582921ffe3b04ff502e1c3a05088ba2902e0f5bd,rl_coach/agents/value_optimization_agent.py,ValueOptimizationAgent,run_off_policy_evaluation,#ValueOptimizationAgent#,104

Before Change

        
        assert self.ope_manager
        dataset_as_episodes = self.call_memory("get_all_complete_episodes_from_to",
                                               (self.call_memory("get_last_training_set_episode_id") + 1,
                                                self.call_memory("num_complete_episodes")))
        if len(dataset_as_episodes) == 0:
            raise ValueError("train_to_eval_ratio is too high causing the evaluation set to be empty. "

After Change

        
        assert self.ope_manager

        if not isinstance(self.pre_network_filter, NoInputFilter) and len(self.pre_network_filter.reward_filters) != 0:
            raise ValueError("Defining a pre-network reward filter when OPEs are calculated will result in a mismatch "
                             "between q values (which are scaled), and actual rewards, which are not. It is advisable "
                             "to use an input_filter, if possible, instead, which will filter the transitions directly "
                             "in the replay buffer, affecting both the q_values and the rewards themselves. ")

        ips, dm, dr, seq_dr, wis = self.ope_manager.evaluate(
                                  evaluation_dataset_as_episodes=self.memory.evaluation_dataset_as_episodes,
                                  evaluation_dataset_as_transitions=self.memory.evaluation_dataset_as_transitions,
                                  batch_size=self.ap.network_wrappers["main"].batch_size,

Instances