def main(): parser = argparse.ArgumentParser() parser.add_argument('--env', help='environment ID', type=str, default='CartPole-v1') parser.add_argument('-f', '--folder', help='Log folder', type=str, default='trained_agents') parser.add_argument('--algo', help='RL Algorithm', default='ppo2', type=str, required=False, choices=list(ALGOS.keys())) parser.add_argument('-n', '--n-timesteps', help='number of timesteps', default=1000, type=int) parser.add_argument('--n-envs', help='number of environments', default=1, type=int) parser.add_argument('--exp-id', help='Experiment ID (default: -1, no exp folder, 0: latest)', default=-1, type=int) parser.add_argument('--verbose', help='Verbose mode (0: no output, 1: INFO)', default=1, type=int) parser.add_argument('--no-render', action='store_true', default=False, help='Do not render the environment (useful for tests)') parser.add_argument('--deterministic', action='store_true', default=False, help='Use deterministic actions') parser.add_argument('--stochastic', action='store_true', default=False, help='Use stochastic actions (for DDPG/DQN/SAC)') parser.add_argument('--load-best', action='store_true', default=False, help='Load best model instead of last model if available') parser.add_argument('--norm-reward', action='store_true', default=False, help='Normalize reward if applicable (trained with VecNormalize)') parser.add_argument('--seed', help='Random generator seed', type=int, default=0) parser.add_argument('--reward-log', help='Where to log reward', default='', type=str) parser.add_argument('--gym-packages', type=str, nargs='+', default=[], help='Additional external Gym environemnt package modules to import (e.g. gym_minigrid)') parser.add_argument('--render-pybullet', help='Slow down Pybullet simulation to render', default=False) # added by Pierre parser.add_argument('--random-pol', help='Random policy', default=False) # added by Pierre args = parser.parse_args() plot_bool = True plot_dim = 2 log_bool = False if plot_bool: if plot_dim == 2: fig, (ax1, ax2) = plt.subplots(2, 1, sharey=True, figsize=(5, 10)) elif plot_dim == 3: fig = plt.figure() ax = fig.gca(projection='3d') if log_bool: output_df = pd.DataFrame() # Going through custom gym packages to let them register in the global registory for env_module in args.gym_packages: importlib.import_module(env_module) env_id = args.env algo = args.algo folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format(log_path) if not args.random_pol: # added by Pierre model_path = find_saved_model(algo, log_path, env_id, load_best=args.load_best) if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(args.seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams(stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = args.reward_log if args.reward_log != '' else None env = create_test_env(env_id, n_envs=args.n_envs, is_atari=is_atari, stats_path=stats_path, seed=args.seed, log_dir=log_dir, should_render=not args.no_render, hyperparams=hyperparams) # ACER raises errors because the environment passed must have # the same number of environments as the model was trained on. load_env = None if algo == 'acer' else env if not args.random_pol: # added by Pierre model = ALGOS[algo].load(model_path, env=load_env) # if not args.no_render: # env.render(mode="human") # added by Pierre (to work with ReachingJaco-v1) obs = env.reset() # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) deterministic = args.deterministic or algo in ['dqn', 'ddpg', 'sac', 'her', 'td3'] and not args.stochastic episode_reward = 0.0 episode_rewards, episode_lengths = [], [] ep_len = 0 episode = 0 # success_threshold_001 = 0.01 # success_list_001, reachtime_list_001, episode_success_list_001 = [], [], [] # success_threshold_0002 = 0.002 # success_list_0002, reachtime_list_0002, episode_success_list_0002 = [], [], [] # success_threshold_0001 = 0.001 # success_list_0001, reachtime_list_0001, episode_success_list_0001 = [], [], [] # success_threshold_00005 = 0.0005 # success_list_00005, reachtime_list_00005, episode_success_list_00005 = [], [], [] # changed for the paper success_threshold_50 = 0.05 success_list_50, reachtime_list_50, episode_success_list_50 = [], [], [] success_threshold_20 = 0.02 success_list_20, reachtime_list_20, episode_success_list_20 = [], [], [] success_threshold_10 = 0.01 success_list_10, reachtime_list_10, episode_success_list_10 = [], [], [] success_threshold_5 = 0.005 success_list_5, reachtime_list_5, episode_success_list_5 = [], [], [] # For HER, monitor success rate successes = [] state = None for _ in range(args.n_timesteps): if args.random_pol: # Random Agent action = [env.action_space.sample()] else: action, state = model.predict(obs, state=state, deterministic=deterministic) # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) obs, reward, done, infos = env.step(action) if args.render_pybullet: time.sleep(1./30.) # added by Pierre (slow down Pybullet for rendering) if infos[0]['total_distance'] <= success_threshold_50: episode_success_list_50.append(1) else: episode_success_list_50.append(0) if infos[0]['total_distance'] <= success_threshold_20: episode_success_list_20.append(1) else: episode_success_list_20.append(0) if infos[0]['total_distance'] <= success_threshold_10: episode_success_list_10.append(1) else: episode_success_list_10.append(0) if infos[0]['total_distance'] <= success_threshold_5: episode_success_list_5.append(1) else: episode_success_list_5.append(0) if plot_bool: goal = infos[0]['goal position'] tip = infos[0]['tip position'] if plot_dim == 2: ax1.cla() ax1.plot(goal[0], goal[2], marker='x', color='b', linestyle='', markersize=10, label="goal", mew=3) ax1.plot(tip[0], tip[2], marker='o', color='r', linestyle='', markersize=10, label="end effector") circ_1_50 = plt.Circle((goal[0], goal[2]), radius=success_threshold_50, edgecolor='g', facecolor='w', linestyle='--', label="50 mm") circ_1_20 = plt.Circle((goal[0], goal[2]), radius=success_threshold_20, edgecolor='b', facecolor='w', linestyle='--', label="20 mm") circ_1_10 = plt.Circle((goal[0], goal[2]), radius=success_threshold_10, edgecolor='m', facecolor='w', linestyle='--', label="10 mm") circ_1_5 = plt.Circle((goal[0], goal[2]), radius=success_threshold_5, edgecolor='r', facecolor='w', linestyle='--', label="5 mm") ax1.add_patch(circ_1_50) ax1.add_patch(circ_1_20) ax1.add_patch(circ_1_10) ax1.add_patch(circ_1_5) ax1.set_xlim([-0.25, 0.25]) ax1.set_ylim([0, 0.5]) ax1.set_xlabel("x (m)", fontsize=15) ax1.set_ylabel("z (m)", fontsize=15) ax2.cla() ax2.plot(goal[1], goal[2], marker='x', color='b', linestyle='', markersize=10, mew=3) ax2.plot(tip[1], tip[2], marker='o', color='r', linestyle='', markersize=10) circ_2_50 = plt.Circle((goal[1], goal[2]), radius=success_threshold_50, edgecolor='g', facecolor='w', linestyle='--') circ_2_20 = plt.Circle((goal[1], goal[2]), radius=success_threshold_20, edgecolor='b', facecolor='w', linestyle='--') circ_2_10 = plt.Circle((goal[1], goal[2]), radius=success_threshold_10, edgecolor='m', facecolor='w', linestyle='--') circ_2_5 = plt.Circle((goal[1], goal[2]), radius=success_threshold_5, edgecolor='r', facecolor='w', linestyle='--') ax2.add_patch(circ_2_50) ax2.add_patch(circ_2_20) ax2.add_patch(circ_2_10) ax2.add_patch(circ_2_5) ax2.set_xlim([-0.25, 0.25]) ax2.set_ylim([0, 0.5]) ax2.set_xlabel("y (m)", fontsize=15) ax2.set_ylabel("z (m)", fontsize=15) ax1.legend(loc='upper left', bbox_to_anchor=(0, 1.2), ncol=3, fancybox=True, shadow=True) elif plot_dim == 3: ax.cla() ax.plot([tip[0]], [tip[1]], zs=[tip[2]], marker='x', color='b') ax.plot([goal[0]], [goal[1]], zs=[goal[2]], marker='o', color='r', linestyle="None") ax.set_xlim([-0.2, 0.2]) ax.set_ylim([-0.2, 0.2]) ax.set_zlim([0, 0.5]) ax.set_xlabel("x (m)", fontsize=15) ax.set_ylabel("y (m)", fontsize=15) ax.set_zlabel("z (m)", fontsize=15) fig.suptitle("timestep "+str(ep_len)+" | distance to target: "+str(round(infos[0]['total_distance']*1000, 1))+" mm") plt.pause(0.01) # plt.show() if log_bool: dict_log = infos[0] dict_log['action'] = action[0] dict_log['obs'] = obs[0] dict_log['reward'] = reward[0] dict_log['done'] = done[0] dict_log['timestep'] = ep_len dict_log['episode'] = episode output_df = output_df.append(dict_log, ignore_index=True) # if not args.no_render: # env.render('human') episode_reward += reward[0] ep_len += 1 if args.n_envs == 1: # For atari the return reward is not the atari score # so we have to get it from the infos dict if is_atari and infos is not None and args.verbose >= 1: episode_infos = infos[0].get('episode') if episode_infos is not None: print("Atari Episode Score: {:.2f}".format(episode_infos['r'])) print("Atari Episode Length", episode_infos['l']) if done and not is_atari and args.verbose > 0: # NOTE: for env using VecNormalize, the mean reward # is a normalized reward when `--norm_reward` flag is passed print("Episode nb: {} | Episode Reward: {:.2f} | Episode Length: {}".format(episode, episode_reward, ep_len)) # print("Episode Length", ep_len) # commented by Pierre state = None episode_rewards.append(episode_reward) episode_lengths.append(ep_len) # append the last element of the episode success list when episode is done success_list_50.append(episode_success_list_50[-1]) success_list_20.append(episode_success_list_20[-1]) success_list_10.append(episode_success_list_10[-1]) success_list_5.append(episode_success_list_5[-1]) # if the episode is successful and it starts from an unsucessful step, calculate reach time if episode_success_list_50[-1] == True and episode_success_list_50[0] == False: idx = 0 while episode_success_list_50[idx] == False: idx += 1 reachtime_list_50.append(idx) if episode_success_list_20[-1] == True and episode_success_list_20[0] == False: idx = 0 while episode_success_list_20[idx] == False: idx += 1 reachtime_list_20.append(idx) if episode_success_list_10[-1] == True and episode_success_list_10[0] == False: idx = 0 while episode_success_list_10[idx] == False: idx += 1 reachtime_list_10.append(idx) if episode_success_list_5[-1] == True and episode_success_list_5[0] == False: idx = 0 while episode_success_list_5[idx] == False: idx += 1 reachtime_list_5.append(idx) if log_bool: # output_df.to_csv(log_path+"/res_episode_"+str(episode)+".csv", index=False) # slow output_df.to_pickle(log_path+"/res_episode_"+str(episode)+".pkl") # reset for new episode episode_reward = 0.0 ep_len = 0 episode_success_list_50 = [] episode_success_list_20 = [] episode_success_list_10 = [] episode_success_list_5 = [] episode += 1 # Reset also when the goal is achieved when using HER if done or infos[0].get('is_success', False): if args.algo == 'her' and args.verbose > 1: print("Success?", infos[0].get('is_success', False)) # Alternatively, you can add a check to wait for the end of the episode # if done: obs = env.reset() if args.algo == 'her': successes.append(infos[0].get('is_success', False)) episode_reward, ep_len = 0.0, 0 if args.verbose > 0 and len(successes) > 0: print("Success rate: {:.2f}%".format(100 * np.mean(successes))) if args.verbose > 0 and len(episode_rewards) > 0: print("Mean reward: {:.2f} +/- {:.2f}".format(np.mean(episode_rewards), np.std(episode_rewards))) print("success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}".format(success_threshold_50, np.mean(success_list_50), np.mean(reachtime_list_50))) print("success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}".format(success_threshold_20, np.mean(success_list_20), np.mean(reachtime_list_20))) print("success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}".format(success_threshold_10, np.mean(success_list_10), np.mean(reachtime_list_10))) print("success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}".format(success_threshold_5, np.mean(success_list_5), np.mean(reachtime_list_5))) # added