How to call tensorflow::NewSession in c++ - c++

Question

I am using c++ to load and run tensorflow graph. The tensorflow version is 2.1 cpu. I have included necessary header files and the libary _pywrap_tensorflow_internal.lib
I use
unique_ptr<tensorflow::Session> session_inception(tensorflow::NewSession(SessionOptions()));
to create a new session. The compiling has no error.
But when I build the executable, there is a link error saying that NewSession is an unresolved external symbol. I guess the function "NewSession" is not in the library file "_pywrap_tensorflow_internal.lib".
How to call NewSession in c++ environment? or maybe using new APIs to run the graph (instead of using session)? Thank you very much.

Answer 1

I assume you're using this code from 2016: https://gist.github.com/kyrs/9adf86366e9e4f04addb which I found by googling for "session_inception". This code predates TensorFlow 2.1 and 2.0 (which was released in September 2019).
The TensorFlow 2.1 documentation does not list any function or type named Session or NewSession - I'm not a TensorFlow user but I think ClientSession is the current type you want.
Here's the sample code from the TensorFlow 2.1 documentation
Scope root = Scope::NewRootScope();
auto a = Placeholder(root, DT_INT32);
auto c = Add(root, a, {41});
ClientSession session(root);
std::vector outputs;
Status s = session.Run({ {a, {1}} }, {c}, &outputs);
if (!s.ok()) { ... }
That said, if you want to run this code from March 2016 you should just download the closest version of TensorFlow released around that date, which is version v0.7.1 available here: https://github.com/tensorflow/tensorflow/releases/tag/v0.7.1 - this is probably inadvisable as TensorFlow's documentation website only has documentation for versions released after v1.0
As an aside, I assume you're relatively new to working in C++ because I found this information quickly in a matter of seconds - had you done the same you wouldn't have posted this question to SO. If you are inexperienced working with C++ or TensorFlow then I advise you that it would be a bad idea to go wading around obsolete and unsupported C++ libraries as you'll be SOL if anything goes wrong - and things will go wrong, especially with prelease versions of TensorFlow.