Skip to content

TF2 - Implement tf.saved_model.load #63

Open

Description

@IamSierraCharlie

Hello, as per the subject, are there plans to implement TF2 functions?

For example:

   tf.saved_model.load 

This is part of the code I would like to implement in C# from python

   image_np = load_image_into_numpy_array(image_path)
   input_tensor = tf.convert_to_tensor(image_np)
   input_tensor = input_tensor[tf.newaxis, ...]
   detections = detect_fn(input_tensor)
   num_detections = int(detections.pop('num_detections'))
   detections = {key: value[0, :num_detections].numpy() for key, value in detections.items()}
   detections['num_detections'] = num_detections
   detections['detection_classes'] = detections['detection_classes'].astype(np.int64)

Some of this I can do, but now with TF2 being the standard (no frozen inference graph), it is difficult to proceed with building models in the old standard.

Just want to know if there are any plans as I cant find anything in the documentation that matches this use case from python.

Many thanks for your work so far.

Activity

  1. Esther2013 commented on Jan 7, 2022

    @Esther2013
    Member

    I'm using below alternate solution:

    var model = BuildModel();
    model.load_weights("");
    
  2. IamSierraCharlie commented on Jan 7, 2022

    @IamSierraCharlie
    Author

    Thanks for your response.
    Can you elaborate on "BuildModel()"

    Are you using this function? or something like it?

    Functional BuildModel()
            {
                tf.Context.reset_context();
                var inputs = keras.Input(shape: 2);
    
                // 1st dense layer
                var DenseLayer = keras.layers.Dense(1, activation: keras.activations.Sigmoid);
                var outputs = DenseLayer.Apply(inputs);
    
                // build keras model
                Functional model = keras.Model(inputs, outputs, name: Guid.NewGuid().ToString());
                // show model summary
                model.summary();
    
                // compile keras model into tensorflow's static graph
                model.compile(loss: keras.losses.MeanSquaredError(name: Guid.NewGuid().ToString()),
                    optimizer: keras.optimizers.Adam(name: Guid.NewGuid().ToString()),
                    metrics: new[] { "accuracy" });
                return model;
            }
    

    as found here? https://github.com/SciSharp/TensorFlow.NET/blob/425b258ccebf10e21f033e5c11da7a64bfc21a9f/test/TensorFlowNET.Keras.UnitTest/MultiThreadsTest.cs

  3. Esther2013 commented on Jan 7, 2022

    @Esther2013
    Member

    Correct.

  4. IamSierraCharlie commented on Jan 24, 2022

    @IamSierraCharlie
    Author

    Thankyou for your direction. I'm just not sure I understand how it works. Are there any other examples or further explanation on how to do this?
    Is it just a matter of setting keras.Input(shape: 2) to be the same shape as what I would input into my own model or is there more to it than that?

    I note that when I export my tf built model, I see this warning:

    WARNING:tensorflow:Skipping full serialization of Keras layer <keras.layers.core.lambda_layer.Lambda object at 0x7f5c9c357910>, because it is not built.
    W0124 21:12:34.170078 140039099339136 save_impl.py:71] Skipping full serialization of Keras layer <keras.layers.core.lambda_layer.Lambda object at 0x7f5c9c357910>, because it is not built.

    Should I still be able to build / convert my model in this suggested way?

    Also, I really would like to know the answer to my first question:

    are there plans to implement TF2 functions?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions