class DQN(OffPolicyRLModel): """ The DQN model class. DQN paper: https://arxiv.org/pdf/1312.5602.pdf :param policy: (DQNPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param 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. :param checkpoint_path: (str) replacement path used if you need to log to somewhere else than a temporary directory. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float) alpha parameter for prioritized replay buffer :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance """ def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, target_network_update_freq=500, 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, verbose=0, tensorboard_log=None, _init_setup_model=True): # TODO: replay_buffer refactoring super(DQN, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, policy_base=DQNPolicy, requires_vec_env=False) self.checkpoint_path = checkpoint_path self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.checkpoint_freq = checkpoint_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.tensorboard_log = tensorboard_log self.graph = None self.sess = None self._train_step = None self.step_model = None self.update_target = None self.act = None self.proba_step = None self.replay_buffer = None self.beta_schedule = None self.exploration = None self.params = None self.summary = None self.episode_reward = None if _init_setup_model: self.setup_model() def setup_model(self): with SetVerbosity(self.verbose): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf_util.make_session(graph=self.graph) optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = deepq.build_train( q_func=self.policy, ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess) self.proba_step = self.step_model.proba_step self.params = find_trainable_variables("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN"): with SetVerbosity(self.verbose), TensorboardWriter( self.graph, self.tensorboard_log, tb_log_name) as writer: self._setup_learn(seed) # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=1.0, final_p=self.exploration_final_eps) episode_rewards = [0.0] obs = self.env.reset() reset = True self.episode_reward = np.zeros((1, )) for step in range(total_timesteps): if callback is not None: callback(locals(), globals()) # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(step) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(step) + self.exploration.value(step) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, _ = self.env.step(env_action) # Store transition in the replay buffer. self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_rew, ep_done, writer, step) episode_rewards[-1] += rew if done: if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True if step > self.learning_starts and step % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(step)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + step) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata(run_metadata, 'step%d' % step) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, step) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) if step > self.learning_starts and step % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", step) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(step))) logger.dump_tabular() return self def predict(self, observation, state=None, mask=None, deterministic=True): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) with self.sess.as_default(): actions, _, _ = self.step_model.step(observation, deterministic=deterministic) if not vectorized_env: actions = actions[0] return actions, None def action_probability(self, observation, state=None, mask=None): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) actions_proba = self.proba_step(observation, state, mask) if not vectorized_env: if state is not None: raise ValueError( "Error: The environment must be vectorized when using recurrent policies." ) actions_proba = actions_proba[0] return actions_proba def save(self, save_path): # params data = { "checkpoint_path": self.checkpoint_path, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "checkpoint_freq": self.checkpoint_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "_vectorize_action": self._vectorize_action } params = self.sess.run(self.params) self._save_to_file(save_path, data=data, params=params) @classmethod def load(cls, load_path, env=None, **kwargs): data, params = cls._load_from_file(load_path) model = cls(policy=data["policy"], env=env, _init_setup_model=False) model.__dict__.update(data) model.__dict__.update(kwargs) model.set_env(env) model.setup_model() restores = [] for param, loaded_p in zip(model.params, params): restores.append(param.assign(loaded_p)) model.sess.run(restores) return model
class DQN(OffPolicyRLModel): """ The DQN model class. DQN paper: https://arxiv.org/abs/1312.5602 Dueling DQN: https://arxiv.org/abs/1511.06581 Double-Q Learning: https://arxiv.org/abs/1509.06461 Prioritized Experience Replay: https://arxiv.org/abs/1511.05952 :param policy: (DQNPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param exploration_initial_eps: (float) initial value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param double_q: (bool) Whether to enable Double-Q learning or not. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float)alpha parameter for prioritized replay buffer. It determines how much prioritization is used, with alpha=0 corresponding to the uniform case. :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance :param full_tensorboard_log: (bool) enable additional logging when using tensorboard WARNING: this logging can take a lot of space quickly :param seed: (int) Seed for the pseudo-random generators (python, numpy, tensorflow). If None (default), use random seed. Note that if you want completely deterministic results, you must set `n_cpu_tf_sess` to 1. :param n_cpu_tf_sess: (int) The number of threads for TensorFlow operations If None, the number of cpu of the current machine will be used. """ def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, exploration_initial_eps=1.0, train_freq=1, batch_size=32, double_q=True, learning_starts=1000, target_network_update_freq=500, 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, n_cpu_tf_sess=None, verbose=0, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None, full_tensorboard_log=False, seed=None, use_rmsprop=False, rmsprop_alpha=0.95, rmsprop_epsilon=0.01, exploration_offset=0): # TODO: replay_buffer refactoring super(DQN, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, policy_base=DQNPolicy, requires_vec_env=False, policy_kwargs=policy_kwargs, seed=seed, n_cpu_tf_sess=n_cpu_tf_sess) self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_initial_eps = exploration_initial_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.tensorboard_log = tensorboard_log self.full_tensorboard_log = full_tensorboard_log self.double_q = double_q self.use_rmsprop = use_rmsprop self.rmsprop_alpha = rmsprop_alpha self.rmsprop_epsilon = rmsprop_epsilon self.exploration_offset = exploration_offset self.graph = None self.sess = None self._train_step = None self.step_model = None self.update_target = None self.act = None self.proba_step = None self.replay_buffer = None self.beta_schedule = None self.exploration = None self.params = None self.summary = None self.episode_reward = None if _init_setup_model: self.setup_model() def _get_pretrain_placeholders(self): policy = self.step_model return policy.obs_ph, tf.placeholder(tf.int32, [None]), policy.q_values def setup_model(self): with SetVerbosity(self.verbose): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.set_random_seed(self.seed) self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph) if self.use_rmsprop: optimizer = tf.train.RMSPropOptimizer( learning_rate=self.learning_rate, decay=self.rmsprop_alpha, epsilon=self.rmsprop_epsilon, centered=True ) else: optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = build_train( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log=self.full_tensorboard_log, double_q=self.double_q ) self.proba_step = self.step_model.proba_step self.params = tf_util.get_trainable_vars("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer(self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule(schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=self.exploration_initial_eps, final_p=self.exploration_final_eps) episode_rewards = [0.0] episode_successes = [] obs = self.env.reset() reset = True self.episode_reward = np.zeros((1,)) for _ in range(total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps - self.exploration_offset) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps - self.exploration_offset) + self.exploration.value(self.num_timesteps - self.exploration_offset) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, info = self.env.step(env_action) # Store transition in the replay buffer. self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew if done: maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. # pytype:disable=bad-unpacking if self.prioritized_replay: assert self.beta_schedule is not None, \ "BUG: should be LinearSchedule when self.prioritized_replay True" experience = self.replay_buffer.sample(self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(self.batch_size) weights, batch_idxes = np.ones_like(rewards), None # pytype:enable=bad-unpacking if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata(run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs(td_errors) + self.prioritized_replay_eps assert isinstance(self.replay_buffer, PrioritizedReplayBuffer) self.replay_buffer.update_priorities(batch_idxes, new_priorities) if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round(float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len(episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * self.exploration.value(self.num_timesteps - self.exploration_offset))) logger.dump_tabular() self.num_timesteps += 1 return episode_rewards def predict(self, observation, state=None, mask=None, deterministic=True): observation = np.array(observation) vectorized_env = self._is_vectorized_observation(observation, self.observation_space) observation = observation.reshape((-1,) + self.observation_space.shape) with self.sess.as_default(): actions, _, _ = self.step_model.step(observation, deterministic=deterministic) if not vectorized_env: actions = actions[0] return actions, None def action_probability(self, observation, state=None, mask=None, actions=None, logp=False): observation = np.array(observation) vectorized_env = self._is_vectorized_observation(observation, self.observation_space) observation = observation.reshape((-1,) + self.observation_space.shape) actions_proba = self.proba_step(observation, state, mask) if actions is not None: # comparing the action distribution, to given actions actions = np.array([actions]) assert isinstance(self.action_space, gym.spaces.Discrete) actions = actions.reshape((-1,)) assert observation.shape[0] == actions.shape[0], "Error: batch sizes differ for actions and observations." actions_proba = actions_proba[np.arange(actions.shape[0]), actions] # normalize action proba shape actions_proba = actions_proba.reshape((-1, 1)) if logp: actions_proba = np.log(actions_proba) if not vectorized_env: if state is not None: raise ValueError("Error: The environment must be vectorized when using recurrent policies.") actions_proba = actions_proba[0] return actions_proba def get_parameter_list(self): return self.params def save(self, save_path, cloudpickle=False): # params data = { "double_q": self.double_q, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "n_cpu_tf_sess": self.n_cpu_tf_sess, "seed": self.seed, "_vectorize_action": self._vectorize_action, "policy_kwargs": self.policy_kwargs } params_to_save = self.get_parameters() self._save_to_file(save_path, data=data, params=params_to_save, cloudpickle=cloudpickle)
