[Image Classification Predictions (Pretrained TensorFlow model scoring)] Example crashes when loading custom trained .pb model instead of Inception .pb file - "Invalid GraphDef"

Hi there,

I tried the "Image Classification Predictions (Pretrained TensorFlow model scoring)" example. When I load the given Inception model, everything works fine. But when I replace the tensorflow_inception_graph.pb file with a different model I trained myself in Keras, the following exception appears when calling the method LoadTensorFlowModel:

`System.FormatException: "Tensorflow exception triggered while loading model." TensorflowException: Invalid GraphDef`

My custom Keras model is saved to a SavedModel format (also a .pb file with some other files) like this:

`tf.saved_model.save(my_custom_model, "/path/to/saved/model")`

Is the .pb file from the saved_model format different from the .pb format of the .pb-file of the inception model? What can I do to fix this?

This is the call stack:

`Tensorflow.Status.Check(bool)`
`Tensorflow.Graph.Import(byte[], string)`
`Tensorflow.Graph.Import(string, string)`
`Microsoft.ML.TensorFlow.TensorFlowUtils.LoadTFSessionByModelFilePath(Microsoft.ML.Runtime.IExceptionContext, string, bool)`

These are the nuget versions of the ML packages:

Microsoft.ML v1.4.0
Microsoft.ML.ImageAnalytics v1.4.0
Microsoft.ML.TensorFlow v1.4.0
SciSharp.TensorFlow.Redist v1.14.0 (also tried v1.15.0)

Target framework is .NET Core 3.1

Any help is much appreciated, thank you in advance

Inizia con l’esempio Image Classification Predictions e la chiamata LoadTensorFlowModel, quindi segui TensorFlowUtils.LoadTFSessionByModelFilePath fino a Tensorflow.Graph.Import. Confronta l’input Inception .pb previsto con la struttura Keras SavedModel riportata. Done deve documentare o implementare il formato di modello supportato e fornire un percorso funzionante o un errore chiaro per i file non supportati.

Scritto dal modello di indicizzazione a partire dal testo della issue.