Load tensorflow model created by python error while 'Visual Studio Tools for AI' well

[slmgong]

System information

• OS version/distro:
  windows 10
  ml.net v0.10.0

• .NET Version (eg., dotnet --info):
  dotcore v2.1

Issue

• What did you do?
• What happened?
• What did you expect?

Source code / logs

I have a tensorflow model, when I use Visual Studio Tools for AI, it auto generator code like this

`

private static List<long> serving_defaultFeaturesShapeForSingleInput = new List<long> { 1, 6, 8 };
private static List<string> serving_defaultInputNames = new List<string> { "features" };
private static List<string> serving_defaultOutputNames = new List<string> { "predictions" };
etc...
/// <summary>
/// Runs inference on abc model for a single input data.
/// </summary>
/// <param name="features">From signature: input 1 with tensor name features; Shape of the input: { 1, 6, 8 }</param>
public IEnumerable<float> Serving_default(IEnumerable<float> features)
{
    List<Tensor> result = manager.RunModel(
        modelName,
        int.MaxValue,
        serving_defaultInputNames,
        new List<Tensor> { new Tensor(features.ToList(), serving_defaultFeaturesShapeForSingleInput) },
        serving_defaultOutputNames
    );
    List<float> r0 = new List<float>();
    result[0].CopyTo(r0);
    return r0;
}     

`
and everything goes well, I run this instance and got a return.

when I use the ML.Net v0.10.0 try a test the same tensorflow model :

`

public class TensorData
{
[VectorType(1, 4, 6)]
[ColumnName("lstm_1_input")]
public float[] input { get; set; }
}

etc.....

var pipeline = mlContext.Transforms.ScoreTensorFlowModel(
   tensorModel, new[] { "lstm_1_input" }, new[] { "lstm_1_input" });
var data = GetTensorData();
var idv = mlContext.Data.ReadFromEnumerable(data);
var trainedModel = pipeline.Fit(idv);
var predicted = trainedModel.Transform(idv);

`

it throw a exception:

System.InvalidOperationException:“Input shape mismatch: Input 'lstm_1_input' has shape [1, 4, 6], but input data is Vec<R4, 1, 4, 6>.”

1、 how to caputure the inputColumnNames/outputColumnNames from a exists model , when Visual Studio Tools for AI use different input/output names ?
2、how to construct a shape [1, 4, 6], when I change input type as float[,,] ,> it throw a

System.InvalidOperationException:“Variable length input columns not supported”.

so how I can correctly use the tensorflow model ? please give a example, thanks much.

[Ivanidzo4ka]

I believe this should be closed by this PR:
#2475
It just didn't get into 0.10 nuget.
Any chance you can test our daily nugets?

[zeahmed]

Yes #2475 should fix the problem you are facing. Regarding inspecting the input you can follow the example here.

machinelearning/docs/samples/Microsoft.ML.Samples/Dynamic/TensorFlow/TextClassification.cs

Line 48 in eb959c3

var modelInfo = TensorFlowUtils.LoadTensorFlowModel(mlContext, modelLocation);

After loading the model with LoadTensorFlowModel, you can query the input schema using GetInputSchema method. It should give you all the input variables in the tf graph with all the necessary information i.e. nodes with optype Placeholder.

[zeahmed]

I am closing it. Please reopen it if necessary.