Handling unconnected gradients in Scipy optimizer

Feature request

Our Scipy optimizer wrapper should explicitly handle unconnected gradients.

Motivation

Currently, the GPflow Scipy optimizer wrapper (gpflow.optimizers.Scipy) simply errors (with one of the typically unhelpful TensorFlow error messages) when one of the passed-in variables is not connected to the loss to be minimized. This can easily happen in a valid case, for example when using an SVGP model with a Constant kernel (which might be used as a very simple baseline) but having inducing locations not explicitly set to non-trainable. This could also happen erroneously in a bug case, e.g. when calling Scipy().minimize(model1.training_loss, model2.trainable_variables).

Proposal

Describe the solution you would like An explicit error message when the variables are not connected to the loss (and hence the grads list contains None entries).

What alternatives have you considered? We could also simply change the call for the gradient computation to use unconnected_gradients=tf.UnconnectedGradients.ZERO. This might silently hide bug cases (see above), so perhaps this option could be an extra flag to the minimize() call?

[st--]

Would you like to open a PR so we can discuss more concretely on the code? 😃

[antonykamp]

Ah! You mean to give a warning, yet continue running?

As you pointed out, sometimes its desired behavior eg. in bug cases. Am I right? 😃

Moving the check was a great idea ofc 😄 I would add the new flag as a simple attribute of Scipy-objects (and uncon_gradients_handle stays?)