TensorFlow: Unable to find entry point 'TF_StringDecode'

[PMGruber]

System information

• OS version/distro: Win10 x64
• .NET Version (eg., dotnet --info): .net Core 3.1
• ML.NET Version: 1.7.0 with SciSharp.TensorFlow.Redist 2.7.0

Issue

• What did you do?
  I load a pretrained TensorFlow model and execute it.

• What happened?
  When executing predict method, I receive an exception "System.EntryPointNotFoundException: 'Unable to find an entry point named 'TF_StringDecode' in DLL 'tensorflow'.'"

• What did you expect?
  Execution without error

Details

I have a simple Tensorflow model that makes a classification. As input the model gets an image (as float array). As output, the model provides a list of probabilities (one entry for each class. Datatype: float array) and a list of classnames (string array).

When I run the prediction, I receive the exception from above with source "TensorFlow.NET" and stack trace

   at Tensorflow.c_api.TF_StringDecode(Byte* src, UInt64 src_len, Byte** dst, UInt64& dst_len, SafeStatusHandle status)
   at Tensorflow.Tensor.StringData()
   at Microsoft.ML.TensorFlow.TensorFlowUtils.FetchStringData[T](Tensor tensor, Span`1 result)
   at Microsoft.ML.Transforms.TensorFlowTransformer.Mapper.<>c__DisplayClass11_0`1.<MakeGetter>b__1(VBuffer`1& dst)
   at Microsoft.ML.Data.TypedCursorable`1.TypedRowBase.<>c__DisplayClass7_0`2.<CreateConvertingVBufferSetter>b__0(TRow row)
   at Microsoft.ML.Data.TypedCursorable`1.TypedRowBase.FillValues(TRow row)
   at Microsoft.ML.Data.TypedCursorable`1.RowImplementation.FillValues(TRow row)
   at Microsoft.ML.PredictionEngineBase`2.FillValues(TDst prediction)
   at Microsoft.ML.PredictionEngine`2.Predict(TSrc example, TDst& prediction)
   at Microsoft.ML.PredictionEngineBase`2.Predict(TSrc example)
   at TensorflowString.Program.Main(String[] args) in 

When I execute the model and just read the column with float array it executes fine (not suprising, since exception states a problem with string type).

Here a code snippet from loading, preparing and executing the model:

            string _assetsPath = Path.Combine(Environment.CurrentDirectory, "assets");
            string _image = Path.Combine(_assetsPath, "input.png");
            string _model = Path.Combine(_assetsPath, "frozen_graph.pb");

            var bitmapImage = new Bitmap(_image);
            var floatImage = VerySlowBitmapTo1dFloatArrayConverter(bitmapImage);

            MLContext mlContext = new MLContext();
            var tensorflowModel = mlContext.Model.LoadTensorFlowModel(_model);

            var inputSchemaDefinition = SchemaDefinition.Create(typeof(ImageData));
            inputSchemaDefinition["Image"].ColumnType = new VectorDataViewType(NumberDataViewType.Single, 750 *500*3);
            inputSchemaDefinition["Image"].ColumnName = "x";

            var outputSchemaDefinition = SchemaDefinition.Create(typeof(ImagePrediction));
            outputSchemaDefinition["Result"].ColumnType = new VectorDataViewType(NumberDataViewType.Single, 15);
            outputSchemaDefinition["Result"].ColumnName = "Identity";
            outputSchemaDefinition["Classnames"].ColumnType = new VectorDataViewType(TextDataViewType.Instance, 15);
            outputSchemaDefinition["Classnames"].ColumnName = "Identity_1";

            var pipeline = tensorflowModel.ScoreTensorFlowModel(outputColumnNames: new[] { "Identity", "Identity_1" }, inputColumnNames: new[] { "x" }, addBatchDimensionInput: false);
            var data = mlContext.Data.LoadFromEnumerable(new List<ImageData>(), inputSchemaDefinition);
            var processingModel = pipeline.Fit(data);

            var imageData = new ImageData()
            {
                Image = floatImage
            };

            var predictor = mlContext.Model.CreatePredictionEngine<ImageData, ImagePrediction>(processingModel, inputSchemaDefinition: inputSchemaDefinition, outputSchemaDefinition: outputSchemaDefinition);
            var prediction = predictor.Predict(imageData);

I uploaded the test project to a private repository (I'm not allowed to make model public). If somebody wants it for a deeper look, I grant access.

I downloaded libtensorflow-cpu-windows-x86_64-2.7.0 from googleapis and had a look at the header files. I just found a comment with "// TF_StringEncode and TF_StringDecode facilitate this encoding." but no definition of "TF_StringDecode".
My C/C++ knowlege is quite low level so I'm stuck at this point.

Therefore I would appreciate any hints or ideas.

Thanks,
Philipp

[luisquintanilla]

Hi @PMGruber,

Sorry to hear you're running into this issue. Can you try using version 2.3.1 of the SciSharp.TensorFlow.Redist NuGet package and see if it's still an issue?

https://www.nuget.org/packages/SciSharp.TensorFlow.Redist/2.3.1

[PMGruber]

Hello @luisquintanilla,

thank you for your fast response.
I installed SciSharp.TensorFlow.Redist 2.3.1 and it is executing, no more exception. Thanks for the hint.

This sounds like a breaking change in a later verstion to me. Unfortunately I can't find any information regarding TensorFlow version that is supported in ML.NET.
Is this kind of information somewhere in documentation? If not, what's the best way to find out (that I can check this if I run in such a situation again)?

Regards
Philipp

[michaelgsharp]

So don't have it in our documentation but it is in some tutorials. We are planning on adding it to some documentation though since this has been coming up more frequently. You can always look in the versions.props file for the exact versions we are using, but hopefully the documentation update will help.

As a side note, I am also working on updating to the latest version. The main blocker is the breaking changes you mentioned. I'm currently in the process of working through them, but it will take a bit of time.

Docs issue tracked here

[PMGruber]

versions.props works fine for me :)
An update to support latest TF version would be awesome.

Thanks a lot for your effeort. You guys making a great job