class DqnAtml(DQN): def setup_model(self): with SetVerbosity(self.verbose): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf_util.make_session(graph=self.graph) optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = build_train_atml( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log=self.full_tensorboard_log) self.proba_step = self.step_model.proba_step self.params = tf_util.get_trainable_vars("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def get_actions_vec(self, actions_prims, actions_inputs, actions_mf): with self.sess.as_default(): self.embedd_matrix = self.step_model.embedding.get_weights() invalid_action = np.zeros(self.embedd_matrix[0].shape[1]) - 1 self.embedd_matrix = np.vstack([self.embedd_matrix[0], invalid_action]) embedded_steps = self.embedd_matrix[actions_prims.astype(int)] actions_inputs = actions_inputs.reshape(len(actions_prims), -1) actions_mf = actions_mf.reshape(len(actions_prims), -1) concat_actions = np.concatenate( (embedded_steps, actions_inputs, actions_mf), axis=1) flatten_act = concat_actions.reshape(-1) return flatten_act def process_state_vec(self, obs, state_info): # transform actions representation with embeddings with self.sess.as_default(): self.embedd_matrix = self.step_model.embedding.get_weights() ind1 = state_info['grid_prims_size'] ind2 = ind1 + state_info['relations_size'] ind3 = ind2 + state_info['ff_state_size'] ind4 = ind3 + state_info['action_prims'] ind5 = ind4 + state_info['action_inputs'] ind6 = ind5 + state_info['action_mf'] cells_num = state_info['cells_num'] actions_prims = obs[ind3:ind4] actions_inputs = obs[ind4:ind5] actions_mf = obs[ind5:] flatten_act = self.get_actions_vec(actions_prims, actions_inputs, actions_mf) final_obs = np.concatenate((obs[:ind3], flatten_act)) return final_obs def hierarchical_step(self, obs, ds_rewards, cnt, kwargs, update_eps): register = False while not register: with self.sess.as_default(): action = self.predict(np.array(obs)[None])[0][0] env_action = action reset = False new_obs, rew, done, info = self.env.step(env_action) level = info.get('hier_level') register = info.get('register') self.actions_container.append(env_action) self.actions_weights.append(level) if rew < 0 or register: with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] if rew < 0 and not register: self.actions_container = self.actions_container[:-1] self.actions_weights = self.actions_weights[:-1] rep_action = np.zeros(self.action_space.n) rep_action[action] = 1.0 if register: if rew > 0: ds_rewards.append([cnt, rew]) cnt += 1 self.actions_container = np.array(self.actions_container) self.actions_weights = np.array( self.actions_weights) / level b = np.zeros( (len(self.actions_container), self.action_space.n)) b[np.arange(len(self.actions_container)), self.actions_container.astype(int)] = 1 act_replay = np.sum((self.actions_weights * b.T).T, axis=0) rep_action = act_replay / np.sum(act_replay) self.actions_container = [] self.actions_weights = [] self.replay_buffer.add(obs, rep_action, rew, new_obs, float(done)) break obs = new_obs obs = new_obs return obs, new_obs, rew, action, done, reset def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, initial_p=1.0): self.actions_weights = [] self.actions_container = [] new_tb_log = self._init_num_timesteps(reset_num_timesteps) cnt = 0 ds_rewards = [[0, 0]] with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=initial_p, final_p=self.exploration_final_eps) episode_rewards = [0.0] obs = self.env.reset() reset = True self.episode_reward = np.zeros((1, )) for _ in range(total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True ''' Hierarchical Step (Start) ''' obs, new_obs, rew, action, done, reset = self.hierarchical_step( obs, ds_rewards, cnt, kwargs, update_eps) ''' Hierarchical Step (End) ''' if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew if done: if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True if self.num_timesteps > self.learning_starts and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: try: new_priorities = np.array([ abs(x) for x in td_errors.tolist() ]) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) except AssertionError: print(td_errors) if self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.dump_tabular() self.num_timesteps += 1 return self, ds_rewards
class MADQN(OffPolicyRLModel): """ The DQN model class. DQN paper: https://arxiv.org/abs/1312.5602 Dueling DQN: https://arxiv.org/abs/1511.06581 Double-Q Learning: https://arxiv.org/abs/1509.06461 Prioritized Experience Replay: https://arxiv.org/abs/1511.05952 :param policy: (DQNPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param exploration_initial_eps: (float) initial value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param double_q: (bool) Whether to enable Double-Q learning or not. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float)alpha parameter for prioritized replay buffer. It determines how much prioritization is used, with alpha=0 corresponding to the uniform case. :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance :param full_tensorboard_log: (bool) enable additional logging when using tensorboard WARNING: this logging can take a lot of space quickly :param seed: (int) Seed for the pseudo-random generators (python, numpy, tensorflow). If None (default), use random seed. Note that if you want completely deterministic results, you must set `n_cpu_tf_sess` to 1. :param n_cpu_tf_sess: (int) The number of threads for TensorFlow operations If None, the number of cpu of the current machine will be used. """ def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, exploration_initial_eps=1.0, train_freq=1, batch_size=32, double_q=True, learning_starts=1000, target_network_update_freq=500, 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, n_cpu_tf_sess=None, verbose=0, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None, full_tensorboard_log=False, seed=None, num_agents=1): # MA-MOD # TODO: replay_buffer refactoring super(MADQN, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, policy_base=DQNPolicy, requires_vec_env=False, policy_kwargs=policy_kwargs, seed=seed, n_cpu_tf_sess=n_cpu_tf_sess) # print("POLICY TYPE", policy) if self.observation_space: obs_sp_low = self.observation_space.low[0, :] obs_sp_high = self.observation_space.high[0, :] self.observation_space = gym.spaces.Box(low=obs_sp_low, high=obs_sp_high) self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_initial_eps = exploration_initial_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.tensorboard_log = tensorboard_log self.full_tensorboard_log = full_tensorboard_log self.double_q = double_q self.num_agents = num_agents self.graph = None self.sess = None self._train_step = [] # MA-MOD self.step_model = [] # MA-MOD self.update_target = [] # MA-MOD self.act = [] # MA-MOD self.proba_step = [] # MA-MOD self.replay_buffer = None # TODO: Possibly try seperate replay buffer. If everything symmetric, OK for one. # If you have the same Value function, its fine. If you have seperate functions, if you have one replay buffer, they learn from the same data. self.beta_schedule = None self.exploration = None self.params = None self.summary = None if _init_setup_model: self.setup_model() def _get_pretrain_placeholders(self): assert False, "MAKE SURE THIS FUNCTION ISNT CALLED" policy = self.step_model return policy.obs_ph, tf.placeholder(tf.int32, [None]), policy.q_values def setup_model(self): with SetVerbosity(self.verbose): for i in range(self.num_agents): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy # print(test_policy.type) assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.set_random_seed(self.seed) self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph) self.params = [] print("AC SPC", self.action_space) for i in range(self.num_agents): with tf.variable_scope("agent" + str(i)): optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) act, _train_step, update_target, step_model = build_train( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log= False, #self.full_tensorboard_log, double_q=self.double_q) self.act.append(act) self._train_step.append(_train_step) self.step_model.append(step_model) self.proba_step.append(step_model.proba_step) self.update_target.append(update_target) self.params.extend( tf_util.get_trainable_vars("agent" + str(i) + "/deepq")) print(self.params) # Initialize the parameters and copy them to the target network. tf_util.initialize( self.sess ) # TODO: copy this file, make two versions of the algorithm. for i in range(self.num_agents): self.update_target[i]( sess=self.sess ) # TODO: Not sure, seems like the best thing to do is try using each agents own target first. # self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) # callback = self._init_callback(callback) # with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ # as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=self.exploration_initial_eps, final_p=self.exploration_final_eps) episode_rewards = [[0.0] * self.num_agents] #MA-MOD episode_successes = [] #callback.on_training_start(locals(), globals()) #callback.on_rollout_start() reset = True obs = self.env.reset() for _ in range(total_timesteps): # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): env_action = [] # MA-MOD for i in range(self.num_agents ): # MA-MOD. This is fine for one policy. action = self.act[i]( np.array(obs[i])[None], update_eps=update_eps, **kwargs )[0] # TODO: Is this the correct way to get the correct agent obs? env_action.append(action) reset = False new_obs, rew, done, info = self.env.step( env_action ) # NOUPDATE - env.step should take a vector of actions ''' Obs: x_me, x_opp --- agent 1. In env: x_1, x_2 Obs: x_me, x_opp -- agent 2. In env: x_2, x_1 Env: (n_agents, state_dim) ''' self.num_timesteps += 1 # Stop training if return value is False # if callback.on_step() is False: # break # Store transition in the replay buffer. # Loop for replay buffer -- either separate or joined. obs[agent_index], action[agent_index], reward[agent_index] # Joey: Does this look right to you? # print(obs, action, rew, new_obs, done) #print("obs",obs[0]) #print(action) #print("ac", action[0]) #print("rew", rew[0]) #print("done", done[0]) for num_agent in range(self.num_agents): self.replay_buffer.add(obs[num_agent], env_action[num_agent], rew[num_agent], new_obs[num_agent], float(done[num_agent])) obs = new_obs # if writer is not None: # ep_rew = np.array([rew]).reshape((1, -1)) # ep_done = np.array([done]).reshape((1, -1)) # tf_util.total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer, # self.num_timesteps) # TODO: current episode_rewards is a list, make it a list of lists where each list is the reward for each agent in all timesteps # append the newest reward to the end of each list for each agent for num_agent in range(self.num_agents): #MA-MOD episode_rewards[-1][num_agent] += rew[num_agent] if done.any(): maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append([0.0] * self.num_agents) reset = True # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # callback.on_rollout_end() for i in range(self.num_agents): # MA-MOD # Minimize the error in Bellman's equation on a batch sampled from replay buffer. # pytype:disable=bad-unpacking if self.prioritized_replay: assert self.beta_schedule is not None, \ "BUG: should be LinearSchedule when self.prioritized_replay True" experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None # pytype:enable=bad-unpacking # if writer is not None: # # run loss backprop with summary, but once every 100 steps save the metadata # # (memory, compute time, ...) # if (1 + self.num_timesteps) % 100 == 0: # run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) # run_metadata = tf.RunMetadata() # summary, td_errors = self._train_step[i](obses_t, actions, rewards, obses_tp1, obses_tp1, # dones, weights, sess=self.sess, options=run_options, # run_metadata=run_metadata) # writer.add_run_metadata(run_metadata, 'step%d_agent%d' % (self.num_timesteps, i)) # else: # summary, td_errors = self._train_step[i](obses_t, actions, rewards, obses_tp1, obses_tp1, # dones, weights, sess=self.sess) # writer.add_summary(summary, self.num_timesteps) # else: td_errors = self._train_step[i](obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: # NOUPDATE - not inside main agent for loop new_priorities = np.abs( td_errors) + self.prioritized_replay_eps # NOUPDATE assert isinstance(self.replay_buffer, PrioritizedReplayBuffer) self.replay_buffer.update_priorities( batch_idxes, new_priorities) # callback.on_rollout_start() if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. for i in range(self.num_agents): self.update_target[i](sess=self.sess) # MA-MOD if len(episode_rewards[-101:-1]) == 0: # MA-MOD mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) #MA-MOD # below is what's logged in terminal. num_episodes = len(episode_rewards) #MA-MOD if self.verbose >= 1 and done.any( ) and log_interval is not None and len( episode_rewards) % log_interval == 0: #MA-MOD logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.dump_tabular() return self def predict( self, observation, agent_idx, state=None, mask=None, deterministic=True): # MA-MOD - added `agent_idx` as a parameter observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) with self.sess.as_default(): actions, _, _ = self.step_model[agent_idx].step( observation, deterministic=deterministic) if not vectorized_env: actions = actions[0] return actions, None # No one ever calls this, so we don't need it? def action_probability(self, observation, state=None, mask=None, actions=None, logp=False): print("Should not be called") return None ''' observation = np.array(observation) vectorized_env = self._is_vectorized_observation(observation, self.observation_space) observation = observation.reshape((-1,) + self.observation_space.shape) actions_proba = self.proba_step(observation, state, mask) if actions is not None: # comparing the action distribution, to given actions actions = np.array([actions]) assert isinstance(self.action_space, gym.spaces.Discrete) actions = actions.reshape((-1,)) assert observation.shape[0] == actions.shape[0], "Error: batch sizes differ for actions and observations." actions_proba = actions_proba[np.arange(actions.shape[0]), actions] # normalize action proba shape actions_proba = actions_proba.reshape((-1, 1)) if logp: actions_proba = np.log(actions_proba) if not vectorized_env: if state is not None: raise ValueError("Error: The environment must be vectorized when using recurrent policies.") actions_proba = actions_proba[0] return actions_proba ''' def get_parameter_list(self): print(self.params) return self.params def save(self, save_path, cloudpickle=False): # params data = { "double_q": self.double_q, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "n_cpu_tf_sess": self.n_cpu_tf_sess, "seed": self.seed, "_vectorize_action": self._vectorize_action, "policy_kwargs": self.policy_kwargs, "num_agents": self.num_agents } params_to_save = self.get_parameters() # print(params_to_save) self._save_to_file(save_path, data=data, params=params_to_save, cloudpickle=cloudpickle)
class DQNUniqueActions(DQN): """Subclass with a modified "learn" method which only allows actions to be taken once per episode. """ def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=self.exploration_initial_eps, final_p=self.exploration_final_eps) episode_rewards = [0.0] episode_successes = [] obs = self.env.reset() reset = True ############################################################ # MODIFICATION: # Track list of actions taken each episode. This is # intentionally not a set so that we can use np.isin. action_list = list() ############################################################ for _ in range(total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): #################################################### # MODIFICATION: # Rename variable from original, since it's now # going to come back as an array due to the # modified build_act function being used to # construct everything. action_arr = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] #################################################### # ORIGINAL: # action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] ######################################################## # MODIFICATION: # Get the best action that has not yet been taken this # episode. action = \ action_arr[np.argmin(np.isin(action_arr, action_list))] # Add this action to the list. action_list.append(action) ######################################################## env_action = action reset = False new_obs, rew, done, info = self.env.step(env_action) # Store transition in the replay buffer. self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew if done: #################################################### # MODIFICATION: # Clear the list. action_list.clear() #################################################### maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. # pytype:disable=bad-unpacking if self.prioritized_replay: assert self.beta_schedule is not None, \ "BUG: should be LinearSchedule when self.prioritized_replay True" experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None # pytype:enable=bad-unpacking if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps assert isinstance(self.replay_buffer, PrioritizedReplayBuffer) self.replay_buffer.update_priorities( batch_idxes, new_priorities) if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.dump_tabular() self.num_timesteps += 1 return self
class DqnAtml(DQN): def get_actions_vec(self, actions_prims, actions_inputs, actions_mf): with self.sess.as_default(): self.embedd_matrix = self.step_model.embedding.get_weights() invalid_action = np.zeros(self.embedd_matrix[0].shape[1]) - 1 self.embedd_matrix = np.vstack([self.embedd_matrix[0], invalid_action]) embedded_steps = self.embedd_matrix[actions_prims.astype(int)] actions_inputs = actions_inputs.reshape(len(actions_prims), -1) actions_mf = actions_mf.reshape(len(actions_prims), -1) concat_actions = np.concatenate( (embedded_steps, actions_inputs, actions_mf), axis=1) flatten_act = concat_actions.reshape(-1) return flatten_act def process_state_vec(self, obs, state_info): # transform actions representation with embeddings with self.sess.as_default(): self.embedd_matrix = self.step_model.embedding.get_weights() ind1 = state_info['grid_prims_size'] ind2 = ind1 + state_info['relations_size'] ind3 = ind2 + state_info['ff_state_size'] ind4 = ind3 + state_info['action_prims'] ind5 = ind4 + state_info['action_inputs'] ind6 = ind5 + state_info['action_mf'] cells_num = state_info['cells_num'] actions_prims = obs[ind3:ind4] actions_inputs = obs[ind4:ind5] actions_mf = obs[ind5:] flatten_act = self.get_actions_vec(actions_prims, actions_inputs, actions_mf) final_obs = np.concatenate((obs[:ind3], flatten_act)) return final_obs def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, initial_p=1.0): new_tb_log = self._init_num_timesteps(reset_num_timesteps) cnt = 0 ds_rewards = [[0, 0]] with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=initial_p, final_p=self.exploration_final_eps) episode_rewards = [0.0] obs = self.env.reset() reset = True self.episode_reward = np.zeros((1, )) for _ in range(total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True ''' Hierarchical Step (Start) ''' register = False while not register: with self.sess.as_default(): action = self.predict(np.array(obs)[None])[0][0] env_action = action reset = False new_obs, rew, done, info = self.env.step(env_action) # self.env.render() register = info.get('register') if register: if rew > 0: ds_rewards.append([cnt, rew]) cnt += 1 with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] break obs = new_obs self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs ''' Hierarchical Step (End) ''' if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew if done: if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True if self.num_timesteps > self.learning_starts and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) if self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.dump_tabular() self.num_timesteps += 1 return self, ds_rewards
class MyDQN(DQN): def __init__(self, *args, **kwargs): super(MyDQN, self).__init__(*args, **kwargs) def setup_model(self): with SetVerbosity(self.verbose): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf_util.make_session(graph=self.graph) optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = my_build_train( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log=self.full_tensorboard_log, double_q=self.double_q) self.proba_step = self.step_model.proba_step self.params = tf_util.get_trainable_vars("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn(seed) # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=1.0, final_p=self.exploration_final_eps) episode_rewards = [0.0] episode_successes = [] Globals.env = self.env obs = self.env.reset() reset = True self.episode_reward = np.zeros((1, )) timesteps_last_log = 0 avr_ep_len_per_log = None sleep = 0.045 for _ in range(total_timesteps): if Globals.loading: Globals.loading = False while Globals.pause_game: pass if Globals.exit_learning: break if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, info = self.env.step(env_action) # Store transition in the replay buffer. self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew if done: maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample: sleep = 0.035 time.sleep(sleep) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) if len(episode_rewards) % log_interval == 0: avr_ep_len_per_log = (self.num_timesteps - timesteps_last_log) / log_interval timesteps_last_log = self.num_timesteps num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.record_tabular("avr length of last logged ep", avr_ep_len_per_log) logger.dump_tabular() self.num_timesteps += 1 Globals.steps -= 1 return self def evaluate(self, n_episodes=2): logging.basicConfig(level=logging.INFO) id = 'BreakoutNoFrameskip-v4' num_env = 1 n_stack = 4 left_lives = 5 seed = 0 episodes = 0 score = 0 frames = 0 frames_per_episode = list() scores = [list() for i in range(n_episodes)] env = make_atari_env(id, num_env=num_env, seed=seed) env = VecFrameStack(env, n_stack=n_stack) obs = env.reset() while (n_episodes - episodes) > 0: frames += 1 action, _states = self.predict(obs) obs, rewards, dones, info = env.step(action) env.render() score += rewards[0] if dones: logging.debug('You died') logging.debug(f'Score = {score}') scores[episodes].append(score) score = 0 left_lives -= 1 if not left_lives: logging.debug('Episode ended') logging.info(f'Scores per life: {scores[episodes]}') frames_per_episode.append(frames) frames = 0 episodes += 1 left_lives = 5 s = list(map(sum, scores)) avg_s = int(sum(s) / len(s)) avg_f = int(sum(frames_per_episode) / len(frames_per_episode)) logging.info(f'Played {n_episodes} episodes') logging.info(f'Scores per episode : {s}') logging.info(f'Average score per episode : {avg_s}') logging.info(f'Average number of frames per episode : {avg_f}') return avg_f, avg_s
