def make_obs_ph(name): """ makes the observation placeholder :param name: (str) the placeholder name :return: (TensorFlow Tensor) the placeholder """ return ObservationInput(observation_space_shape, name=name)
def make_obs_ph(name): return ObservationInput(observation_space, name=name)
mode = input("Press 1 for NN expert mode\nPress 2 for Human expert mode\nAnything else for no expert mode\n--> ") if mode == "1": with open('expert_demonstrations_NN.csv', 'r', newline='') as csvfile: data_reader = csv.reader(csvfile, delimiter=',') exp_demo = ([r for r in data_reader]) elif mode == "2": with open('expert_demonstrations_Human.csv', 'r', newline='') as csvfile: data_reader = csv.reader(csvfile, delimiter=',') exp_demo = ([r for r in data_reader]) else: exp_demo = [] # Create all the functions necessary to train the model act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name), q_func=model, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=5e-4), gamma=0.99, ) # Create the replay buffer replay_buffer = ReplayBuffer(50000) # Create the schedule for exploration starting from 1 (every action is random) down to # 0.02 (98% of actions are selected according to values predicted by the model). exploration = LinearSchedule(schedule_timesteps=10000, initial_p=1.0, final_p=0.02) # Initialize the parameters and copy them to the target network. U.initialize() update_target()
def train_policy(arglist): with U.single_threaded_session(): # Create the environment if arglist.use_dense_rewards: print("Will use env MineRLNavigateDense-v0") env = gym.make("MineRLNavigateDense-v0") env_name = "MineRLNavigateDense-v0" else: print("Will use env MineRLNavigate-v0") env = gym.make('MineRLNavigate-v0') env_name = "MineRLNavigate-v0" if arglist.force_forward: env = MineCraftWrapperSimplified(env) else: env = MineCraftWrapper(env) if not arglist.use_demonstrations: # Use stack of last 4 frames as obs env = FrameStack(env, 4) # Create all the functions necessary to train the model act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name), q_func=build_q_func('conv_only', dueling=True), num_actions=env.action_space.n, gamma=0.9, optimizer=tf.train.AdamOptimizer(learning_rate=5e-4), ) # Create the replay buffer(s) (TODO: Use prioritized replay buffer) if arglist.use_demonstrations: replay_buffer = ReplayBuffer(int(arglist.replay_buffer_len / 2)) demo_buffer = load_demo_buffer(env_name, int(arglist.replay_buffer_len / 2)) else: replay_buffer = ReplayBuffer(arglist.replay_buffer_len) # Create the schedule for exploration starting from 1 (every action is random) down to # 0.02 (98% of actions are selected according to values predicted by the model). exploration = LinearSchedule( schedule_timesteps=arglist.num_exploration_steps * arglist.num_episodes * arglist.max_episode_steps, initial_p=1.0, final_p=arglist.final_epsilon) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] n_episodes = 0 n_steps = 0 obs = env.reset() log_path = "./learning_curves/minerl_" + str(date.today()) + "_" + str( time.time()) + ".dat" log_file = open(log_path, "a") for episode in range(arglist.num_episodes): print("Episode: ", str(episode)) done = False episode_steps = 0 while not done: # Take action and update exploration to the newest value action = act(obs[None], update_eps=exploration.value(n_steps))[0] new_obs, rew, done, _ = env.step(action) n_steps += 1 episode_steps += 1 # Break episode if episode_steps > arglist.max_episode_steps: done = True # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs # Store rewards episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0) n_episodes += 1 # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if (n_steps > arglist.learning_starts_at_steps) and (n_steps % 4 == 0): obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( 32) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) if arglist.use_demonstrations: if (n_steps < arglist.learning_starts_at_steps) and ( n_steps % 4 == 0): obses_t, actions, rewards, obses_tp1, dones = demo_buffer.sample( 32) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) if (n_steps > arglist.learning_starts_at_steps) and ( n_steps % 4 == 0): obses_t, actions, rewards, obses_tp1, dones = demo_buffer.sample( 32) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) # Update