Keras VariationalGaussianProcess-Layer

Hallo, I've implemented with tf a VGP-fit on my data and I would like to do the same with a keras-model and a VGP-layer. But I can't find any examples, where the kernel is a combination from several kernels and how to pass the result of tfd.VariationalGaussianProcess.optimal_variational_posterior() to the layer.

So, what I have:

variational_loc, variational_scale = (
            tfd.VariationalGaussianProcess.optimal_variational_posterior(
                mean_fn=mean_fn,
                kernel=kernel,
                inducing_index_points=inducing_index_points,
                observation_index_points=timestamp_train,
                observations=train_data.values,
                observation_noise_variance=observation_noise_variance))

observation_noise_variance = tfp.util.TransformedVariable(
                                    initial_value=1,
                                    bijector=tfb.Chain([tfb.Shift(np.float64(1e-6)), tfb.Softplus()]),
                                    dtype=dtype,
                                    name='observation_noise_variance') 

        periodic_amplitude_1y = tfp.util.TransformedVariable(
                                    initial_value=1,
                                    bijector=tfb.Chain([tfb.Shift(np.float64(1e-6)), tfb.Softplus()]),
                                    dtype=dtype,
                                    name='periodic_amplitude_1y')

        periodic_length_scale_1y = tfp.util.TransformedVariable(
                                    initial_value=1,
                                    bijector=tfb.Chain([tfb.Shift(np.float64(1e-6)), tfb.Softplus()]),
                                    dtype=dtype,
                                    name='periodic_length_scale_1y')

        periodic_period_1y = tfp.util.TransformedVariable(
                                    initial_value=24*365,
                                    bijector=tfb.Chain([tfb.Shift(np.float64(1e-6)), tfb.Softplus()]),
                                    dtype=dtype,
                                    name='periodic_period_1y')  # period of 24h * 365d

        # periodic kernel:
        periodic_kernel_1y = (tfk.ExpSinSquared(
                                    amplitude=periodic_amplitude_1y,
                                    length_scale=periodic_length_scale_1y,
                                    period=periodic_period_1y)
        )
...
        kernel = (periodic_kernel_1d * periodic_kernel_1w * periodic_kernel_1y)

And I would like to pass these to the following:

model = tf.keras.Sequential([
            tf.keras.layers.InputLayer(input_shape=[1], dtype=train_data.dtype),
            tf.keras.layers.Dense(1, kernel_initializer='ones', use_bias=False),
            tfp.layers.VariationalGaussianProcess(
                num_inducing_points=num_inducing_points,
                kernel_provider=kernel,
                event_shape=[1],
                inducing_index_points_initializer=inducing_index_points,
                variational_inducing_observations_scale_initializer=variational_scale,
                unconstrained_observation_noise_variance_initializer=observation_noise_variance,
            ),
        ])

My question ist here also, how would I pass variational_loc?

Many thanks for your help!

Start with the tfp.layers.VariationalGaussianProcess entry point and the tfd.VariationalGaussianProcess.optimal_variational_posterior call shown in the issue. Check how kernel_provider and the variational scale initializer are expected to receive values, then determine how variational_loc is represented. Done means a documented working example covering the composite kernel and both variational parameters.