What if just using the setter of the property Weights of an Layer object?

I'm not sure if this is the right way to do it.
Currently it's throwing a NullReferenceException.

The following code snippet is an example:

        var layers = new LayersApi();

        var sequential = keras.Sequential();
        var inputs = keras.Input(shape: new Shape(8));
        sequential.add(inputs);
        sequential.add(layers.Dense(8, "relu"));

        sequential.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
                            optimizer: keras.optimizers.Adam(),
                            metrics: new[] { "accuracy" });

        float[,] myArray = new float[8, 8]
        {
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
            { 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f },
        };

        Tensor myTensor = new Tensor(myArray);
        sequential.Weights[0] = new ResourceVariable(myTensor);

It's maybe wrong to use the ResourceVariable constructor?

Please use Tensor myTensor = tf.convert_to_tensor(myArray); because directly using constructor of Tensor won't produce an EagerTensor.

I'm not sure if this is the right way to do it.

I agree and we'll add the set_weights API. If you met any other problem when using the setter of Weights, please tell us and we'll help you.

BTW, tf.Variable is recommended compared to using constructor of ResourceVariable.

It works well when use sequential.Weights[0].assign(myTensor);

Thank you all for the useful suggestions.
The following works fine:

Tensor myTensor = tf.convert_to_tensor(myArray);
sequential.Weights[0].assign(myTensor);

Instead with:

Tensor myTensor = new Tensor (myArray);
sequential.Weights[0].assign(myTensor);

I get the following exception:

Tensorflow.RuntimeError: 'Attempting to capture an EagerTensor without building a function.'

The tf.convert_to_tensor function seems to be mandatory.
Provided you're using it, other formulations of the second row (like tf.Variable, or ResourseVariable constructor) seem not to throw exceptions, but weights are actually not updated.

The set_weights API has already been added, close this issue as completed. You can reopen this issue if there's any problem with it.