def test_rnnsac_compilation(self): """Test whether RNNSAC can be built on all frameworks.""" config = ( sac.RNNSACConfig().rollouts(num_rollout_workers=0).training( # Wrap with an LSTM and use a very simple base-model. model={"max_seq_len": 20}, policy_model_config={ "use_lstm": True, "lstm_cell_size": 64, "fcnet_hiddens": [10], "lstm_use_prev_action": True, "lstm_use_prev_reward": True, }, q_model_config={ "use_lstm": True, "lstm_cell_size": 64, "fcnet_hiddens": [10], "lstm_use_prev_action": True, "lstm_use_prev_reward": True, }, replay_buffer_config={ "type": "MultiAgentPrioritizedReplayBuffer", "replay_burn_in": 20, "zero_init_states": True, }, lr=5e-4, )) num_iterations = 1 # Test building an RNNSAC agent in all frameworks. for _ in framework_iterator(config, frameworks="torch"): trainer = config.build(env="CartPole-v0") for i in range(num_iterations): results = trainer.train() print(results) check_compute_single_action( trainer, include_state=True, include_prev_action_reward=True, )
def _import_rnnsac(): from ray.rllib.algorithms import sac return sac.RNNSAC, sac.RNNSACConfig().to_dict()