tensorflow C++ equivalent of tf.trainable_variables()?

Question

My goal is to get a list of names with all the trainable variables from the C++ API. In Python this would be down with tf.trainable_variables().

So far I tried this approach. I have a tensorflow::GraphDef object and I can see all nodes that have been created like this:

for (int i = 0; i < graphDef.node_size(); i++) {
    graphDef.node(i).PrintDebugString();
}

which is great. Some of those nodes refer to trainable variables, but I don't know how do I get that information / or if it's possible.

Answer 1

That information is not available in the GraphDef object. tf.trainable_variables just returns the graph collection with key tf.GraphKeys.TRAINABLE_VARIABLES , but graph collections are not saved to the GraphDef , only to the MetaGraphDef (see Exporting and Importing a MetaGraph ). If you want to access trainable variables in a saved graph from C++, you have to either export and import the MetaGraph instead or, maybe, use a consistent naming scheme to differentiate them.

Note, by the way, that graph collections will be deprecated in TensorFlow 2.x. See Deprecating collections for more information.