target network periodically. if n_steps % arglist.target_net_update_freq == 0: update_target() # Log data for analysis if done and len(episode_rewards) % 10 == 0: logger.record_tabular("steps", n_steps) logger.record_tabular("episodes", len(episode_rewards)) logger.record_tabular( "mean episode reward", round(np.mean(episode_rewards[-101:-1]), 1)) logger.record_tabular( "% time spent exploring", int(100 * exploration.value(n_steps))) logger.dump_tabular() #TODO: Save checkpoints if n_steps % arglist.checkpoint_rate == 0: checkpoint_path = "./checkpoints/minerl_" + str( episode) + "_" + str(date.today()) + "_" + str( time.time()) + ".pkl" save_variables(checkpoint_path) print("%s,%s,%s,%s" % (n_steps, episode, round(np.mean(episode_rewards[-101:-1]), 1), int(100 * exploration.value(n_steps))), file=log_file) log_file.close()
def make_obs_ph(name): #return ObservationInput(ob_space, name=name) return ObservationInput(Box(low=0.0, high=screen_dim[0], shape=(screen_dim[0], screen_dim[1], 1)), name=name)
def learn(env, network, seed=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=3000, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=3000, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, **network_kwargs ): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. batch_size: int size of a batch sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name), q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=0.99, double_q=False #grad_norm_clipping=10, # param_noise=param_noise ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(10000), initial_p=1.0, final_p=0.02) # Initialize the parameters and copy them to the target network. U.initialize() update_target() old_state = None formula_LTLf_1 = "!d U(g)" monitoring_RightToLeft = MonitoringSpecification( ltlf_formula=formula_LTLf_1, r=0, c=-0.01, s=10, f=-10 ) formula_LTLf_2 = "F(G(bb)) " # break brick monitoring_BreakBrick = MonitoringSpecification( ltlf_formula=formula_LTLf_2, r=10, c=-0.01, s=10, f=0 ) monitoring_specifications = [monitoring_BreakBrick, monitoring_RightToLeft] def RightToLeftConversion(observation) -> TraceStep: done=False global old_state if arrays_equal(observation[-9:], np.zeros((len(observation[-9:])))): ### Checking if all Bricks are broken # print('goal reached') goal = True # all bricks are broken done = True else: goal = False dead = False if done and not goal: dead = True order = check_ordered(observation[-9:]) if not order: # print('wrong order', state[5:]) dead=True done = True if old_state is not None: # if not the first state if not arrays_equal(old_state[-9:], observation[-9:]): brick_broken = True # check_ordered(state[-9:]) # print(' a brick is broken') else: brick_broken = False else: brick_broken = False dictionary={'g': goal, 'd': dead, 'o': order, 'bb':brick_broken} #print(dictionary) return dictionary multi_monitor = MultiRewardMonitor( monitoring_specifications=monitoring_specifications, obs_to_trace_step=RightToLeftConversion ) episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() reset = True # initialize done = False #monitor.get_reward(None, False) # add first state in trace with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) episodeCounter=0 num_episodes=0 for t in itertools.count(): # Take action and update exploration to the newest value action = act(obs[None], update_eps=exploration.value(t))[0] #print(action) #print(action) new_obs, rew, done, _ = env.step(action) done=False #done=False ## FOR FIRE ONLY #print(new_obs) #new_obs.append() start_time = time.time() rew, is_perm = multi_monitor(new_obs) #print("--- %s seconds ---" % (time.time() - start_time)) old_state=new_obs #print(rew) done=done or is_perm # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew is_solved = t > 100 and np.mean(episode_rewards[-101:-1]) >= 200 if episodeCounter % 100 == 0 or episodeCounter<1: # Show off the result #print("coming here Again and Again") env.render() if done: episodeCounter+=1 num_episodes+=1 obs = env.reset() old_state=None episode_rewards.append(0) multi_monitor.reset() #monitor.get_reward(None, False) else: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if t > 1000: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(64) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) # Update target