class DeepQ(BaseRLModel): """ The DQN model class. DQN paper: https://arxiv.org/pdf/1312.5602.pdf :param policy: (function (TensorFlow Tensor, int, str, bool): TensorFlow Tensor) the policy that takes the following inputs: - observation_in: (object) the output of observation placeholder - num_actions: (int) number of actions - scope: (str) - reuse: (bool) should be passed to outer variable scope and returns a tensor of shape (batch_size, num_actions) with values of every action. :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param 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. :param checkpoint_path: (str) replacement path used if you need to log to somewhere else than a temporary directory. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float) alpha parameter for prioritized replay buffer :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance """ def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, target_network_update_freq=500, 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, verbose=0, _init_setup_model=True): super(DeepQ, self).__init__(policy=policy, env=env, requires_vec_env=False, verbose=verbose) assert not isinstance(policy, ActorCriticPolicy), \ "Error: DeepQ does not support the actor critic policies, please use the " \ "'stable_baselines.deepq.models.mlp' and 'stable_baselines.deepq.models.cnn_to_mlp' " \ "functions to create your policies." self.checkpoint_path = checkpoint_path self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.checkpoint_freq = checkpoint_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.graph = None self.sess = None self._train_step = None self.update_target = None self.act = None self.replay_buffer = None self.beta_schedule = None self.exploration = None self.params = None if _init_setup_model: self.setup_model() def setup_model(self): with SetVerbosity(self.verbose): assert isinstance(self.action_space, gym.spaces.Discrete), \ "Error: DeepQ cannot output a {} action space, only spaces.Discrete is supported."\ .format(self.action_space) self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf_util.make_session(graph=self.graph) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = self.observation_space def make_obs_ph(name): """ makes the observation placeholder :param name: (str) the placeholder name :return: (TensorFlow Tensor) the placeholder """ return ObservationInput(observation_space, name=name) self.act, self._train_step, self.update_target, _ = deepq.build_train( make_obs_ph=make_obs_ph, q_func=self.policy, num_actions=self.action_space.n, optimizer=tf.train.AdamOptimizer( learning_rate=self.learning_rate), gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise) self.params = find_trainable_variables("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) def learn(self, total_timesteps, callback=None, seed=None, log_interval=100): with SetVerbosity(self.verbose): self._setup_learn(seed) # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=1.0, final_p=self.exploration_final_eps) episode_rewards = [0.0] obs = self.env.reset() reset = True for step in range(total_timesteps): if callback is not None: callback(locals(), globals()) # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(step) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(step) + self.exploration.value(step) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, _ = self.env.step(env_action) # Store transition in the replay buffer. self.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True if step > self.learning_starts and step % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(step)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) if step > self.learning_starts and step % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", step) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(step))) logger.dump_tabular() return self def predict(self, observation, state=None, mask=None): observation = np.array(observation).reshape( self.observation_space.shape) with self.sess.as_default(): action = self.act(observation[None])[0] if self._vectorize_action: return [action], [None] else: return action, None def action_probability(self, observation, state=None, mask=None): observation = np.array(observation).reshape( (-1, ) + self.observation_space.shape) # Get the tensor just before the softmax function in the TensorFlow graph, # then execute the graph from the input observation to this tensor. tensor = self.graph.get_tensor_by_name( 'deepq/q_func/fully_connected_2/BiasAdd:0') if self._vectorize_action: return self._softmax( self.sess.run(tensor, feed_dict={'deepq/observation:0': observation})) else: return self._softmax( self.sess.run(tensor, feed_dict={'deepq/observation:0': observation}))[0] def save(self, save_path): # params data = { "checkpoint_path": self.checkpoint_path, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "checkpoint_freq": self.checkpoint_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "_vectorize_action": self._vectorize_action } params = self.sess.run(self.params) self._save_to_file(save_path, data=data, params=params) @classmethod def load(cls, load_path, env=None, **kwargs): data, params = cls._load_from_file(load_path) model = cls(policy=data["policy"], env=env, _init_setup_model=False) model.__dict__.update(data) model.__dict__.update(kwargs) model.set_env(env) model.setup_model() restores = [] for param, loaded_p in zip(model.params, params): restores.append(param.assign(loaded_p)) model.sess.run(restores) return model
class DQN(OffPolicyRLModel): """ The DQN model class. DQN paper: https://arxiv.org/abs/1312.5602 Dueling DQN: https://arxiv.org/abs/1511.06581 Double-Q Learning: https://arxiv.org/abs/1509.06461 Prioritized Experience Replay: https://arxiv.org/abs/1511.05952 :param policy: (DQNPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param exploration_initial_eps: (float) initial value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param double_q: (bool) Whether to enable Double-Q learning or not. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float)alpha parameter for prioritized replay buffer. It determines how much prioritization is used, with alpha=0 corresponding to the uniform case. :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance :param full_tensorboard_log: (bool) enable additional logging when using tensorboard WARNING: this logging can take a lot of space quickly :param seed: (int) Seed for the pseudo-random generators (python, numpy, tensorflow). If None (default), use random seed. Note that if you want completely deterministic results, you must set `n_cpu_tf_sess` to 1. :param n_cpu_tf_sess: (int) The number of threads for TensorFlow operations If None, the number of cpu of the current machine will be used. """ def _create_replay_wrapper(self, env): """ Wrap the environment in a HERGoalEnvWrapper if needed and create the replay buffer wrapper. """ if not isinstance(env, HERGoalEnvWrapper): env = HERGoalEnvWrapper(env) self.env = env self.n_sampled_goal = 4 self.goal_selection_strategy = 'future' # NOTE: we cannot do that check directly with VecEnv # maybe we can try calling `compute_reward()` ? # assert isinstance(self.env, gym.GoalEnv), "HER only supports gym.GoalEnv" self.replay_wrapper = functools.partial( HindsightExperienceReplayWrapper, n_sampled_goal=self.n_sampled_goal, goal_selection_strategy=self.goal_selection_strategy, wrapped_env=self.env) def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, exploration_initial_eps=1.0, train_freq=1, batch_size=32, double_q=True, learning_starts=1000, target_network_update_freq=500, 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, n_cpu_tf_sess=None, verbose=0, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None, full_tensorboard_log=False, seed=None, n_actions=2): # TODO: replay_buffer refactoring super(DQN, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, policy_base=DQNPolicy, requires_vec_env=False, policy_kwargs=policy_kwargs, seed=seed, n_cpu_tf_sess=n_cpu_tf_sess) # for HER algorithm if self.env is not None and 'Fetch' in self.env.__str__(): self._create_replay_wrapper(self.env) self.observation_space = self.env.observation_space self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_initial_eps = exploration_initial_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.tensorboard_log = tensorboard_log self.full_tensorboard_log = full_tensorboard_log self.double_q = double_q self.graph = None self.sess = None self._train_step = None self.step_model = None self.update_target = None self.act = None self.proba_step = None self.replay_buffer = None self.beta_schedule = None self.exploration = None self.params = None self.summary = None self.n_actions = n_actions self.macro_len = 5 self.macro_count = 0 # when macro_count % macro_len == 0, resample macro action if _init_setup_model: self.setup_model() def _get_pretrain_placeholders(self): policy = self.step_model return policy.obs_ph, tf.placeholder(tf.int32, [None]), policy.q_values def setup_model(self): with SetVerbosity(self.verbose): # decision net: produce categorical distribution self.action_space = gym.spaces.discrete.Discrete(self.n_actions) assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.set_random_seed(self.seed) self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph) optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = build_train( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log=self.full_tensorboard_log, double_q=self.double_q) self.proba_step = self.step_model.proba_step self.params = tf_util.get_trainable_vars("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None, distinct_replay_buffer=False): new_tb_log = self._init_num_timesteps(reset_num_timesteps) for i, m in enumerate(self.sub_models): m.learning_rate = get_schedule_fn(m.learning_rate) if len(self.replay_wrappers) != 0: m.replay_buffer = self.replay_wrappers[i](m.replay_buffer) m._setup_learn() with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn() # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=self.exploration_initial_eps, final_p=self.exploration_final_eps) episode_rewards = [0.0] episode_successes = [] obs = self.env.reset() reset = True macro_count = 0 macro_len = self.macro_len macro_choices = [] n_updates = 0 for step in range(total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True with self.sess.as_default(): if reset or macro_count % macro_len == 0: macro_action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] # macro_action = 1 macro_obs = obs reward_in_one_macro = 0 macro_count += 1 macro_choices.append(macro_action) # use sub_model to decide action # env_action = self.sub_models[macro_action] current_sub = self.sub_models[macro_action] if self.num_timesteps < self.learning_starts or np.random.rand( ) < current_sub.random_exploration: # actions sampled from action space are from range specific to the environment # but algorithm operates on tanh-squashed actions therefore simple scaling is used unscaled_action = self.env.action_space.sample() action = scale_action(self.env.action_space, unscaled_action) else: action = current_sub.policy_tf.step( obs[None], deterministic=False).flatten() # Add noise to the action (improve exploration, # not needed in general) if current_sub.action_noise is not None: action = np.clip(action + current_sub.action_noise(), -1, 1) # inferred actions need to be transformed to environment action_space before