by Pierre print("path:", log_path) d = { "Eval mean reward": np.mean(episode_rewards), "Eval std": np.std(episode_rewards), "success ratio 50mm": np.mean(success_list_50), "Average reach time 50mm": np.mean(reachtime_list_50), "success ratio 20mm": np.mean(success_list_20), "Average reach time 20mm": np.mean(reachtime_list_20), "success ratio 10mm": np.mean(success_list_10), "Average reach time 10mm": np.mean(reachtime_list_10), "success ratio 5mm": np.mean(success_list_5), "Average reach time 5mm": np.mean(reachtime_list_5), } df = pd.DataFrame(d, index=[0]) if args.random_pol: df.to_csv("logs/random_policy_0.2M/"+env_id+"/stats.csv", index=False) # make path naming more robust else: df.to_csv(log_path+"/stats.csv", index=False) if args.verbose > 0 and len(episode_lengths) > 0: print("Mean episode length: {:.2f} +/- {:.2f}".format(np.mean(episode_lengths), np.std(episode_lengths))) # Workaround for https://github.com/openai/gym/issues/893 if not args.no_render: if args.n_envs == 1 and 'Bullet' not in env_id and not is_atari and isinstance(env, VecEnv): # DummyVecEnv # Unwrap env while isinstance(env, VecNormalize) or isinstance(env, VecFrameStack): env = env.venv env.envs[0].env.close() else: # SubprocVecEnv env.close()
def main(): parser = argparse.ArgumentParser() parser.add_argument('--env', help='environment ID', type=str, default='CartPole-v1') parser.add_argument('-f', '--folder', help='Log folder', type=str, default='trained_agents') parser.add_argument('--algo', help='RL Algorithm', default='ppo2', type=str, required=False, choices=list(ALGOS.keys())) # parser.add_argument('-n', '--n-timesteps', help='number of timesteps', default=1000, # type=int) parser.add_argument('-n', '--n-episodes', help='number of episodes to collect', default=20, type=int) parser.add_argument('--n-envs', help='number of environments', default=1, type=int) parser.add_argument( '--exp-id', help='Experiment ID (default: -1, no exp folder, 0: latest)', default=-1, type=int) parser.add_argument('--verbose', help='Verbose mode (0: no output, 1: INFO)', default=1, type=int) parser.add_argument( '--no-render', action='store_true', default=False, help='Do not render the environment (useful for tests)') # for deterministic (bool type) parser.add_argument('--deterministic', dest='deterministic', action='store_true') parser.add_argument('--no-deterministic', dest='deterministic', action='store_false') parser.set_defaults(deterministic=True) # true by default # parser.add_argument('--deterministic', action='store_true', default=False, # help='Use deterministic actions') # parser.add_argument('--stochastic', action='store_true', default=False, # help='Use stochastic actions (for DDPG/DQN/SAC)') parser.add_argument( '--norm-reward', action='store_true', default=False, help='Normalize reward if applicable (trained with VecNormalize)') parser.add_argument('--seed', help='Random generator seed', type=int, default=0) parser.add_argument('--reward-log', help='Where to log reward', default='', type=str) parser.add_argument( '--gym-packages', type=str, nargs='+', default=[], help= 'Additional external Gym environemnt package modules to import (e.g. gym_minigrid)' ) args = parser.parse_args() # Going through custom gym packages to let them register in the global registory for env_module in args.gym_packages: importlib.import_module(env_module) env_id = args.env algo = args.algo folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format( log_path) stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams( stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = args.reward_log if args.reward_log != '' else None if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(args.seed) is_atari = 'NoFrameskip' in env_id env = create_test_env(env_id, n_envs=args.n_envs, is_atari=is_atari, stats_path=stats_path, seed=args.seed, log_dir=log_dir, should_render=not args.no_render, hyperparams=hyperparams) model_path = find_saved_model(algo, log_path, env_id) model = ALGOS[algo].load(model_path, env=env) # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) # deterministic = args.deterministic or algo in ['dqn', 'ddpg', 'sac', 'her', 'td3'] and not args.stochastic deterministic = args.deterministic save_dir = os.path.join("expert_trajs_by_info_deterministic_with_std", algo) if not os.path.isdir(save_dir): os.makedirs(save_dir) runner(env, env_id, model, args.n_episodes, deterministic, save=True, save_dir=save_dir)
def main(): parser = argparse.ArgumentParser() parser.add_argument('--env', help='environment ID', type=str, default='CartPole-v1') parser.add_argument('-f', '--folder', help='Log folder', type=str, default='trained_agents') parser.add_argument('--algo', help='RL Algorithm', default='ppo2', type=str, required=False, choices=list(ALGOS.keys())) parser.add_argument('-n', '--n-timesteps', help='number of timesteps', default=1000, type=int) parser.add_argument('--n-envs', help='number of environments', default=1, type=int) parser.add_argument( '--exp-id', help='Experiment ID (default: -1, no exp folder, 0: latest)', default=-1, type=int) parser.add_argument('--verbose', help='Verbose mode (0: no output, 1: INFO)', default=1, type=int) parser.add_argument( '--no-render', action='store_true', default=False, help='Do not render the environment (useful for tests)') parser.add_argument('--deterministic', action='store_true', default=False, help='Use