network periodically. if t % 1000 == 0: update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) if done and len(episode_rewards) % 10 == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", len(episode_rewards)) logger.record_tabular("currentEpisodeReward", episode_rewards[-1]) logger.record_tabular("mean 100 episode reward", round(np.mean(episode_rewards[-101:-1]), 1)) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) act.save_act() #save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward # if model_saved: # if print_freq is not None: # logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) # load_variables(model_file) return act
def cart(target_update_steps=1000, minibatch_size=32): with U.make_session(8): # Create the environment env = gym.make("CartPole-v0") # Create all the functions necessary to train the model act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name), q_func=model, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=5e-4), ) # Create the replay buffer replay_buffer = ReplayBuffer(50000) # Create the schedule for exploration starting from 1 (every action is random) down to # 0.02 (98% of actions are selected according to values predicted by the model). exploration = LinearSchedule(schedule_timesteps=10000, initial_p=1.0, final_p=0.02) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] obs = env.reset() for t in itertools.count(): # Take action and update exploration to the newest value action = act(obs[None], update_eps=exploration.value(t))[0] new_obs, rew, done, _ = env.step(action) # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0) is_solved = t > 100 and np.mean(episode_rewards[-101:-1]) >= 200 if is_solved: pass # Show off the result #env.render() else: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if t > 1000: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( minibatch_size) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) # Update target network periodically. if t % target_update_steps == 0: update_target() if done and len(episode_rewards) % 10 == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", len(episode_rewards)) logger.record_tabular( "mean episode reward", round(np.mean(episode_rewards[-101:-1]), 1)) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
def make_obs_ph(name): return ObservationInput(observation_space, name=name, extra_channels=extra_channels)
def make_bel_ph(name): return ObservationInput(belief_space, name=name)
def make_obs_ph(name): print("ENV.OBSERVATION_SPACE: {}".format(env.observation_space)) return ObservationInput(env.observation_space, name=name)
def main(): # configure logger, disable logging in child MPI processes (with rank > 0) arg_parser = common_arg_parser() args, unknown_args = arg_parser.parse_known_args() extra_args = parse_cmdline_kwargs(unknown_args) if MPI is None or MPI.COMM_WORLD.Get_rank() == 0: rank = 0 logger.configure() else: logger.configure(format_strs=[]) rank = MPI.COMM_WORLD.Get_rank() model, env, debug = train(args, extra_args) # Get the trained model env.close() if args.save_path is not None and rank == 0: save_path = osp.expanduser(args.save_path) model.save(save_path) if args.adv_alg: # If attack is applied, build the function for crafting adversarial observations g = tf.Graph() with g.as_default(): with tf.Session() as sess: q_func = build_q_func(network='conv_only') craft_adv_obs = build_adv( make_obs_tf=lambda name: ObservationInput(env.observation_space, name=name), q_func=q_func, num_actions=env.action_space.n, epsilon=args.epsilon, attack=args.adv_alg ) if args.save_info: # Save all the information in a csv filter name = args.info_name csv_file = open('/Users/harry/Documents/info/' + name, mode='a' ) fieldnames = ['episode', 'diff_type', 'diff', 'epsilon', 'steps', 'attack rate', 'success rate', 'score'] writer = csv.DictWriter(csv_file, fieldnames=fieldnames) writer.writeheader() if args.play: logger.log("Running trained model") env = build_env(args) obs = env.reset() action_meanings = env.unwrapped.get_action_meanings() def initialize_placeholders(nlstm=128,**kwargs): return np.zeros((args.num_env or 1, 2*nlstm)), np.zeros((1)) state, dones = initialize_placeholders(**extra_args) num_episodes = 0 num_moves = 0 num_success_attack = 0 num_attack = 0 step = 0 