stepping unscaled_action = unscale_action(self.env.action_space, action) assert action.shape == self.env.action_space.shape reset = False new_obs, rew, done, info = self.env.step(unscaled_action) episode_rewards[-1] += rew # rew -= self.args.policy_cost_coef * self.args.sub_policy_costs[macro_action] reward_in_one_macro += rew - self.args.policy_cost_coef * self.args.sub_policy_costs[ macro_action] # Store transition in the replay buffer. if macro_count % macro_len == 0 or done: self.replay_buffer.add(macro_obs, macro_action, reward_in_one_macro, new_obs, float(done)) for i, m in enumerate(self.sub_models): if distinct_replay_buffer: if i == macro_action: m.replay_buffer.add(obs, action, rew, new_obs, float(done)) else: m.replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) # print("step: %d, done: %d" % (self.num_timesteps, done)) if done: maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True macro_action = None macro_count = 0 prev_macro_choices = macro_choices macro_choices = [] # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. # pytype:disable=bad-unpacking if self.prioritized_replay: assert self.beta_schedule is not None, \ "BUG: should be LinearSchedule when self.prioritized_replay True" experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None # pytype:enable=bad-unpacking if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps assert isinstance(self.replay_buffer, PrioritizedReplayBuffer) self.replay_buffer.update_priorities( batch_idxes, new_priorities) if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if step % self.sub_models[0].train_freq == 0: mb_infos_vals = [] for m in self.sub_models: # Update policy, critics and target networks for grad_step in range(m.gradient_steps): # Break if the warmup phase is not over # or if there are not enough samples in the replay buffer if not m.replay_buffer.can_sample(m.batch_size) \ or self.num_timesteps < m.learning_starts: break n_updates += 1 # Compute current learning_rate frac = 1.0 - step / total_timesteps current_lr = m.learning_rate(frac) # Update policy and critics (q functions) mb_infos_vals.append( m._train_step(step, writer, current_lr)) # Update target network if (step + grad_step) % m.target_update_interval == 0: # Update target network m.sess.run(m.target_update_op) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) # print(done, log_interval, len(episode_rewards), self.num_timesteps) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) prev_macro_choices = np.array(prev_macro_choices) macro_choices_ratio = [ '%.2f' % ((prev_macro_choices[prev_macro_choices == i]).size / prev_macro_choices.size) for i in range(self.n_actions) ] logger.record_tabular("macro choices", macro_choices_ratio) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.logkv("n_updates_of_sub", n_updates) logger.dump_tabular() print("macro choices", prev_macro_choices) self.num_timesteps += 1 return self def eval(self, total_episodes, callback=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: episode_rewards = [0.0] episode_successes = [] obs = self.env.reset() reset = True macro_count = 0 macro_len = self.macro_len macro_choices = [] n_updates = 0 macro_action = None # for step in range(total_timesteps): while True: with self.sess.as_default(): if reset or macro_count % macro_len == 0: # macro_action = self.act(np.array(obs)[None], update_eps=0, **{})[0] macro_actions, _, _ = self.step_model.step( np.array(obs)[None], deterministic=False) macro_action = macro_actions[0] # macro_action = 1 macro_obs = obs reward_in_one_macro = 0 macro_count += 1 macro_choices.append(macro_action) current_sub = self.sub_models[macro_action] action = current_sub.policy_tf.step( obs[None], deterministic=True).flatten() # Add noise to the action (improve exploration, # not needed in general) if current_sub.action_noise is not None: action = np.clip(action + current_sub.action_noise(), -1, 1) # inferred actions need to be transformed to environment action_space before stepping unscaled_action = unscale_action(self.env.action_space, action) assert action.shape == self.env.action_space.shape reset = False new_obs, rew, done, info = self.env.step(unscaled_action) episode_rewards[-1] += rew rew -= self.args.policy_cost_coef * self.args.sub_policy_costs[ macro_action] reward_in_one_macro += rew obs = new_obs # print("step: %d, done: %d" % (self.num_timesteps, done)) if done: maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True macro_action = None macro_count = 0 print("=" * 70) print("macro_choices:", macro_choices) print("return:", episode_rewards[-2]) print("=" * 70) prev_macro_choices = macro_choices macro_choices = [] if len(episode_rewards) - 1 == total_episodes: break if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) # print(done, log_interval, len(episode_rewards), self.num_timesteps) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("macro choices", np.mean(prev_macro_choices)) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.logkv("n_updates_of_sub", n_updates) logger.dump_tabular() print("macro choices", prev_macro_choices) self.num_timesteps += 1 return self def predict(self, observation, state=None, mask=None, deterministic=True, args=None): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) with self.sess.as_default(): if self.macro_count % self.macro_len == 0: macro_actions, _, _ = self.step_model.step( observation, deterministic=deterministic) # macro_actions = self.act(observation, update_eps=0) self.macro_act = macro_actions[ 0] # not supporting vectorized_env if args.eval_certain_sub != None: self.macro_act = args.eval_certain_sub self.macro_count += 1 # Sample from sub_policy current_sub = self.sub_models[self.macro_act] action = current_sub.policy_tf.step(observation, deterministic=deterministic) action = action.reshape( (-1, ) + self.env.action_space.shape) # reshape to the correct action shape # inferred actions need to be transformed to environment action_space before stepping unscaled_action = unscale_action(self.env.action_space, action) unscaled_action = unscaled_action[0] # not supporting vectorized_env return self.macro_act, unscaled_action, None def action_probability(self, observation, state=None, mask=None, actions=None, logp=False): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) actions_proba = self.proba_step(observation, state, mask) if actions is not None: # comparing the action distribution, to given actions actions = np.array([actions]) assert isinstance(self.action_space, gym.spaces.Discrete) actions = actions.reshape((-1, )) assert observation.shape[0] == actions.shape[ 0], "Error: batch sizes differ for actions and observations." actions_proba = actions_proba[np.arange(actions.shape[0]), actions] # normalize action proba shape actions_proba = actions_proba.reshape((-1, 1)) if logp: actions_proba = np.log(actions_proba) if not vectorized_env: if state is not None: raise ValueError( "Error: The environment must be vectorized when using recurrent policies." ) actions_proba = actions_proba[0] return actions_proba def get_parameter_list(self): return self.params def save(self, save_path, cloudpickle=False): # params data = { "double_q": self.double_q, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "n_cpu_tf_sess": self.n_cpu_tf_sess, "seed": self.seed, "_vectorize_action": self._vectorize_action, "policy_kwargs": self.policy_kwargs } params_to_save = self.get_parameters() self._save_to_file(save_path, data=data, params=params_to_save, cloudpickle=cloudpickle)
class TD3(OffPolicyRLModel): """ Twin Delayed DDPG (TD3) Addressing Function Approximation Error in Actor-Critic Methods. Original implementation: https://github.com/sfujim/TD3 Paper: https://arxiv.org/pdf/1802.09477.pdf Introduction to TD3: https://spinningup.openai.com/en/latest/algorithms/td3.html :param policy: (TD3Policy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) the discount factor :param learning_rate: (float or callable) learning rate for adam optimizer, the same learning rate will be used for all networks (Q-Values and Actor networks) it can be a function of the current progress (from 1 to 0) :param buffer_size: (int) size of the replay buffer :param batch_size: (int) Minibatch size for each gradient update :param tau: (float) the soft update coefficient ("polyak update" of the target networks, between 0 and 1) :param policy_delay: (int) Policy and target networks will only be updated once every policy_delay steps per training steps. The Q values will be updated policy_delay more often (update every training step). :param action_noise: (ActionNoise) the action noise type. Cf DDPG for the different action noise type. :param target_policy_noise: (float) Standard deviation of gaussian noise added to target policy (smoothing noise) :param target_noise_clip: (float) Limit for absolute value of target policy smoothing noise. :param train_freq: (int) Update the model every `train_freq` steps. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param gradient_steps: (int) How many gradient update after each step :param random_exploration: (float) Probability of taking a random action (as in an epsilon-greedy strategy) This is not needed for TD3 normally but can help exploring when using HER + TD3. This hack was present in the original OpenAI Baselines repo (DDPG + HER) :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance :param policy_kwargs: (dict) additional arguments to be passed to the policy on creation :param full_tensorboard_log: (bool) enable additional logging when using tensorboard Note: this has no effect on TD3 logging for now """ def __init__(self, policy, env, gamma=0.99, learning_rate=3e-4, buffer_size=50000, buffer_type=ReplayBuffer, prioritization_starts=0, beta_schedule=None, learning_starts=100, train_freq=100, gradient_steps=100, batch_size=128, tau=0.005, policy_delay=2, action_noise=None, action_l2_scale=0, target_policy_noise=0.2, target_noise_clip=0.5, random_exploration=0.0, verbose=0, write_freq=1, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None, full_tensorboard_log=False, time_aware=False): super(TD3, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, write_freq=write_freq, policy_base=TD3Policy, requires_vec_env=False, policy_kwargs=policy_kwargs) self.prioritization_starts = prioritization_starts self.beta_schedule = beta_schedule self.buffer_is_prioritized = buffer_type.__name__ in [ "PrioritizedReplayBuffer", "RankPrioritizedReplayBuffer" ] self.loss_history = None self.buffer_type = buffer_type self.buffer_size = buffer_size self.learning_rate = learning_rate self.learning_starts = learning_starts self.train_freq = train_freq self.batch_size = batch_size self.tau = tau self.gradient_steps = gradient_steps self.gamma = gamma self.action_noise = action_noise self.action_l2_scale = action_l2_scale self.random_exploration = random_exploration self.policy_delay = policy_delay self.target_noise_clip = target_noise_clip self.target_policy_noise = target_policy_noise self.time_aware = time_aware self.graph = None self.replay_buffer = None self.episode_reward = None self.sess = None self.tensorboard_log = tensorboard_log self.verbose = verbose self.params = None self.summary = None self.policy_tf = None self.full_tensorboard_log = full_tensorboard_log self.obs_target = None self.target_policy_tf = None self.actions_ph = None self.rewards_ph = None self.terminals_ph = None self.observations_ph = None self.action_target = None self.next_observations_ph = None self.is_weights_ph = None self.step_ops = None self.target_ops = None self.infos_names = None self.target_params = None self.learning_rate_ph = None self.processed_obs_ph = None self.processed_next_obs_ph = None self.policy_out = None self.policy_train_op = None self.policy_loss = None self.active_sampling = False if _init_setup_model: self.setup_model() def _get_pretrain_placeholders(self): policy = self.policy_tf # Rescale policy_out = self.policy_out * np.abs(self.action_space.low) return policy.obs_ph, self.actions_ph, policy_out def setup_model(self): with SetVerbosity(self.verbose): self.graph = tf.Graph() with self.graph.as_default(): n_cpu = multiprocessing.cpu_count() if sys.platform == 'darwin': n_cpu //= 2 self.sess = tf_util.make_session(num_cpu=n_cpu, graph=self.graph) self.buffer_is_prioritized = self.buffer_type.