Retrained graph unusable

[Robertb84]

Hi, I'am trying to use a retrained graph, out of tensorflows tutorial in a Unity project with the webcamtexture. But as a first test i simply tried to Load a picture and let it run through the graph. I basically tried what is shown in example "ExampleInceptionInference"

I used optimize for inference and transform_graph on it (with different calls and settings, i "collected" some graphs)

The graphs work perfectly in python after and before each editing step (retrain, optimize and transform) My guess is that i'am doing something wrong when i try to transform the picture to a tensor. I'am using the code out of the example with slight changes. (W,H, Mean are different)

        private static void ConstructGraphToNormalizeImage(out TFGraph graph, out TFOutput input, out TFOutput output, TFDataType destinationDataType = TFDataType.Float)
        {

            const int W = 299;
            const int H = 299;
            const float Mean = 128;
            const float Scale = 1;
			graph = new TFGraph();
			input = graph.Placeholder(TFDataType.String);
            output = graph.Cast(graph.Div(
				x: graph.Sub(
                    x: graph.ResizeBilinear(
                        images: graph.ExpandDims(
                            input: graph.Cast(
                                graph.DecodeJpeg(contents: input, channels: 3), DstT: TFDataType.Float),
                            dim: graph.Const(0, "make_batch")),
                        size: graph.Const(new int[] { W, H }, "size")),
                    y: graph.Const(Mean, "mean")),
                y: graph.Const(Scale, "scale")), destinationDataType);
        }

I Init my class with

        var model = File.ReadAllBytes(Utils.getFilePath("dnn/10k_trans_t.pb"));
        g.Import(model, "");
        labels = File.ReadAllLines(Utils.getFilePath("dnn/retrained_labels_t.txt"));
        session = new TFSession(g);
        g_input = g["Mul"][0];
        g_output = g["final_result"][0];

        resultObject = "not sure";
        resultProb = 0;

        _file = Utils.getFilePath("Passion.jpg");

and basically after a few seconds i trigger

 public void HandleJPG()
    {
        
        Debug.Log($"HandlePictures {TensorPics.Count}");
        
        var tensor = ImageUtil.CreateTensorFromImageFile(_file);
        var runner = session.GetRunner();

        runner.AddInput(g_input, tensor).Fetch(g_output);
        var output = runner.Run();

        var bestIdx = 0;
        float best = 0;
        var result = output[0];
        var rshape = result.Shape;
        var probabilities = ((float[][])result.GetValue(jagged: true))[0];
        for (int r = 0; r < probabilities.Length; r++)
        {
            if (probabilities[r] > best)
            {
                bestIdx = r;
                best = probabilities[r];
            }
        }
        Debug.Log("Tensorflow thinks this is: " + labels[bestIdx] + " Prob : " + best * 100);
  }

But in TensorflowSharp i get completely different results.

so in short i have two questions:

1. Has anyone a idea what has to be done differently on a retrained graph?
2. Has anyone tried to transform the webcamtexture into a TFTensor? (could also use Opencvs Mats as a helper)

[hwvs]

I was able to fix this on my own model by changing const float Scale = 1 to const float Scale = 128f

See if it works for you.

[adragoset]

@hwvs I just spent 4 days troubleshooting why my trained models coming out of a python system weren't working correctly in my c# inference code and this was the culprit. Pre processing between c# and python are identical otherwise.

[mg021]

@Robertb84 I Have two Result with python My Model Work Well . with TensorflowSharp my model Is not Goood . Help Me . Is TensorflowSharp Good For Using or ... ??