C++ memory leak after session->Close()

[pil...@gmail.com]

Hi,

Using a TF graph and session via C++ API, closing the session does not appear to release the memory (RAM).

Here is a code snippet:

```

tensorflow::GraphDef graph_def;

tensorflow::Status graphLoadedStatus = ReadBinaryProto(tensorflow::Env::Default(),graphFile,&graph_def);

tensorflow::SessionOptions options;

std::unique_ptr<tensorflow::Session> session = std::unique_ptr<tensorflow::Session>(tensorflow::NewSession(options));

tensorflow::Status session_create_status = _session->Create(graph_def);                                                        

...                                                                               

tensorflow::Status run_status  = session->Run({{_inputLayer,*(vtfinputs.begin())}},{_outputLayer},{},&finalOutput)  /* runs file, results are what they should be, and are acquired via finalOutput. */

...

session->Close();                                                                                                          

session.reset();

/* RAM keeps building up if code above put into a loop */

```

I've read all I could find, and I understand that the closing of the session does not release the underlying graph. In Python it seems that `reset_default_graph()` releases the graph resources.

What is / would be the equivalent of `reset_default_graph()` in C++ ? Looking at the `python/client/ops.py` I was unable to link the stack release to the C++ counterpart. 

As a side note, using ```session->Reset(options,containers)``` does not appear to compile (Reset does not exist for Session). Looking at ```session.h``` and ```direct_session.h``` I don't understand why this is the case. 

Could someone explain what exact Session subclass is ran by the C++ APi by any chance.

(GPU memory appears to not been released when looking it up with `nvidia-smi` but I believe it is an `nvidia-smi` problem as in practice the GPU memory appears to be reusable, so no problem on this side).

I've spent a decent amount of time on this one, so any help and pointers are very much appreciated !

Thanks,

Em.

[bharat.wor...@gmail.com]

Hey,

I am facing the same issue.

Were you able to find a solution ?

Thank you,

Bharat