Save a model in TensorFlowSharp

I can load python-built models or build models myself but is there currently a good way of saving the whole graph that I do not know of? Right now i am loading a pretrained model, run the init operation and then train it myself which is working fine. However, i would like to save it to disc also, as is possible in python.

Only thing i found was the TFGraph.Save() method, but i would have to specify what tensors to save and wouldnt know how to reload it then (or if it is even doing what i think its doing).

Any help would be appreciated!

[migueldeicaza]

I added these convenience methods to TFSession

[plasticalligator]

I was frustrated earlier and worked out a temporary solution until the library is in a better state.

This works for what I need, but it can be improved a lot: there is a RestoreV2 function which can restore multiple tensors in a single go, but it wants a parameter named shape_and_slices, I haven’t a clue what formatting it wants the numbers in, and right now I want to throw my laptop out a window.

    public static class Extensions
    {
        static public TFOutput RestoreTensor(this TFSession session, string filename, string tensor, TFDataType type)        
            => session.Graph.Restore(session.Graph.Const(TFTensor.CreateString(Encoding.ASCII.GetBytes(filename))), 
                                     session.Graph.Const(TFTensor.CreateString(Encoding.ASCII.GetBytes(tensor))), 
                                     type);

        static public TFTensor[] SaveTensors(this TFSession session, string filename,params ValueTuple<string,TFOutput>[] tensors) =>
            session.GetRunner().AddTarget(session.Graph.Save(session.Graph.Const(TFTensor.CreateString(Encoding.ASCII.GetBytes(filename)), TFDataType.String),
                                          session.Graph.Concat(session.Graph.Const(0), tensors.Select(T =>
                                          {
                                              TFTensor clone = TFTensor.CreateString(Encoding.ASCII.GetBytes(T.Item1));
                                              return session.Graph.Const(new TFTensor(TFDataType.String,
                                                                                      new long[] { 1 },
                                                                                      clone.Data,
                                                                                      clone.TensorByteSize, 
                                                                                      null, IntPtr.Zero));
                                          }).ToArray()), tensors.Select(d=>d.Item2).ToArray() )).Run();
    }

e.g

            using (var session = new TFSession())
            {
                    var a = session.Graph.Const(30, "a");
                    var b = session.Graph.Const(12, "b");
                    var multiplyResults = session.GetRunner().Run(session.Graph.Add(a, b));
                    var multiplyResultValue = multiplyResults.GetValue();
                    Console.WriteLine("a*b={0}", multiplyResultValue);
                    session.SaveTensors($"saved.tsf", ("a", a), ("b", b));
            }

            using (var session = new TFSession())
            {
                    var a = session.RestoreTensor("saved.tsf", "a", TFDataType.Int32);
                    var b = session.RestoreTensor("saved.tsf", "b", TFDataType.Int32);                
                    var multiplyResults = session.GetRunner().Run(session.Graph.Add(a, b));
                    var multiplyResultValue = multiplyResults.GetValue();
                    Console.WriteLine("a*b={0}", multiplyResultValue);
            }