Fixing optimization variables

[maximilianmordig]

Hi,

I refer to the code at https://github.com/ICL-SML/Doubly-Stochastic-DGP/blob/master/demos/using_natural_gradients.ipynb, specifically:

ng_vars = [[model_nat_grads.layers[-1].q_mu, model_nat_grads.layers[-1].q_sqrt]]
for v in ng_vars[0]:
    v.set_trainable(False)
ng_action = NatGradOptimizer(gamma=1.).make_optimize_action(model_nat_grads, var_list=ng_vars)
adam_action = AdamOptimizer(0.001).make_optimize_action(model_nat_grads)
Loop([ng_action, adam_action], stop=iterations)()

Am I understanding correctly that the NatGradOptimizer ignores the set_trainable(False) whilst AdamOptimizer takes it into account? Is this because it acts at a lower level than GPFlow?

[hughsalimbeni]

The NatGradOptimizer optimizes only the variables in var_list, whereas all the tensorflow optimizers grab all variables with the trainable flag set to True. As you say, the NatGradOptimizer is a lower level implementation (it calls tf.assign directly, rather than let the tensorflow optimizer take care of it) as the update rule cannot be implemented through ordinary gradients.