retrained image classifer in tensorflowsharp?

[bzburr]

Hi there

I've retrained an image classifier using tensorflow ImageNet retraining sample. I have my files

-{DIR}bottleneck -{DIR}retrain_logs -checkpoint -output_graph.pb -output_labels.txt -_retrain_checkpoint.data-00000-of-00001 -_retrain_checkpoint.index -_retrain_checkpoint.meta

When i try and load the .pb file into TensorflowSharp Graph object - get the following mess of an object

What do i need to do to load the retrained imagenet model into Tensorflow sharp and how do i do it?

thanks :)

[migueldeicaza]

Did you use TFSession.FromSavedModel to load?

It needs to be saved like this:

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md

[bzburr]

no i used... theGraph = new TFGraph(); var model = File.ReadAllBytes(classifierModelPath); theGraph.Import(new TFBuffer(model)); ...

[bzburr]

So i have _tfGraph = new TFGraph(); using (var tmpSess = new TFSession(_tfGraph)) using (var tfSessionOptions = new TFSessionOptions()) using (var metaGraphUnused = new TFBuffer()) { //for some reason FromSavedModel is not static _tfSession = tmpSess.FromSavedModel(tfSessionOptions, null, @"D:\AI\savedfolder\", new[] { "serve" }, _tfGraph, metaGraphUnused); }

_tfGraph doesnt look great

my saved folder looks like

my instantiation looks like

 _tfGraph = new TFGraph();
            using (var tmpSess = new TFSession(_tfGraph))
            using (var tfSessionOptions = new TFSessionOptions())
            using (var metaGraphUnused = new TFBuffer())
            {
                //for some reason FromSavedModel is not static
                _tfSession = tmpSess.FromSavedModel(tfSessionOptions, null, @"D:\AI\savedfolder\", new[] { "serve" }, _tfGraph, metaGraphUnused);
            }

my function to call the classifier is

public ClassificationResult ClassifyImageFromSavedModel(string theTempFile)
        {
            var bestIdx = 0;
            float best = 0;
            using (var session = new TFSession(_tfGraph))
            {
                var tensor = ClasssifierImageUtil.CreateTensorFromImageFile(theTempFile, 299);
                var runner = _tfSession.GetRunner();
                runner.AddInput(_tfGraph["image"][0], tensor).Fetch(_tfGraph["prediction"][0]);
                var output = runner.Run();
                var result = output[0];

                var probabilities = ((float[][])result.GetValue(jagged: true))[0];
                for (int i = 0; i < probabilities.Length; i++)
                {
                    if (probabilities[i] > best)
                    {
                        bestIdx = i;
                        best = probabilities[i];
                    }
                }
            }

when i hit

runner.AddInput(_tfGraph["image"][0], tensor).Fetch(_tfGraph["prediction"][0]);

i get an exception..which to be fair isnt unxpected

I'm assuming theres an issue in my saved model...?

[bzburr]

saved_model_cli show --dir thefolder --all gave me the info i needed,.... runner.AddInput(theGraph["Placeholder"][0], tensor).Fetch(theGraph["final_result"][0]);

hurray

Now to test tensorflowsharp vs label_image.py

[bzburr]

well....running label_image.py works and correctly classifies the image given

but when i run the classifier from tensorflowsharp... i get an incorrect classification....hmm tfsharp seems to want to classify any image given to it as index 8.

[bzburr]

This is turning into quite the stream of consciousness - for this that follow you have to head into ConstructGraphToNormalizeImage and you need to examime const float Mean and const float Scale ...and find the values used when the model was trained. Whack those in and the model works a lot better :)

[diwadu]

Hi, could you provide code example how to improve efficiency in comparison to python? In may case is like 49% in TF# vs 90% in Python, so quite a lot.