deterministic actions') parser.add_argument('--stochastic', action='store_true', default=False, help='Use stochastic actions (for DDPG/DQN/SAC)') parser.add_argument( '--norm-reward', action='store_true', default=False, help='Normalize reward if applicable (trained with VecNormalize)') parser.add_argument('--seed', help='Random generator seed', type=int, default=0) parser.add_argument('--reward-log', help='Where to log reward', default='', type=str) parser.add_argument( '--gym-packages', type=str, nargs='+', default=[], help= 'Additional external Gym environemnt package modules to import (e.g. gym_minigrid)' ) args = parser.parse_args() # Going through custom gym packages to let them register in the global registory for env_module in args.gym_packages: importlib.import_module(env_module) env_id = args.env algo = args.algo folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format( log_path) model_path = find_saved_model(algo, log_path, env_id) if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(args.seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams( stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = args.reward_log if args.reward_log != '' else None env = create_test_env(env_id, n_envs=args.n_envs, is_atari=is_atari, stats_path=stats_path, seed=args.seed, log_dir=log_dir, should_render=not args.no_render, hyperparams=hyperparams) # ACER raises errors because the environment passed must have # the same number of environments as the model was trained on. load_env = None if algo == 'acer' else env model = ALGOS[algo].load(model_path, env=load_env) obs = env.reset() # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) deterministic = args.deterministic or algo in [ 'dqn', 'ddpg', 'sac', 'her', 'td3' ] and not args.stochastic episode_reward = 0.0 episode_rewards = [] ep_len = 0 # For HER, monitor success rate successes = [] for _ in range(args.n_timesteps): action, _ = model.predict(obs, deterministic=deterministic) # Random Agent # action = [env.action_space.sample()] # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) obs, reward, done, infos = env.step(action) if not args.no_render: env.render('human') episode_reward += reward[0] ep_len += 1 if args.n_envs == 1: # For atari the return reward is not the atari score # so we have to get it from the infos dict if is_atari and infos is not None and args.verbose >= 1: episode_infos = infos[0].get('episode') if episode_infos is not None: print("Atari Episode Score: {:.2f}".format( episode_infos['r'])) print("Atari Episode Length", episode_infos['l']) if done and not is_atari and args.verbose > 0: # NOTE: for env using VecNormalize, the mean reward # is a normalized reward when `--norm_reward` flag is passed print("Episode Reward: {:.2f}".format(episode_reward)) print("Episode Length", ep_len) episode_rewards.append(episode_reward) episode_reward = 0.0 ep_len = 0 # Reset also when the goal is achieved when using HER if done or infos[0].get('is_success', False): if args.algo == 'her' and args.verbose > 1: print("Success?", infos[0].get('is_success', False)) # Alternatively, you can add a check to wait for the end of the episode # if done: obs = env.reset() if args.algo == 'her': successes.append(infos[0].get('is_success', False)) episode_reward, ep_len = 0.0, 0 if args.verbose > 0 and len(successes) > 0: print("Success rate: {:.2f}%".format(100 * np.mean(successes))) if args.verbose > 0 and len(episode_rewards) > 0: print("Mean reward: {:.2f}".format(np.mean(episode_rewards))) # Workaround for https://github.com/openai/gym/issues/893 if not args.no_render: if args.n_envs == 1 and 'Bullet' not in env_id and not is_atari and isinstance( env, VecEnv): # DummyVecEnv # Unwrap env while isinstance(env, VecNormalize) or isinstance( env, VecFrameStack): env = env.venv env.envs[0].env.close() else: # SubprocVecEnv env.close()
def main(): seed = 0 num_samples = 20 parser = argparse.ArgumentParser() parser.add_argument('--env', help='environment ID', type=str, default='CartPole-v1') parser.add_argument('-f', '--folder', help='Log folder', type=str, default='rl-baselines-zoo/trained_agents') parser.add_argument('--algo', help='RL Algorithm', default='dqn', type=str, required=False, choices=list(ALGOS.keys())) parser.add_argument('-n', '--n-timesteps', help='number of timesteps', default=2000, type=int) parser.add_argument('--n-envs', help='number of environments', default=1, type=int) parser.add_argument( '--exp-id', help='Experiment ID (default: -1, no exp folder, 0: latest)', default=-1, type=int) parser.add_argument('--verbose', help='Verbose mode (0: no output, 1: INFO)', default=1, type=int) parser.add_argument( '--no-render', action='store_true', default=False, help='Do not render the environment (useful for tests)') parser.add_argument('--deterministic', action='store_true', default=False, help='Use deterministic actions') parser.add_argument('--stochastic', action='store_true', default=False, help='Use stochastic actions (for DDPG/DQN/SAC)') parser.add_argument( '--load-best', action='store_true', default=False, help='Load best model instead of last model if available') parser.add_argument( '--norm-reward', action='store_true', default=False, help='Normalize reward if applicable (trained with VecNormalize)') args = parser.parse_args() env_id = args.env algo = args.algo folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format( log_path) model_path = find_saved_model(algo, log_path, env_id, load_best=args.load_best) if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams( stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = None env_kwargs = {} env = create_test_env(env_id, n_envs=args.n_envs, is_atari=is_atari, stats_path=stats_path, seed=seed, log_dir=log_dir, should_render=not args.no_render, hyperparams=hyperparams, env_kwargs=env_kwargs) # ACER raises errors because the environment passed must have # the same number of environments as the model was trained on. load_env = None if algo == 'acer' else env model = ALGOS[algo].load(model_path, env=load_env) env = gym.make('CartPole-v1') obs = env.reset() # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) deterministic = args.deterministic or algo in [ 'dqn', 'ddpg', 'sac', 'her', 'td3' ] and not args.stochastic episode_reward = 0.0 episode_rewards, episode_lengths = [], [] ep_len = 0 # For HER, monitor success rate successes = [] state = None embedder = indicator_feature halfspaces = {} for i in range(num_samples): print("+" * 10) #sample random state to start in #TODO: maybe reset with random actions? How to make it realistic? Does it matter. Let's just try random for now to test weird edge cases. obs = env.reset(uniform=True) #sample more uniformly than typical print("start state", obs) # input() #obs = env.reset_state(env.observation_space.sample()) #rollout once for each action and compute feature counts start_state = obs.copy() fcount_vectors = [] init_actions = [] ##rollout code: for init_action in range(env.action_space.n): print("ACTION", init_action) obs = env.reset(start_state=start_state) print("init state", obs) env.render() # input() ep_ret = 0 fcounts = embedder(start_state) #do initial action obs, r, done, info = env.step(init_action) # take a random action fcounts += embedder(obs) #TODO: discount?? ep_ret += r #print(r, obs) if done: print("final state", obs) print("return", ep_ret) print("fcounts", fcounts) fcount_vectors.append(fcounts) init_actions.append(init_action) continue #run tester policy thereafter while True: #env.render() #TODO: sample within allowable range of angle and position action, state = model.predict(obs, state=state, deterministic=deterministic) # Random Agent # action = [env.action_space.sample()] # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) #a = env.action_space.sample() #print(obs, action) obs, r, done, info = env.step(action) # take a random action fcounts += embedder(obs) #print(obs) #print(done) ep_ret += r #print(r, obs) if done: print("final state", obs) print("return", ep_ret) print("fcounts", fcounts) fcount_vectors.append(fcounts) init_actions.append(init_action) break print("action {} over {} => fcount diff = {}".format( init_actions[0], init_actions[1], fcount_vectors[0] - fcount_vectors[1])) halfspaces[state, init_actions[0], init_actions[1]] = fcount_vectors[0] - fcount_vectors[1] # input() #TODO: put this inside one of the value alignment verification classes to get sa_fcount_diffs and hopefully reuse that code #then visualize test cases # input() # for _ in range(args.n_timesteps): # action, state = model.predict(obs, state=state, deterministic=deterministic) # # Random Agent # # action = [env.action_space.sample()] # # Clip Action to avoid out of bound errors # if isinstance(env.action_space, gym.spaces.Box): # action = np.clip(action, env.action_space.low, env.action_space.high) # obs, reward, done, infos = env.step(action) # if not args.no_render: # env.render('human') # episode_reward += reward # ep_len += 1 # if args.n_envs == 1: # # For atari the return reward is not the atari score # # so we have to get it from the infos dict # if is_atari and infos is not None and args.verbose >= 1: # episode_infos = infos.get('episode') # if episode_infos is not None: # print("Atari Episode Score: {:.2f}".format(episode_infos['r'])) # print("Atari Episode Length", episode_infos['l']) # if done and not is_atari and args.verbose > 0: # # NOTE: for env using VecNormalize, the mean reward # # is a normalized reward when `--norm_reward` flag is passed # print("Episode Reward: {:.2f}".format(episode_reward)) # print("Episode Length", ep_len) # state = None # episode_rewards.append(episode_reward) # episode_lengths.append(ep_len) # episode_reward = 0.0 # ep_len = 0 # # Reset also when the goal is achieved when using HER # if done or infos.get('is_success', False): # if args.algo == 'her' and args.verbose > 1: # print("Success?", infos[0].get('is_success', False)) # # Alternatively, you can add a check to wait for the end of the episode # # if done: # obs = env.reset() # if args.algo == 'her': # successes.append(infos[0].get('is_success', False)) # episode_reward, ep_len = 0.0, 0 # if args.verbose > 0 and len(successes) > 0: # print("Success rate: {:.2f}%".format(100 * np.mean(successes))) # if args.verbose > 0 and len(episode_rewards) > 0: # print("Mean reward: {:.2f} +/- {:.2f}".format(np.mean(episode_rewards), np.std(episode_rewards))) # if args.verbose > 0 and len(episode_lengths) > 0: # print("Mean episode length: {:.2f} +/- {:.2f}".format(np.mean(episode_lengths), np.std(episode_lengths))) # Workaround for https://github.com/openai/gym/issues/893 if not args.no_render: if args.n_envs == 1 and 'Bullet' not in env_id and not is_atari and isinstance( env, VecEnv): # DummyVecEnv # Unwrap env while isinstance(env, VecNormalize) or isinstance( env, VecFrameStack): env = env.venv env.envs[0].env.close() else: # SubprocVecEnv env.close()