q_value_dict = {} old_diff = 0 diff_type = args.diff_type print("Type of diff: {}. Threshold to launch attack: {}".format(diff_type, args.diff)) print('-------------------------Episode 0 -------------------------') while True: step = step + 1 # Overall steps. Does not reset to 0 when an episode ends num_moves = num_moves + 1 q_values = debug['q_values']([obs]) q_values = np.squeeze(q_values) minus_diff = np.max(q_values) - np.min(q_values) div_diff = np.max(q_values) / np.min(q_values) sec_ord_diff = minus_diff - old_diff old_diff = minus_diff if args.save_q_value: # Save the q value to a file with open('/Users/harry/Documents/q_value_pong_ep' + str(num_episodes+1) + '_diff' + str(args.diff) + '.csv', 'a') as q_value_file: q_value_writter = csv.writer(q_value_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL) q_value_writter.writerow(q_values) if args.adv_alg: diff = minus_diff if args.diff_type == 'diff' else div_diff \ if args.diff_type == 'div_diff' else sec_ord_diff \ if args.diff_type == 'sec_ord_diff' else minus_diff if diff >= args.diff: num_attack = num_attack + 1 with g.as_default(): with tf.Session() as sess: sess.run(tf.global_variables_initializer()) adv_obs = craft_adv_obs([obs])[0] # Get the adversarial observation adv_obs = np.rint(adv_obs) adv_obs = adv_obs.astype(np.uint8) if args.preview_image: # Show a few adversarial images on the screen if num_attack >= 2 and num_attack <= 10: adv_img = Image.fromarray(np.asarray(adv_obs[:,:,0]), mode='L') adv_img.show() if args.save_image: # Save one episode of adversarial images in a folder if num_episodes == 0: img = Image.fromarray(np.asarray(adv_obs[:,:,0]), mode='L') img.save('/Users/harry/Documents/adv_19_99/adv_image_' + str(num_moves) + '.png') prev_state = np.copy(state) action, _, _, _ = model.step(obs,S=prev_state, M=dones) adv_action, _, state, _ = model.step(adv_obs,S=prev_state, M=dones) if (adv_action != action): # Count as a successful atttack # print('Action before: {}, Action after: {}'.format( # action_meanings[action[0]], action_meanings[adv_action[0]])) num_success_attack = num_success_attack + 1 obs, rew, done, info = env.step(adv_action) else: action, _, state, _ = model.step(obs,S=state, M=dones) obs, rew, done, info = env.step(action) if args.save_image: img = Image.fromarray(np.asarray(obs[:,:,0]), mode='L') img.save('/Users/harry/Documents/adv_images_ep' + str(num_episodes+1) + '/' + str(num_moves) + '.png') else: if args.save_image: # Save one episode of normal images in a folder if num_episodes == 0: img = Image.fromarray(np.asarray(obs[:,:,0]), mode='L') img.save('/Users/harry/Documents/normal_obs' + str(num_moves) + '.png') action, _, state, _ = model.step(obs,S=state, M=dones) obs, _, done, info = env.step(action) env.render() done = done.any() if isinstance(done, np.ndarray) else done if done: npc_score = info['episode']['r'] score = 21 if npc_score < 0 else 21 - npc_score obs = env.reset() print('Episode {} takes {} time steps'.format(num_episodes, num_moves)) print('NPC Score: {}'.format(npc_score)) if args.adv_alg: attack_rate = float(num_attack) / num_moves success_rate = float(num_success_attack) / num_attack print('Percentage of attack: {}'.format(100 * attack_rate)) print('Percentage of successful attacks: {}'.format(100 * success_rate)) info_dict = {'episode': num_episodes+1,'diff_type': args.diff_type, 'diff': args.diff, 'epsilon': args.epsilon, 'steps': num_moves, 'attack rate': attack_rate, 'success rate': success_rate, 'score': score} writer.writerow(info_dict) num_moves = 0 num_transfer = 0 num_episodes = num_episodes + 1 num_attack = 0 num_success_attack = 0 print(f'-------------------------Episode {num_episodes}-------------------------') env.close()
