Using Tensorflow in C++ with Python/C API results in missing default graph

Question

I want to train a Tensorflow model in C++ using the Python/C API (I am aware of the C++ API of Tensorflow, but it's too restricted). First I create the model in Python and then I export it. Ater that I reload it in C++.

The problem: Unfortunately the Python wrapping in C++ seems to drop the default graph of tensorflow during restoring the Tensorflow session.

Here the working import code in Python:

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf

print("Default graph before:", tf.get_default_graph())
saver = tf.train.import_meta_graph("graph.pb.meta")
sess = tf.Session()
saver.restore(sess, "graph.pb")
print("Default graph after:", tf.get_default_graph())

This outputs the python graph object two times:

Default graph before: <tensorflow.python.framework.ops.Graph object at 0x7f382b762588>
Default graph after: <tensorflow.python.framework.ops.Graph object at 0x7f382b762588>

Now the not working C++ code (sry for this much code, each line in the upper python code is marked with a comment):

int main(){
    // Init python
    Py_Initialize();
    std::cout << "Python version: " << Py_GetVersion() << std::endl;

    // py: import tensorflow
    PyObject *pName = PyUnicode_FromString("tensorflow");
    PyObject *fModule = PyImport_Import(pName);
    if(!fModule) std::cout << "Import tensorflow failed." << std::endl;
    PyObject *pDict = PyModule_GetDict(fModule);

    // py: print("Default graph before:", tf.get_default_graph())
    PyObject* pGetDefaultGraph = PyDict_GetItemString(pDict, "get_default_graph");
    PyObject* pDefaultGraph;
    if(PyCallable_Check(pGetDefaultGraph)){
        pDefaultGraph = PyObject_CallFunction(pGetDefaultGraph, 0);
        std::cout << "Default graph before: ";
        PyObject_Print(pDefaultGraph, stdout, 0);
        std::cout << std::endl;
    }
    else std::cout << "tensorflow.get_default_graph() is not callable.";

    // py: saver = tf.train.import_meta_graph("graph.pb.meta")
    PyObject* trainModule = PyDict_GetItemString(pDict, "train");
    PyObject* importMetaGraph = PyObject_GetAttrString(trainModule,"import_meta_graph");
    PyObject* fSaver;
    if(PyCallable_Check(importMetaGraph)){
        fSaver = PyObject_CallFunction(importMetaGraph, "(s)", "graph.pb.meta");
    }
    else std::cout << "Cannot create Tensorflow saver from imported meta graph" << std::endl;
    if(fSaver==0) std::cout << "Tensorflow model failed to load from file \"" << "graph.pb.meta" << "\"" << std::endl;

    // py: sess = tf.Session()
    PyObject* session = PyDict_GetItemString(pDict, "Session");
    PyObject* fSession;
    if(PyCallable_Check(session)){
        fSession = PyObject_CallObject(session, 0);
    }
    else std::cout << "Cannot create Tensorflow session" << std::endl;
    if(fSession==0) std::cout << "Tensorflow session points to zero, failed to create session" << std::endl;

    // py: saver.restore(sess, "graph.pb")
    PyObject_CallMethod(fSaver, "restore", "(0s)", fSession, "graph.pb");

    // py: print("Default graph after:", tf.get_default_graph())
    std::cout << "Default graph after: ";
    pDefaultGraph = PyObject_CallFunction(pGetDefaultGraph, 0);
    PyObject_Print(pDefaultGraph, stdout, 0);
    std::cout << std::endl;
}

And here the problem, this results in this, where the default graph disappiers after restoring the session:

Python version: 3.5.1 (default, Mar  3 2016, 09:29:07) [GCC 5.3.0]
Default graph before: <tensorflow.python.framework.ops.Graph object at 0x7f3e8cd572b0>
Default graph after: <nil>

I know, weird question, but I am totally confused!