def main(): parser = argparse.ArgumentParser() parser.add_argument('--env', help='environment ID', type=str, default='CartPole-v1') parser.add_argument('-f', '--folder', help='Log folder', type=str, default='trained_agents') parser.add_argument('--algo', help='RL Algorithm', default='ppo2', type=str, required=False, choices=list(ALGOS.keys())) parser.add_argument('-n', '--n-timesteps', help='number of timesteps', default=1000, type=int) parser.add_argument('--n-envs', help='number of environments', default=1, type=int) parser.add_argument( '--exp-id', help='Experiment ID (default: -1, no exp folder, 0: latest)', default=-1, type=int) parser.add_argument('--verbose', help='Verbose mode (0: no output, 1: INFO)', default=1, type=int) parser.add_argument( '--no-render', action='store_true', default=False, help='Do not render the environment (useful for tests)') parser.add_argument('--deterministic', action='store_true', default=False, help='Use deterministic actions') parser.add_argument('--stochastic', action='store_true', default=False, help='Use stochastic actions (for DDPG/DQN/SAC)') parser.add_argument( '--load-best', action='store_true', default=False, help='Load best model instead of last model if available') parser.add_argument( '--norm-reward', action='store_true', default=False, help='Normalize reward if applicable (trained with VecNormalize)') parser.add_argument('--seed', help='Random generator seed', type=int, default=0) parser.add_argument('--reward-log', help='Where to log reward', default='', type=str) parser.add_argument( '--gym-packages', type=str, nargs='+', default=[], help= 'Additional external Gym environemnt package modules to import (e.g. gym_minigrid)' ) parser.add_argument( '--env-kwargs', type=str, nargs='+', action=StoreDict, help='Optional keyword argument to pass to the env constructor') parser.add_argument('--render-pybullet', help='Slow down Pybullet simulation to render', default=False) # added by Pierre parser.add_argument('--random-pol', help='Random policy', default=False) # added by Pierre parser.add_argument( '--log-dir-random', help='Log directory of the random policy') # added by Pierre args = parser.parse_args() # Going through custom gym packages to let them register in the global registory for env_module in args.gym_packages: importlib.import_module(env_module) env_id = args.env algo = args.algo folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format( log_path) if not args.random_pol: # added by Pierre model_path = find_saved_model(algo, log_path, env_id, load_best=args.load_best) if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(args.seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams( stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = args.reward_log if args.reward_log != '' else None env_kwargs = {} if args.env_kwargs is None else args.env_kwargs env = create_test_env(env_id, n_envs=args.n_envs, is_atari=is_atari, stats_path=stats_path, seed=args.seed, log_dir=log_dir, should_render=not args.no_render, hyperparams=hyperparams, env_kwargs=env_kwargs) # ACER raises errors because the environment passed must have # the same number of environments as the model was trained on. load_env = None if algo == 'acer' else env if not args.random_pol: # added by Pierre model = ALGOS[algo].load(model_path, env=load_env) obs = env.reset() # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) deterministic = args.deterministic or algo in [ 'dqn', 'ddpg', 'sac', 'her', 'td3' ] and not args.stochastic # INITIALISE METRICS episode_reward = 0.0 episode_rewards, episode_lengths = [], [] ep_len = 0 success_threshold_50 = 0.05 success_list_50, reachtime_list_50, episode_success_list_50 = [], [], [] success_threshold_20 = 0.02 success_list_20, reachtime_list_20, episode_success_list_20 = [], [], [] success_threshold_10 = 0.01 success_list_10, reachtime_list_10, episode_success_list_10 = [], [], [] success_threshold_5 = 0.005 success_list_5, reachtime_list_5, episode_success_list_5 = [], [], [] # For HER, monitor success rate successes = [] state = None for _ in range(args.n_timesteps): # Added by Pierre if args.random_pol: action = [env.action_space.sample()] # Random Agent else: action, state = model.predict(obs, state=state, deterministic=deterministic) # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) obs, reward, done, infos = env.step(action) # if args.render_pybullet: # time.sleep(1./30.) # added by Pierre (slow down Pybullet for rendering) # added by Pierre if infos[0]['dist_ft_t'] <= success_threshold_50: episode_success_list_50.append(1) else: episode_success_list_50.append(0) if infos[0]['dist_ft_t'] <= success_threshold_20: episode_success_list_20.append(1) else: episode_success_list_20.append(0) if infos[0]['dist_ft_t'] <= success_threshold_10: episode_success_list_10.append(1) else: episode_success_list_10.append(0) if infos[0]['dist_ft_t'] <= success_threshold_5: episode_success_list_5.append(1) else: episode_success_list_5.append(0) if not args.no_render: env.render('human') # env.render(mode="human") episode_reward += reward[0] ep_len += 1 if args.n_envs == 1: # For atari the return reward is not the atari score # so we have to get it from the infos dict if is_atari and infos is not None and args.verbose >= 1: episode_infos = infos[0].get('episode') if episode_infos is not None: print("Atari Episode Score: {:.2f}".format( episode_infos['r'])) print("Atari Episode Length", episode_infos['l']) if done and not is_atari and args.verbose > 0: # NOTE: for env using VecNormalize, the mean reward # is a normalized reward when `--norm_reward` flag is passed print("Episode Reward: {:.2f}".format(episode_reward)) print("Episode Length", ep_len) episode_rewards.append(episode_reward) episode_lengths.append(ep_len) # Pierre: append the last element of the episode success list when episode is done success_list_50.append(episode_success_list_50[-1]) success_list_20.append(episode_success_list_20[-1]) success_list_10.append(episode_success_list_10[-1]) success_list_5.append(episode_success_list_5[-1]) # if the episode is successful and it starts from an unsucessful step, calculate reach time if episode_success_list_50[ -1] == True