From 4a79f7e5a76bcbdef7fd43dc768882742cb1fb4c Mon Sep 17 00:00:00 2001 From: Antonin RAFFIN Date: Tue, 24 Dec 2019 13:12:04 +0100 Subject: [PATCH] Print std reward for evaluation --- torchy_baselines/common/evaluation.py | 2 +- torchy_baselines/sac/sac.py | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/torchy_baselines/common/evaluation.py b/torchy_baselines/common/evaluation.py index 3d76d9f..bc8fc70 100644 --- a/torchy_baselines/common/evaluation.py +++ b/torchy_baselines/common/evaluation.py @@ -19,4 +19,4 @@ def evaluate_policy(model, env, n_eval_episodes=10, deterministic=True, render=F env.render() episode_rewards.append(episode_reward) - return np.mean(episode_rewards), n_steps + return np.mean(episode_rewards), np.std(episode_rewards) diff --git a/torchy_baselines/sac/sac.py b/torchy_baselines/sac/sac.py index cd54497..e4f24d2 100644 --- a/torchy_baselines/sac/sac.py +++ b/torchy_baselines/sac/sac.py @@ -291,10 +291,10 @@ class SAC(BaseRLModel): if 0 < eval_freq <= timesteps_since_eval and eval_env is not None: timesteps_since_eval %= eval_freq sync_envs_normalization(self.env, eval_env) - mean_reward, _ = evaluate_policy(self, eval_env, n_eval_episodes) + mean_reward, std_reward = evaluate_policy(self, eval_env, n_eval_episodes) evaluations.append(mean_reward) if self.verbose > 0: - print("Eval num_timesteps={}, mean_reward={:.2f}".format(self.num_timesteps, evaluations[-1])) + print("Eval num_timesteps={}, mean_reward={:.2f}, std_reward={:.2f}".format(self.num_timesteps, mean_reward, std_reward)) print("FPS: {:.2f}".format(self.num_timesteps / (time.time() - self.start_time))) return self