Import InceptionV3 throws "Invalid GraphDef" error about tensorflowsharp

[csharpner]

My model is only 84KB and I have another that's 69KB and I get this error on both of them. Latest version of TensorFlowSharp available in NuGet as of 2019-02-01 and TensorFlow 1.12.0. I've not been able to get TensorFlowSharp to work at all. Is there an alternative way to use a TensorFlow .pb model file from a C# app?

[khanyuriy]

I have the same error, but my model less then 64MB.
How can i be sure, that TensorFlowSharp bindings are compatible with savedmodel file i have got?

[migueldeicaza]

Yeah, that seems to be a common problem with Tensorflow and Protobuf. I believe someone suggested a way somewhere to recompile the underlying protocol buffer implementation in libtensorflow with a larger buffer size. I do not recall where I saw that.

This is what stops one of our samples as well, the file is too large.

[migueldeicaza]

This is the tensorflow issue, it should be addressed if you upgrade your TensorFlow runtime:

tensorflow/tensorflow#582

[codingyoutub]

I have the same problem. I made a pb file with a teachable machine, but it seems to be a compatibility problem. I think it's 2.3 version of pb. Did you find a solution?

[migueldeicaza]

I recommend you now use TensorFlow.NET which is being actively developed.

[Nevey]

I recommend you now use TensorFlow.NET which is being actively developed.

I can't explain why, but it's a pain in the ass getting that library to work in Unity. TensorFlowSharp seems a lot easier to deal with, but I'm also running into the TFException "Invalid GraphDef" issue. You said this has to do with the version of TensorFlow used to create the model? Any tips for me? (I'm running a python project to train data with tf version 2.3.1, then I want to read it in Unity)