and episode_success_list_50[0] == False: idx = 0 while episode_success_list_50[idx] == False: idx += 1 reachtime_list_50.append(idx) if episode_success_list_20[ -1] == True and episode_success_list_20[0] == False: idx = 0 while episode_success_list_20[idx] == False: idx += 1 reachtime_list_20.append(idx) if episode_success_list_10[ -1] == True and episode_success_list_10[0] == False: idx = 0 while episode_success_list_10[idx] == False: idx += 1 reachtime_list_10.append(idx) if episode_success_list_5[ -1] == True and episode_success_list_5[0] == False: idx = 0 while episode_success_list_5[idx] == False: idx += 1 reachtime_list_5.append(idx) # RESET FOR NEW EPISODE state = None episode_reward = 0.0 ep_len = 0 episode_success_list_50 = [] episode_success_list_20 = [] episode_success_list_10 = [] episode_success_list_5 = [] # Reset also when the goal is achieved when using HER if done or infos[0].get('is_success', False): if args.algo == 'her' and args.verbose > 1: print("Success?", infos[0].get('is_success', False)) # Alternatively, you can add a check to wait for the end of the episode # if done: obs = env.reset() if args.algo == 'her': successes.append(infos[0].get('is_success', False)) episode_reward, ep_len = 0.0, 0 if args.verbose > 0 and len(successes) > 0: print("Success rate: {:.2f}%".format(100 * np.mean(successes))) if args.verbose > 0 and len(episode_rewards) > 0: print("Mean reward: {:.2f} +/- {:.2f}".format(np.mean(episode_rewards), np.std(episode_rewards))) print( "success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}" .format(success_threshold_50, np.mean(success_list_50), np.mean(reachtime_list_50))) print( "success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}" .format(success_threshold_20, np.mean(success_list_20), np.mean(reachtime_list_20))) print( "success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}" .format(success_threshold_10, np.mean(success_list_10), np.mean(reachtime_list_10))) print( "success threshold: {} | success ratio: {:.2f} | Average reach time: {:.2f}" .format(success_threshold_5, np.mean(success_list_5), np.mean(reachtime_list_5))) if args.verbose > 0 and len(episode_lengths) > 0: print("Mean episode length: {:.2f} +/- {:.2f}".format( np.mean(episode_lengths), np.std(episode_lengths))) # added by Pierre print("path:", log_path) d = { "Eval mean reward": np.mean(episode_rewards), "Eval std": np.std(episode_rewards), "success ratio 50mm": np.mean(success_list_50), "Average reach time 50mm": np.mean(reachtime_list_50), "success ratio 20mm": np.mean(success_list_20), "Average reach time 20mm": np.mean(reachtime_list_20), "success ratio 10mm": np.mean(success_list_10), "Average reach time 10mm": np.mean(reachtime_list_10), "success ratio 5mm": np.mean(success_list_5), "Average reach time 5mm": np.mean(reachtime_list_5), } df = pd.DataFrame(d, index=[0]) if args.random_pol: log_rand = args.log_dir_random df.to_csv(log_rand + "/stats.csv", index=False) else: df.to_csv(log_path + "/stats.csv", index=False) # Workaround for https://github.com/openai/gym/issues/893 if not args.no_render: if args.n_envs == 1 and 'Bullet' not in env_id and not is_atari and isinstance( env, VecEnv): # DummyVecEnv # Unwrap env while isinstance(env, VecNormalize) or isinstance( env, VecFrameStack): env = env.venv env.envs[0].env.close() else: # SubprocVecEnv env.close()
def rollout_halfspaces(env_id='CartPole-v1',algo='dqn',num_samples=20, precision=0.0001, render=False): seed = 0 folder = 'rl-baselines-zoo/trained_agents' n_envs = 1 no_render = False deterministic = True stochastic = False norm_reward=False log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format(log_path) model_path = find_saved_model(algo, log_path, env_id, load_best=False) set_global_seeds(seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams(stats_path, norm_reward=norm_reward, test_mode=True) log_dir = None env_kwargs = {} env = create_test_env(env_id, n_envs=n_envs, is_atari=is_atari, stats_path=stats_path, seed=seed, log_dir=log_dir, should_render=not no_render, hyperparams=hyperparams, env_kwargs=env_kwargs) # ACER raises errors because the environment passed must have # the same number of environments as the model was trained on. load_env = None if algo == 'acer' else env model = ALGOS[algo].load(model_path, env=load_env) env = gym.make('CartPole-v1') obs = env.reset() # Force deterministic for DQN, DDPG, SAC and HER (that is a wrapper around) deterministic = deterministic or algo in ['dqn', 'ddpg', 'sac', 'her', 'td3'] and not stochastic episode_reward = 0.0 episode_rewards, episode_lengths = [], [] ep_len = 0 # For HER, monitor success rate successes = [] state = None embedder = indicator_feature halfspaces = {} for i in range(num_samples): print("+"*10) #sample random state to start in #TODO: maybe reset with random actions? How to make it realistic? Does it matter. Let's just try random for now to test weird edge cases. obs = env.reset(uniform=True) #sample more uniformly than typical start_state = obs.copy() print("start state", obs) #find out the "near optimal" action for this state to compare other actions to opt_action, _ = model.predict(obs, state=state, deterministic=deterministic) #take this action print("TEACHER ACTION", opt_action) obs = env.reset(start_state=start_state) print("init state", obs) if render: env.render() # input() ep_ret = 0 fcounts = embedder(start_state) #do initial action obs, r, done, info = env.step(opt_action) # take a random action fcounts += embedder(obs) #TODO: discount?? ep_ret += r #print(r, obs) if done: #sample again, since started with terminal state continue #run tester policy thereafter while True: #env.render() #TODO: sample within allowable range of