Deferred initialization error

[tdiethe]

When using gluon blocks, it's common to not specify the shape of e.g. the inputs to begin with, say:

        # Create net
        self.net = mx.gluon.nn.HybridSequential(prefix=self.prefix)
        with self.net.name_scope():
            for layer in self.network_shape[1:-1]:
                self.net.add(mx.gluon.nn.Dense(layer, activation="relu"))
            #  Last layer for classification
            self.net.add(mx.gluon.nn.Dense(self.network_shape[-1], flatten=True, in_units=self.network_shape[-2]))
        self.net.hybridize()
        self.net.initialize(mx.init.Xavier(magnitude=2.34), ctx=self.ctx)

then I create my MXFusion model as follows:

        self.model = Model()
        self.model.N = Variable()
        self.model.D = Variable()
        self.model.f = MXFusionGluonFunction(self.net, num_outputs=1, broadcastable=False)
        self.model.x = Variable(shape=(self.model.N, self.model.D))
        self.model.r = self.model.f(self.model.x)
        for v in self.model.r.factor.parameters.values():
            v.set_prior(Normal(mean=mx.nd.array([0]), variance=mx.nd.array([3.])))
            self.model.y = Categorical.define_variable(log_prob=self.model.r, shape=(self.model.N, 1), num_classes=2)

I then define my posterior and initialize:

            data = mx.nd.flatten(batch.data[0]).as_in_context(self.ctx)
            labels = batch.label[0].as_in_context(self.ctx)
            N, D = data.shape

            # pass some data to initialise the net
            self.net.forward(data[0, :])

            observed = [self.model.x, self.model.y]
            q = create_Gaussian_meanfield(model=self.model, observed=observed)
            alg = StochasticVariationalInference(num_samples=5, model=self.model, posterior=q, observed=observed)
            inference = GradBasedInference(inference_algorithm=alg, grad_loop=BatchInferenceLoop())
            inference.initialize(N=N, D=D, y=mx.nd.array(labels), x=mx.nd.array(data))

Now I'd like to set the variances to be some small value:

            for v_name, v in self.model.r.factor.parameters.items():
                qv_mean = q[v.uuid].factor.mean
                qv_variance = q[v.uuid].factor.variance
                a = inference.params.param_dict[qv_variance.uuid].data().asnumpy()
                a[:] = 1e-8

But here I get mxnet.gluon.parameter.DeferredInitializationError, since the shape of q_variance is still (256, 0).