__name__ in [ "PrioritizedReplayBuffer", "RankPrioritizedReplayBuffer" ] if self.replay_buffer is None: if self.buffer_is_prioritized: if self.num_timesteps is not None and self.prioritization_starts > self.num_timesteps or self.prioritization_starts > 0: self.replay_buffer = ReplayBuffer(self.buffer_size) else: buffer_kw = { "size": self.buffer_size, "alpha": 0.7 } if self.buffer_type.__name__ == "RankPrioritizedReplayBuffer": buffer_kw.update({ "learning_starts": self.prioritization_starts, "batch_size": self.batch_size }) self.replay_buffer = self.buffer_type(**buffer_kw) else: self.replay_buffer = self.buffer_type(self.buffer_size) #self.replay_buffer = DiscrepancyReplayBuffer(self.buffer_size, scorer=self.policy_tf.get_q_discrepancy) with tf.variable_scope("input", reuse=False): # Create policy and target TF objects self.policy_tf = self.policy(self.sess, self.observation_space, self.action_space, **self.policy_kwargs) self.target_policy_tf = self.policy( self.sess, self.observation_space, self.action_space, **self.policy_kwargs) # Initialize Placeholders self.observations_ph = self.policy_tf.obs_ph # Normalized observation for pixels self.processed_obs_ph = self.policy_tf.processed_obs self.next_observations_ph = self.target_policy_tf.obs_ph self.processed_next_obs_ph = self.target_policy_tf.processed_obs self.action_target = self.target_policy_tf.action_ph self.terminals_ph = tf.placeholder(tf.float32, shape=(None, 1), name='terminals') self.rewards_ph = tf.placeholder(tf.float32, shape=(None, 1), name='rewards') self.actions_ph = tf.placeholder(tf.float32, shape=(None, ) + self.action_space.shape, name='actions') self.learning_rate_ph = tf.placeholder( tf.float32, [], name="learning_rate_ph") with tf.variable_scope("model", reuse=False): # Create the policy self.policy_out = policy_out = self.policy_tf.make_actor( self.processed_obs_ph) self.policy_test = policy_test = self.policy_tf.make_actor( self.processed_obs_ph, scope="pi_t") # Use two Q-functions to improve performance by reducing overestimation bias qf1, qf2 = self.policy_tf.make_critics( self.processed_obs_ph, self.actions_ph) # Q value when following the current policy qf1_pi, qf2_pi = self.policy_tf.make_critics( self.processed_obs_ph, policy_out, reuse=True) qf1_pi_t, _ = self.policy_tf.make_critics( self.processed_obs_ph, policy_test, reuse=True) with tf.variable_scope("target", reuse=False): # Create target networks target_policy_out = self.target_policy_tf.make_actor( self.processed_next_obs_ph) # Target policy smoothing, by adding clipped noise to target actions target_noise = tf.random_normal( tf.shape(target_policy_out), stddev=self.target_policy_noise) target_noise = tf.clip_by_value(target_noise, -self.target_noise_clip, self.target_noise_clip) # Clip the noisy action to remain in the bounds [-1, 1] (output of a tanh) noisy_target_action = tf.clip_by_value( target_policy_out + target_noise, -1, 1) # Q values when following the target policy qf1_target, qf2_target = self.target_policy_tf.make_critics( self.processed_next_obs_ph, noisy_target_action) with tf.variable_scope("loss", reuse=False): # Take the min of the two target Q-Values (clipped Double-Q Learning) min_qf_target = tf.minimum(qf1_target, qf2_target) # Targets for Q value regression q_backup = tf.stop_gradient(self.rewards_ph + (1 - self.terminals_ph) * self.gamma * min_qf_target) # Compute Q-Function loss if self.buffer_is_prioritized: self.is_weights_ph = tf.placeholder(tf.float32, shape=(None, 1), name="is_weights") qf1_loss = tf.reduce_mean(self.is_weights_ph * (q_backup - qf1)**2) qf2_loss = tf.reduce_mean(self.is_weights_ph * (q_backup - qf2)**2) else: qf1_loss = tf.reduce_mean((q_backup - qf1)**2) qf2_loss = tf.reduce_mean((q_backup - qf2)**2) q_discrepancy = tf.abs(qf1_pi - qf2_pi) self.q_disc_strength_ph = tf.placeholder( tf.float32, [], name="q_disc_strength_ph") self.q_disc_strength_schedule = ExponentialSchedule( int(1e5), 30, 0, rate=10) qvalues_losses = qf1_loss + qf2_loss rew_loss = tf.reduce_mean(qf1_pi) q_disc_loss = tf.reduce_mean( q_discrepancy ) #self.q_disc_strength_ph * tf.reduce_mean(q_discrepancy) action_loss = self.action_l2_scale * tf.nn.l2_loss( self.policy_out) # Policy loss: maximise q value self.policy_loss = policy_loss = -( rew_loss + q_disc_loss) + action_loss self.policy_loss_t = policy_loss_t = -tf.reduce_mean( qf1_pi_t) # Policy train op # will be called only every n training steps, # where n is the policy delay policy_optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate_ph) policy_train_op = policy_optimizer.minimize( policy_loss, var_list=get_vars('model/pi')) self.policy_train_op = policy_train_op policy_optimizer_t = tf.train.AdamOptimizer( learning_rate=self.learning_rate_ph) policy_train_op_t = policy_optimizer_t.minimize( policy_loss_t, var_list=get_vars('model/pi_t')) self.policy_train_op_t = policy_train_op_t # Q Values optimizer qvalues_optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate_ph) qvalues_params = get_vars('model/values_fn/') # Q Values and policy target params source_params = get_vars("model/") target_params = get_vars("target/") source_params = [ param for param in source_params if "pi_t" not in param.name ] # Polyak averaging for target variables self.target_ops = [ tf.assign(target, (1 - self.tau) * target + self.tau * source) for target, source in zip(target_params, source_params) ] # Initializing target to match source variables target_init_op = [ tf.assign(target, source) for target, source in zip(target_params, source_params) ] train_values_op = qvalues_optimizer.minimize( qvalues_losses, var_list=qvalues_params) self.infos_names = ['qf1_loss', 'qf2_loss'] # All ops to call during one training step self.step_ops = [ qf1_loss, qf2_loss, qf1, qf2, train_values_op, q_discrepancy ] # Monitor losses and entropy in tensorboard tf.summary.scalar("rew_loss", rew_loss) tf.summary.scalar("q_disc_loss", q_disc_loss) tf.summary.scalar("action_loss", action_loss) tf.summary.scalar('policy_loss', policy_loss) tf.summary.scalar("policy_loss_t", policy_loss_t) tf.summary.scalar('qf1_loss', qf1_loss) tf.summary.scalar('qf2_loss', qf2_loss) tf.summary.scalar('learning_rate', tf.reduce_mean(self.learning_rate_ph)) # Retrieve parameters that must be saved self.params = get_vars("model") self.target_params = get_vars("target/") # Initialize Variables and target network with self.sess.as_default(): self.sess.run(tf.global_variables_initializer()) self.sess.run(target_init_op) self.summary = tf.summary.merge_all() def _train_step(self, step, writer, learning_rate, update_policy): # Sample a batch from the replay buffer if self.buffer_is_prioritized and self.num_timesteps >= self.prioritization_starts: batch = self.replay_buffer.sample(self.batch_size, beta=self.beta_schedule( self.num_timesteps)) batch_obs, batch_actions, batch_rewards, batch_next_obs, batch_dones, batch_weights, batch_idxs = batch if len(batch_weights.shape) == 1: batch_weights = np.expand_dims(batch_weights, axis=1) else: batch = self.replay_buffer.sample(self.batch_size) batch_obs, batch_actions, batch_rewards, batch_next_obs, batch_dones = batch if self.buffer_is_prioritized: batch_weights = np.ones(shape=(self.batch_size, 1)) feed_dict = { self.observations_ph: batch_obs, self.actions_ph: batch_actions, self.next_observations_ph: batch_next_obs, self.rewards_ph: batch_rewards.reshape(self.batch_size, -1), self.terminals_ph: batch_dones.reshape(self.batch_size, -1), self.learning_rate_ph: learning_rate, self.q_disc_strength_ph: self.q_disc_strength_schedule(self.num_timesteps) } if self.buffer_is_prioritized: feed_dict[self.is_weights_ph] = batch_weights step_ops = self.step_ops if update_policy: # Update policy and target networks step_ops = step_ops + [ self.policy_train_op, self.policy_train_op_t, self.target_ops, self.policy_loss, self.policy_loss_t ] # Do one gradient step # and optionally compute log for tensorboard if writer is not None: out = self.sess.run([self.summary] + step_ops, feed_dict) summary = out.pop(0) writer.add_summary(summary, step) else: out = self.sess.run(step_ops, feed_dict) # Unpack to monitor losses q_discrepancies = out.pop(5) qf1_loss, qf2_loss, *_values = out if self.buffer_is_prioritized and self.num_timesteps >= self.prioritization_starts: if isinstance(self.replay_buffer, HindsightExperienceReplayWrapper): self.replay_buffer.replay_buffer.update_priorities( batch_idxs, q_discrepancies) else: self.replay_buffer.update_priorities(batch_idxs, q_discrepancies) return qf1_loss, qf2_loss, q_discrepancies def learn(self, total_timesteps, callback=None, seed=None, log_interval=4, tb_log_name="TD3", reset_num_timesteps=True, replay_wrapper=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) last_replay_update = 0 if replay_wrapper is not None: self.replay_buffer = replay_wrapper(self.replay_buffer) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn(seed) # Transform to callable if needed self.learning_rate = get_schedule_fn(self.learning_rate) # Initial learning rate current_lr = self.learning_rate(1) start_time = time.time() episode_rewards = [0.0] episode_successes = [] if self.action_noise is not None: self.action_noise.reset() obs = self.env.reset() self.episode_reward = np.zeros((1, )) ep_info_buf = deque(maxlen=100) n_updates = 0 infos_values = [] self.active_sampling = False initial_step = self.num_timesteps if self.buffer_is_prioritized and \ ((replay_wrapper is not None and self.replay_buffer.replay_buffer.__name__ == "ReplayBuffer") or (replay_wrapper is None and self.replay_buffer.__name__ == "ReplayBuffer")) \ and self.num_timesteps >= self.prioritization_starts: self._set_prioritized_buffer() for step in range(initial_step, total_timesteps): if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Before training starts, randomly sample actions # from a uniform distribution for better exploration. # Afterwards, use the learned policy # if random_exploration is set to 0 (normal setting) if (self.num_timesteps < self.learning_starts or np.random.rand() < self.random_exploration): # No need to rescale when sampling random action rescaled_action = action = self.env.action_space.sample() else: action = self.policy_tf.step(obs[None]).flatten() # Add noise to the action, as the policy # is deterministic, this is required for exploration if self.action_noise is not None: action = np.clip(action + self.action_noise(), -1, 1) # Rescale from [-1, 1] to the correct bounds rescaled_action = action * np.abs(self.action_space.low) assert action.shape == self.env.action_space.shape new_obs, reward, done, info = self.env.step(rescaled_action) # Store transition in the replay buffer. self.replay_buffer.add( obs, action, reward, new_obs, float(done if not self.time_aware else done and info["termination"] != "steps")) obs = new_obs if ((replay_wrapper is not None and self.replay_buffer.replay_buffer.__name__ == "RankPrioritizedReplayBuffer")\ or self.replay_buffer.