angle and position action, state = model.predict(obs, state=state, deterministic=deterministic) # Random Agent # action = [env.action_space.sample()] # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) #a = env.action_space.sample() #print(obs, action) obs, r, done, info = env.step(action) # take a random action fcounts += embedder(obs) #print(obs) #print(done) ep_ret += r #print(r, obs) if done: print("final state", obs) print("return", ep_ret) print("fcounts", fcounts) opt_fcounts = fcounts break # input() #obs = env.reset_state(env.observation_space.sample()) #rollout once for each action and compute feature counts fcount_vectors = [] init_actions = [] ##rollout code: for init_action in range(env.action_space.n): if init_action == opt_action: #don't need to roll this out since we already did continue print("ACTION", init_action) obs = env.reset(start_state=start_state) print("init state", obs) if render: env.render() # input() ep_ret = 0 fcounts = embedder(start_state) #do initial action obs, r, done, info = env.step(init_action) # take a random action fcounts += embedder(obs) #TODO: discount?? ep_ret += r #print(r, obs) if done: print("final state", obs) print("return", ep_ret) print("fcounts", fcounts) fcount_vectors.append(fcounts) init_actions.append(init_action) continue #run tester policy thereafter while True: #env.render() #TODO: sample within allowable range of angle and position action, state = model.predict(obs, state=state, deterministic=deterministic) # Random Agent # action = [env.action_space.sample()] # Clip Action to avoid out of bound errors if isinstance(env.action_space, gym.spaces.Box): action = np.clip(action, env.action_space.low, env.action_space.high) #a = env.action_space.sample() #print(obs, action) obs, r, done, info = env.step(action) # take a random action fcounts += embedder(obs) #print(obs) #print(done) ep_ret += r #print(r, obs) if done: print("final state", obs) print("return", ep_ret) print("fcounts", fcounts) break normal_vector = opt_fcounts - fcounts print("action {} over {} => fcount diff = {}".format(opt_fcounts, init_action, normal_vector)) if np.linalg.norm(normal_vector) > precision: halfspaces[tuple(start_state), init_action, opt_action] = normal_vector input() #TODO: put this inside one of the value alignment verification classes to get sa_fcount_diffs and hopefully reuse that code #then visualize test cases # input() # for _ in range(args.n_timesteps): # action, state = model.predict(obs, state=state, deterministic=deterministic) # # Random Agent # # action = [env.action_space.sample()] # # Clip Action to avoid out of bound errors # if isinstance(env.action_space, gym.spaces.Box): # action = np.clip(action, env.action_space.low, env.action_space.high) # obs, reward, done, infos = env.step(action) # if not args.no_render: # env.render('human') # episode_reward += reward # ep_len += 1 # if args.n_envs == 1: # # For atari the return reward is not the atari score # # so we have to get it from the infos dict # if is_atari and infos is not None and args.verbose >= 1: # episode_infos = infos.get('episode') # if episode_infos is not None: # print("Atari Episode Score: {:.2f}".format(episode_infos['r'])) # print("Atari Episode Length", episode_infos['l']) # if done and not is_atari and args.verbose > 0: # # NOTE: for env using VecNormalize, the mean reward # # is a normalized reward when `--norm_reward` flag is passed # print("Episode Reward: {:.2f}".format(episode_reward)) # print("Episode Length", ep_len) # state = None # episode_rewards.append(episode_reward) # episode_lengths.append(ep_len) # episode_reward = 0.0 # ep_len = 0 # # Reset also when the goal is achieved when using HER # if done or infos.get('is_success', False): # if args.algo == 'her' and args.verbose > 1: # print("Success?", infos[0].get('is_success', False)) # # Alternatively, you can add a check to wait for the end of the episode # # if done: # obs = env.reset() # if args.algo == 'her': # successes.append(infos[0].get('is_success', False)) # episode_reward, ep_len = 0.0, 0 # if args.verbose > 0 and len(successes) > 0: # print("Success rate: {:.2f}%".format(100 * np.mean(successes))) # if args.verbose > 0 and len(episode_rewards) > 0: # print("Mean reward: {:.2f} +/- {:.2f}".format(np.mean(episode_rewards), np.std(episode_rewards))) # if args.verbose > 0 and len(episode_lengths) > 0: # print("Mean episode length: {:.2f} +/- {:.2f}".format(np.mean(episode_lengths), np.std(episode_lengths))) # Workaround for https://github.com/openai/gym/issues/893 if not no_render: if n_envs == 1 and 'Bullet' not in env_id and not is_atari and isinstance(env, VecEnv): # DummyVecEnv # Unwrap env while isinstance(env, VecNormalize) or isinstance(env, VecFrameStack): env = env.venv env.envs[0].env.close() else: # SubprocVecEnv env.close() return halfspaces
folder = args.folder if args.exp_id == 0: args.exp_id = get_latest_run_id(os.path.join(folder, algo), env_id) print('Loading latest experiment, id={}'.format(args.exp_id)) # Sanity checks if args.exp_id > 0: log_path = os.path.join(folder, algo, '{}_{}'.format(env_id, args.exp_id)) else: log_path = os.path.join(folder, algo) assert os.path.isdir(log_path), "The {} folder was not found".format(log_path) model_path = find_saved_model(algo, log_path, env_id, load_best=args.load_best) if algo in ['dqn', 'ddpg', 'sac', 'td3']: args.n_envs = 1 set_global_seeds(args.seed) is_atari = 'NoFrameskip' in env_id stats_path = os.path.join(log_path, env_id) hyperparams, stats_path = get_saved_hyperparams(stats_path, norm_reward=args.norm_reward, test_mode=True) log_dir = args.reward_log if args.reward_log != '' else None env_kwargs = {} if args.env_kwargs is None else args.env_kwargs