def learn(env, network, seed=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=3000, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=3000, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, **network_kwargs ): sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name), q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), # gamma=gamma, # grad_norm_clipping=10, # param_noise=param_noise ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(100000), initial_p=1.0, final_p=0.02) # Initialize the paramete print(type(act))rs and copy them to the target network. U.initialize() update_target() old_state = None formula_LTLf_1 = "!F(die)" monitoring_RightToLeft = MonitoringSpecification( ltlf_formula=formula_LTLf_1, r=1, c=-10, s=1, f=-10 ) monitoring_specifications = [monitoring_RightToLeft] stepCounter = 0 done = False def RightToLeftConversion(observation) -> TraceStep: print(stepCounter) if(done and not(stepCounter>=199)): die=True else: die=False dictionary={'die': die} print(dictionary) return dictionary multi_monitor = MultiRewardMonitor( monitoring_specifications=monitoring_specifications, obs_to_trace_step=RightToLeftConversion ) episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() reset = True with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) episodeCounter=0 num_episodes=0 for t in itertools.count(): # Take action and update exploration to the newest value action = act(obs[None], update_eps=exploration.value(t))[0] #print(action) new_obs, rew, done, _ = env.step(action) stepCounter+=1 rew, is_perm = multi_monitor(new_obs) old_state=new_obs # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew is_solved = t > 100 and np.mean(episode_rewards[-101:-1]) >= 200 if episodeCounter % 100 == 0 or episodeCounter<1: # Show off the result #print("coming here Again and Again") env.render() if done: episodeCounter+=1 num_episodes+=1 obs = env.reset() episode_rewards.append(0) multi_monitor.reset() stepCounter=0 else: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if t > 1000: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(32) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) # Update target network periodically. if t % 1000 == 0: update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) if done and len(episode_rewards) % 10 == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", len(episode_rewards)) logger.record_tabular("mean 100 episode reward", round(np.mean(episode_rewards[-101:-1]), 1)) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 500 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) act.save_act() #save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward # if model_saved: # if print_freq is not None: # logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) # load_variables(model_file) return act
def make_obs_ph(name): if aug_type == 'constraint_state': return ConstraintStateAugmentedInput(observation_space, constraints, name=name) return ObservationInput(observation_space, name=name)
def learn(self): with U.make_session(8): # Create the environment env = gym.make(self._args.env) # Create all the functions necessary to train the model act, train, update_target, debug = deepq.build_train( make_obs_ph=lambda name: ObservationInput( env.observation_space, name=name), q_func=self.model, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer( learning_rate=self._args.learning_rate), ) # Create the replay buffer replay_buffer = ReplayBuffer(self._args.replay_buffer_size) # Create the schedule for exploration starting from 1 till min_exploration_rate. exploration = LinearSchedule( schedule_timesteps=self._args.exploration_duration, initial_p=1.0, final_p=self._args.min_exploration_rate) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] obs = env.reset() for t in itertools.count(): # Take action and update exploration to the newest value action = act(obs[None], update_eps=exploration.value(t))[0] new_obs, rew, done, _ = env.step(action) # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0) mean_episode_reward = np.mean(episode_rewards[-101:-1]) # Show learned agent: if mean_episode_reward >= self._render_reward_threshold: env.render() # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if t > 1000: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( 32) train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards)) # Update target network periodically. if t % 1000 == 0: update_target() if done and len(episode_rewards) % 10 == 0: self._reward_buffer_mutex.acquire() self._reward_buffer.append(mean_episode_reward) logger.record_tabular("steps", t) logger.record_tabular("episodes", len(episode_rewards)) logger.record_tabular("mean episode reward", round(mean_episode_reward, 1)) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() self._reward_buffer_changed = True self._reward_buffer_mutex.release()