How to compute gradient of output wrt input in Tensorflow 2.0 when there are multi-labels?

Question

In my case, the input shape is (None, 16) (sample size, feature size), and my output is (None, 3). It means that my DNN is a mapping from input space to 3-D output space. In tha