in graphN)rNnodesds r_assert_nodes_are_presentrWs* 7a 62Q66 7r!ct}t|}|r-|d}|d=||vr|j|||vr|||z }|r-|S)z;Breadth first search for reachable nodes from target nodes.r)setrIadd) target_nodesrM nodes_to_keep next_to_visitr4s r_bfs_for_reachable_nodesr^sg%-|$-  Da }d !!)$//m  r!zgraph_util.extract_sub_graphcjt|tjstdt |dt|t rtdt |dt |\}}t||t||}tt|fd}tj}|D]4}|jjtj||g6|jj!|j|j"j!|j"|S)aDExtract the subgraph that can reach any of the nodes in 'dest_nodes'. Args: graph_def: A graph_pb2.GraphDef proto. dest_nodes: An iterable of strings specifying the destination node names. Returns: The GraphDef of the sub-graph. Raises: TypeError: If 'graph_def' is not a graph_pb2.GraphDef proto. z;graph_def must be a graph_pb2.GraphDef proto, but got type .z8dest_nodes must be an iterable of strings, but got type c|SNrT)r=rOs rz#extract_sub_graph..s );r!)key)r,rr TypeErrorr1strrRrWr^sortedrIr4extendcopydeepcopylibraryCopyFromversions) rL dest_nodesrMrNr\nodes_to_keep_listoutr=rOs @rextract_sub_graphrqs!" Iy11 2  O,A/ 00 C  ,-Q0 117M73lOL*5*:7IJ- =;= # 6aHHOOT]]<?3456++y((),, **+ *r!z*graph_util.tensor_shape_from_node_def_namec`d|vr|dz}n|}|j|}|j}|S)zBConvenience function to get a shape from a NodeDef's input string.r:z:0)get_tensor_by_name get_shape)graph input_namecanonical_nametensorshapes rtensor_shape_from_node_def_namerzs> $&NN  # #N 3&    % ,r!z graph_util.remove_training_nodesc|sg}ddi}|j}i}|D]/}|j|vs|j|vs!d||j<1g}|D]}|j|vrtj}|j ||j }|j dd=|D]9} tjdd| } | |vr|j j| ;|j|ddd} t} t} t}|D]}|j D]C}d|vs| j|jdd| j|jEd|jvsd|jdjjD]}|jt!|i}|D]Z}|j| vs|j|vs!|j|vr0|j| vs?|j d ||j<\|j#Dcic] \}}|| vr||}}}g}|D]}|j|vrtj}|j ||j }|j dd=|D]Y} tjdd| } | |vr!|| } tjdd| } | |vr!|j j| [|j|t%j&}|jj)||Scc}}w) aPrunes out nodes that aren't needed for inference. There are nodes like Identity and CheckNumerics that are only useful during training, and can be removed in graphs that will be used for nothing but inference. Here we identify and remove them, returning an equivalent graph. To be specific, CheckNumerics nodes are always removed, and Identity nodes that aren't involved in control edges are spliced out so that their input and outputs are directly connected. Args: input_graph: Model to analyze and prune. protected_nodes: An optional list of names of nodes to be kept unconditionally. This is for example useful to preserve Identity output nodes. Returns: A list of nodes with the unnecessary ones removed. CheckNumericsTNz^\^)r StopGradientr8rFr)r4rrGrr/rlrHresubrKrYrZreplacer0rIrJrDitemsrr rh) input_graphprotected_nodestypes_to_remove input_nodesnames_to_remover4nodes_after_removalnew_nodeinput_before_removalfull_input_namervtypes_to_splicecontrol_input_namesnode_names_with_control_inputnode_in_colocated node_inputrBnames_to_splicerGvaluenodes_after_splicing output_graphs rremove_training_nodesrs]. O$d+/  +/(d ww/!dii&F#'odii ( )d yyO###%H d::q/-66&"o6j  &nnO, - x( )"&trDrRrWr^rqrzrboolrrTr!rrs /2:.+.6 %9 %i&8&8 9  &"*  # L !# +,-).#)f ;.7 " !# -./$ 0#$ N !# ;<= >#  !# 123]4#]@ 5"=1644)"4"444"+"4"444)-44:>44>44r!