As described here, TF is inflexible when it comes to access to the gradients:
http://stackoverflow.com/questions/35731506/unaggregated-gradients-gradients-per-example-in-tensorflow?rq=1

Please provide a method where the user can retrieve the raw gradients, not the averaged gradients. Requiring the user to compute their own gradients is impractical -- the framework should work for the user, not the other way around.

Use case: this is needed for reinforcement learning, where the gradients of one net needs to be backpropagated through another net (in separate steps).