__name__ == "RankPrioritizedReplayBuffer") and \ self.num_timesteps % self.buffer_size == 0: self.replay_buffer.rebalance() # Retrieve reward and episode length if using Monitor wrapper maybe_ep_info = info.get('episode') if maybe_ep_info is not None: ep_info_buf.extend([maybe_ep_info]) if writer is not None: # Write reward per episode to tensorboard ep_reward = np.array([reward]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_reward, ep_done, writer, self.num_timesteps) if step % self.train_freq == 0: mb_infos_vals = [] # Update policy, critics and target networks for grad_step in range(self.gradient_steps): # Break if the warmup phase is not over # or if there are not enough samples in the replay buffer if not self.replay_buffer.can_sample(self.batch_size) \ or self.num_timesteps < self.learning_starts: break n_updates += 1 # Compute current learning_rate frac = 1.0 - self.num_timesteps / total_timesteps current_lr = self.learning_rate(frac) # Update policy and critics (q functions) # Note: the policy is updated less frequently than the Q functions # this is controlled by the `policy_delay` parameter step_writer = writer if grad_step % self.write_freq == 0 else None mb_infos_vals.append( self._train_step(step, step_writer, current_lr, (step + grad_step) % self.policy_delay == 0)) # Log losses and entropy, useful for monitor training if len(mb_infos_vals) > 0: infos_values = np.mean(mb_infos_vals, axis=0) episode_rewards[-1] += reward if done: if isinstance(self.replay_buffer, DiscrepancyReplayBuffer ) and n_updates - last_replay_update >= 5000: self.replay_buffer.update_priorities() last_replay_update = n_updates if self.action_noise is not None: self.action_noise.reset() if not isinstance(self.env, VecEnv): if self.active_sampling: sample_obs, sample_state = self.env.get_random_initial_states( 25) obs_discrepancies = self.policy_tf.get_q_discrepancy( sample_obs) obs = self.env.reset( **sample_state[np.argmax(obs_discrepancies)]) else: obs = self.env.reset() episode_rewards.append(0.0) maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if len(episode_rewards[-101:-1]) == 0: mean_reward = -np.inf else: mean_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) self.num_timesteps += 1 if self.buffer_is_prioritized and \ ((replay_wrapper is not None and self.replay_buffer.replay_buffer.__name__ == "ReplayBuffer") or (replay_wrapper is None and self.replay_buffer.__name__ == "ReplayBuffer"))\ and self.num_timesteps >= self.prioritization_starts: self._set_prioritized_buffer() # Display training infos if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: fps = int(step / (time.time() - start_time)) logger.logkv("episodes", num_episodes) logger.logkv("mean 100 episode reward", mean_reward) if len(ep_info_buf) > 0 and len(ep_info_buf[0]) > 0: logger.logkv( 'ep_rewmean', safe_mean( [ep_info['r'] for ep_info in ep_info_buf])) logger.logkv( 'eplenmean', safe_mean( [ep_info['l'] for ep_info in ep_info_buf])) logger.logkv("n_updates", n_updates) logger.logkv("current_lr", current_lr) logger.logkv("fps", fps) logger.logkv('time_elapsed', int(time.time() - start_time)) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) if len(infos_values) > 0: for (name, val) in zip(self.infos_names, infos_values): logger.logkv(name, val) logger.logkv("total timesteps", self.num_timesteps) logger.dumpkvs() # Reset infos: infos_values = [] return self def action_probability(self, observation, state=None, mask=None, actions=None, logp=False): _ = np.array(observation) if actions is not None: raise ValueError("Error: TD3 does not have action probabilities.") # here there are no action probabilities, as DDPG does not use a probability distribution warnings.warn( "Warning: action probability is meaningless for TD3. Returning None" ) return None def predict(self, observation, state=None, mask=None, deterministic=True): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) actions = self.policy_tf.step(observation, test=deterministic) if self.action_noise is not None and not deterministic: actions = np.clip(actions + self.action_noise(), -1, 1) actions = actions.reshape( (-1, ) + self.action_space.shape) # reshape to the correct action shape actions = actions * np.abs( self.action_space.low) # scale the output for the prediction if not vectorized_env: actions = actions[0] return actions, None def _get_env(self): env = self.env env = env.env return env def _set_prioritized_buffer(self): buffer_kw = {"size": self.buffer_size, "alpha": 0.7} if self.buffer_type.__name__ == "RankPrioritizedReplayBuffer": buffer_kw.update({ "learning_starts": self.prioritization_starts, "batch_size": self.batch_size }) r_buf = self.buffer_type(**buffer_kw) for i, transition in enumerate(self.replay_buffer._storage): r_buf.add(*transition) r_buf.update_priorities([i], self.policy_tf.get_q_discrepancy( transition[0])[0]) if r_buf.__name__ == "RankPrioritizedReplayBuffer": r_buf.rebalance() if isinstance(self.replay_buffer, HindsightExperienceReplayWrapper): self.replay_buffer.replay_buffer = r_buf else: self.replay_buffer = r_buf self.learning_rate = get_schedule_fn( self.learning_rate(1) / 4) # TODO: will not work with non-constant self.beta_schedule = get_schedule_fn(self.beta_schedule) print("Enabled prioritized replay buffer") def get_parameter_list(self): return (self.params + self.target_params) def save(self, save_path): data = { "learning_rate": self.learning_rate, "buffer_size": self.buffer_size, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "batch_size": self.batch_size, "tau": self.tau, # Should we also store the replay buffer? # this may lead to high memory usage # with all transition inside "replay_buffer": self.replay_buffer, "policy_delay": self.policy_delay, "target_noise_clip": self.target_noise_clip, "target_policy_noise": self.target_policy_noise, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "action_noise": self.action_noise, "random_exploration": self.random_exploration, "_vectorize_action": self._vectorize_action, "policy_kwargs": self.policy_kwargs, "num_timesteps": self.num_timesteps } params_to_save = self.get_parameters() self._save_to_file(save_path, data=data, params=params_to_save)
class DQN(OffPolicyRLModel): """ The DQN model class. DQN paper: https://arxiv.org/pdf/1312.5602.pdf :param policy: (DQNPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, LnMlpPolicy, ...) :param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str) :param gamma: (float) discount factor :param learning_rate: (float) learning rate for adam optimizer :param buffer_size: (int) size of the replay buffer :param exploration_fraction: (float) fraction of entire training period over which the exploration rate is annealed :param exploration_final_eps: (float) final value of random action probability :param train_freq: (int) update the model every `train_freq` steps. set to None to disable printing :param batch_size: (int) size of a batched sampled from replay buffer for training :param 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. :param checkpoint_path: (str) replacement path used if you need to log to somewhere else than a temporary directory. :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param target_network_update_freq: (int) update the target network every `target_network_update_freq` steps. :param prioritized_replay: (bool) if True prioritized replay buffer will be used. :param prioritized_replay_alpha: (float)alpha parameter for prioritized replay buffer. It determines how much prioritization is used, with alpha=0 corresponding to the uniform case. :param prioritized_replay_beta0: (float) initial value of beta for prioritized replay buffer :param 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 max_timesteps. :param prioritized_replay_eps: (float) epsilon to add to the TD errors when updating priorities. :param param_noise: (bool) Whether or not to apply noise to the parameters of the policy. :param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug :param tensorboard_log: (str) the log location for tensorboard (if None, no logging) :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance :param full_tensorboard_log: (bool) enable additional logging when using tensorboard WARNING: this logging can take a lot of space quickly """ def __init__(self, policy, env, gamma=0.99, learning_rate=5e-4, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, target_network_update_freq=500, 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, verbose=0, tensorboard_log=None, _init_setup_model=True, policy_kwargs=None, full_tensorboard_log=False): # TODO: replay_buffer refactoring super(DQN, self).__init__(policy=policy, env=env, replay_buffer=None, verbose=verbose, policy_base=DQNPolicy, requires_vec_env=False, policy_kwargs=policy_kwargs) self.checkpoint_path = checkpoint_path self.param_noise = param_noise self.learning_starts = learning_starts self.train_freq = train_freq self.prioritized_replay = prioritized_replay self.prioritized_replay_eps = prioritized_replay_eps self.batch_size = batch_size self.target_network_update_freq = target_network_update_freq self.checkpoint_freq = checkpoint_freq self.prioritized_replay_alpha = prioritized_replay_alpha self.prioritized_replay_beta0 = prioritized_replay_beta0 self.prioritized_replay_beta_iters = prioritized_replay_beta_iters self.exploration_final_eps = exploration_final_eps self.exploration_fraction = exploration_fraction self.buffer_size = buffer_size self.learning_rate = learning_rate self.gamma = gamma self.tensorboard_log = tensorboard_log self.full_tensorboard_log = full_tensorboard_log self.graph = None self.sess = None self._train_step = None self.step_model = None self.update_target = None self.act = None self.proba_step = None self.replay_buffer = None self.beta_schedule = None self.exploration = None self.params = None self.summary = None self.episode_reward = None if _init_setup_model: self.setup_model() def _get_pretrain_placeholders(self): policy = self.step_model return policy.obs_ph, tf.placeholder(tf.int32, [None]), policy.q_values @staticmethod def _is_vectorized_observation(observation, observation_space): """ For every observation type, detects and validates the shape, then returns whether or not the observation is vectorized. :param observation: (np.ndarray) the input observation to validate :param observation_space: (gym.spaces) the observation space :return: (bool) whether the given observation is vectorized or not """ if isinstance(observation_space, gym.spaces.Box): if observation.shape == observation_space.shape: return False elif observation.shape[1:] == observation_space.shape: return True else: raise ValueError( "Error: Unexpected observation shape {} for ".format( observation.shape) + "Box environment, please use {} ".format( observation_space.shape) + "or (n_env, {}) for the observation shape.".format( ", ".join(map(str, observation_space.shape)))) elif isinstance(observation_space, gym.spaces.Discrete): if observation.shape == ( ): # A numpy array of a number, has shape empty tuple '()' return False elif len(observation.shape) == 1: return True else: raise ValueError( "Error: Unexpected observation shape {} for ".format( observation.shape) + "Discrete environment, please use (1,) or (n_env, 1) for the observation shape." ) elif isinstance(observation_space, gym.spaces.MultiDiscrete): if observation.shape == (len(observation_space.nvec), ): return False elif len(observation.shape) == 2 and observation.shape[1] == len( observation_space.nvec): return True else: raise ValueError( "Error: Unexpected observation shape {} for