Dynamic placement

The tf.device() function can be provided with a function name in place of a device string. If a function name is provided, then the function has to return the device string. This way of providing a device string through a custom function allows complex algorithms to be used for placing the variables on different devices. For example, TensorFlow provides a round robin device setter function in tf.train.replica_device_setter().