tf.function decorator removal resolves “Cannot convert a symbolic Tensor (concat:0) to a numpy array” in optimization step

Question

I am supplying different minibatches to optimize a gpflow model (SVGP), if I decorate the optimization_step with tf.function I get the fol