TF_Graph
tensorflow: TF_Graph
A TensorFlow Graph, holding the operations a Session executes.
A TF_Graph owns the Graph it wraps and releases it when the object
is destroyed. A TF_Session keeps a reference to the Graph it was
created over, so a Graph is never released while a Session still needs it,
whatever order the variables are cleared in.
Source Code: TF_Graph
The TF_Graph class contains the following properties:
The TF_Graph class offers the following public methods:
TF_Graph: obj = TF_Graph ()
TF_Graph: names = operationNames (obj)
The graph of a loaded model can be inspected, which is how the operations to feed and to read are found when a model does not follow the usual names. Placeholders are fed; here serving_default_x is the model input and saver_filename belongs to the checkpoint saver.
sess = TF_Session.fromSavedModel (__tf_test_model__ ()); names = sess.Graph.operationNames (); types = sess.Graph.operationTypes (); names(strcmp (types, "Placeholder"))
ans =
{
[1,1] = serving_default_x
[1,2] = saver_filename
}
Every operation of the graph, with its type.
sess = TF_Session.fromSavedModel (__tf_test_model__ ());
names = sess.Graph.operationNames ();
types = sess.Graph.operationTypes ();
for i = 1:numel (names)
printf ("%-32s %s\n", names{i}, types{i});
endfor
b VarHandleOp b/Read/ReadVariableOp ReadVariableOp w VarHandleOp w/Read/ReadVariableOp ReadVariableOp serving_default_x Placeholder StatefulPartitionedCall StatefulPartitionedCall NoOp NoOp Const Const saver_filename Placeholder StatefulPartitionedCall_1 StatefulPartitionedCall StatefulPartitionedCall_2 StatefulPartitionedCall
TF_Graph: types = operationTypes (obj)
TF_Graph: tf = hasOperation (obj, name)
TF_Graph: n = numOutputs (obj, name)
TF_Graph: delete (obj)
This is called automatically when no variable refers to the object any more, and calling it a second time does nothing.