MultiDiscrete " .format(observation.shape) + "environment, please use ({},) or ".format( len(observation_space.nvec)) + "(n_env, {}) for the observation shape.".format( len(observation_space.nvec))) elif isinstance(observation_space, gym.spaces.MultiBinary): if observation.shape == (observation_space.n, ): return False elif len(observation.shape ) == 2 and observation.shape[1] == observation_space.n: return True else: raise ValueError( "Error: Unexpected observation shape {} for MultiBinary ". format(observation.shape) + "environment, please use ({},) or ".format( observation_space.n) + "(n_env, {}) for the observation shape.".format( observation_space.n)) elif isinstance(observation_space, gym.spaces.Dict): import logging # print("the observation is the space type {} " # "and the observation is in vector with shape {}.".format(observation_space, observation.shape)) pass else: raise ValueError( "Error: Cannot determine if the observation is vectorized with the space type {}." .format(observation_space)) def setup_model(self): with SetVerbosity(self.verbose): assert not isinstance(self.action_space, gym.spaces.Box), \ "Error: DQN cannot output a gym.spaces.Box action space." # If the policy is wrap in functool.partial (e.g. to disable dueling) # unwrap it to check the class type if isinstance(self.policy, partial): test_policy = self.policy.func else: test_policy = self.policy assert issubclass(test_policy, DQNPolicy), "Error: the input policy for the DQN model must be " \ "an instance of DQNPolicy." self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf_util.make_session(graph=self.graph) optimizer = tf.train.AdamOptimizer( learning_rate=self.learning_rate) self.act, self._train_step, self.update_target, self.step_model = deepq.build_train( q_func=partial(self.policy, **self.policy_kwargs), ob_space=self.observation_space, ac_space=self.action_space, optimizer=optimizer, gamma=self.gamma, grad_norm_clipping=10, param_noise=self.param_noise, sess=self.sess, full_tensorboard_log=self.full_tensorboard_log) self.proba_step = self.step_model.proba_step self.params = tf_util.get_trainable_vars("deepq") # Initialize the parameters and copy them to the target network. tf_util.initialize(self.sess) self.update_target(sess=self.sess) self.summary = tf.summary.merge_all() def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN", reset_num_timesteps=True, replay_wrapper=None, learning_curve=False, test_t=None): new_tb_log = self._init_num_timesteps(reset_num_timesteps) with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \ as writer: self._setup_learn(seed) # Create the replay buffer if self.prioritized_replay: self.replay_buffer = PrioritizedReplayBuffer( self.buffer_size, alpha=self.prioritized_replay_alpha) if self.prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps else: prioritized_replay_beta_iters = self.prioritized_replay_beta_iters self.beta_schedule = LinearSchedule( prioritized_replay_beta_iters, initial_p=self.prioritized_replay_beta0, final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.buffer_size) self.beta_schedule = None if replay_wrapper is not None: assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER" self.replay_buffer = replay_wrapper(self.replay_buffer) # Create the schedule for exploration starting from 1. self.exploration = LinearSchedule( schedule_timesteps=int(self.exploration_fraction * total_timesteps), initial_p=1.0, final_p=self.exploration_final_eps) episode_rewards = [0.0] self.cumul_reward = [0.0] episode_successes = [] obs = self.env.reset() reset = True self.episode_reward = np.zeros((1, )) # variables for test eval ## test_step = test_t * 3 test_results = {'sum': []} test_ts = [] for _ in range(total_timesteps): ## Test eval period ## if learning_curve and _ % test_step == 0 and _ > 0: print("--> Simulating test period") self.env.reset() test_r = 0.0 for i in range(test_t): feasible_actions = AllocationEnv.get_feasible_actions( obs["board_config"]) action_mask = AllocationEnv.get_action_mask( feasible_actions, self.env.action_space.n) action, _states = self.predict(obs, mask=action_mask) action = AllocationEnv.check_action( obs['board_config'], action) obs, rewards, dones, info = self.env.step(action) test_r += rewards test_results["sum"].append(test_r) test_ts.append(_) self.env.reset() # plot test eval progress plt.plot(test_ts, test_results["sum"]) # plt.errorbar(iteration_cuts, results["mean"], yerr=results["std"], fmt='.k') plt.xlabel("Iteration count") plt.ylabel("Total (sum) test reward") plt.savefig("figs/rl-learning-curve-{}.pdf".format( cfg.vals['prj_name'])) plt.clf() plt.close() # write test eval progress write_results = {} for k, v in test_results.items(): write_results[k] = serialize_floats(v) with open( "output/rl-learning-curve-{}.json".format( cfg.vals['prj_name']), 'w') as f: json.dump(write_results, f) if callback is not None: # Only stop training if return value is False, not when it is None. This is for backwards # compatibility with callbacks that have no return statement. if callback(locals(), globals()) is False: break # Take action and update exploration to the newest value kwargs = {} if not self.param_noise: update_eps = self.exploration.value(self.num_timesteps) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = \ -np.log(1. - self.exploration.value(self.num_timesteps) + self.exploration.value(self.num_timesteps) / float(self.env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True feasible_actions = AllocationEnv.get_feasible_actions( obs["board_config"]) action_mask = AllocationEnv.get_action_mask( feasible_actions, self.action_space.n) with self.sess.as_default(): action = self.act(State.get_vec_observation(obs)[None], update_eps=update_eps, **kwargs, mask=action_mask)[0] reset = False # CHECK IF ACTIONS IS FEASIBLE action = AllocationEnv.check_action(obs['board_config'], action) env_action = action new_obs, rew, done, info = self.env.step(env_action) print("action: {} - reward: {} - eps: {:.4}".format( action, rew, update_eps)) print(new_obs['day_vec']) print(new_obs['board_config']) # Store transition in the replay buffer. self.replay_buffer.add(State.get_vec_observation(obs), action, rew, State.get_vec_observation(new_obs), float(done)) obs = new_obs if writer is not None: ep_rew = np.array([rew]).reshape((1, -1)) ep_done = np.array([done]).reshape((1, -1)) self.episode_reward = total_episode_reward_logger( self.episode_reward, ep_rew, ep_done, writer, self.num_timesteps) episode_rewards[-1] += rew self.cumul_reward.append(self.cumul_reward[-1] + rew) if done: maybe_is_success = info.get('is_success') if maybe_is_success is not None: episode_successes.append(float(maybe_is_success)) if not isinstance(self.env, VecEnv): obs = self.env.reset() episode_rewards.append(0.0) reset = True # Do not train if the warmup phase is not over # or if there are not enough samples in the replay buffer can_sample = self.replay_buffer.can_sample(self.batch_size) if can_sample and self.num_timesteps > self.learning_starts \ and self.num_timesteps % self.train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if self.prioritized_replay: experience = self.replay_buffer.sample( self.batch_size, beta=self.beta_schedule.value(self.num_timesteps)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None if writer is not None: # run loss backprop with summary, but once every 100 steps save the metadata # (memory, compute time, ...) if (1 + self.num_timesteps) % 100 == 0: run_options = tf.RunOptions( trace_level=tf.RunOptions.FULL_TRACE) run_metadata = tf.RunMetadata() summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess, options=run_options, run_metadata=run_metadata) writer.add_run_metadata( run_metadata, 'step%d' % self.num_timesteps) else: summary, td_errors = self._train_step( obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) writer.add_summary(summary, self.num_timesteps) else: _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) if self.prioritized_replay: new_priorities = np.abs( td_errors) + self.prioritized_replay_eps self.replay_buffer.update_priorities( batch_idxes, new_priorities) if can_sample and self.num_timesteps > self.learning_starts and \ self.num_timesteps % self.target_network_update_freq == 0: # Update target network periodically. self.update_target(sess=self.sess) if len(episode_rewards[-101:-1]) == 0: mean_100ep_reward = -np.inf else: mean_100ep_reward = round( float(np.mean(episode_rewards[-101:-1])), 1) num_episodes = len(episode_rewards) if self.verbose >= 1 and done and log_interval is not None and len( episode_rewards) % log_interval == 0: logger.record_tabular("steps", self.num_timesteps) logger.record_tabular("episodes", num_episodes) if len(episode_successes) > 0: logger.logkv("success rate", np.mean(episode_successes[-100:])) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular( "% time spent exploring", int(100 * self.exploration.value(self.num_timesteps))) logger.dump_tabular() print('timestamp: {}'.format(self.num_timesteps, end='\r\n')) self.num_timesteps += 1 return self def predict(self, observation, state=None, mask=None, deterministic=True): if isinstance(observation, dict): observation = State.get_vec_observation(observation)[None] vectorized_env = self._is_vectorized_observation( observation, self.observation_space) with self.sess.as_default(): actions, _, _ = self.step_model.step(observation, deterministic=deterministic, mask=mask) if not vectorized_env: actions = actions[0] return actions, None def action_probability(self, observation, state=None, mask=None, actions=None, logp=False): observation = np.array(observation) vectorized_env = self._is_vectorized_observation( observation, self.observation_space) observation = observation.reshape((-1, ) + self.observation_space.shape) actions_proba = self.proba_step(observation, state, mask) if actions is not None: # comparing the action distribution, to given actions actions = np.array([actions]) assert isinstance(self.action_space, gym.spaces.Discrete) actions = actions.reshape((-1, )) assert observation.shape[0] == actions.shape[ 0], "Error: batch sizes differ for actions and observations." actions_proba = actions_proba[np.arange(actions.shape[0]), actions] # normalize action proba shape actions_proba = actions_proba.reshape((-1, 1)) if logp: actions_proba = np.log(actions_proba) if not vectorized_env: if state is not None: raise ValueError( "Error: The environment must be vectorized when using recurrent policies." ) actions_proba = actions_proba[0] return actions_proba def get_parameter_list(self): return self.params def save(self, save_path): # params data = { "checkpoint_path": self.checkpoint_path, "param_noise": self.param_noise, "learning_starts": self.learning_starts, "train_freq": self.train_freq, "prioritized_replay": self.prioritized_replay, "prioritized_replay_eps": self.prioritized_replay_eps, "batch_size": self.batch_size, "target_network_update_freq": self.target_network_update_freq, "checkpoint_freq": self.checkpoint_freq, "prioritized_replay_alpha": self.prioritized_replay_alpha, "prioritized_replay_beta0": self.prioritized_replay_beta0, "prioritized_replay_beta_iters": self.prioritized_replay_beta_iters, "exploration_final_eps": self.exploration_final_eps, "exploration_fraction": self.exploration_fraction, "learning_rate": self.learning_rate, "gamma": self.gamma, "verbose": self.verbose, "observation_space": self.observation_space, "action_space": self.action_space, "policy": self.policy, "n_envs": self.n_envs, "_vectorize_action": self._vectorize_action, "policy_kwargs": self.policy_kwargs } params_to_save = self.get_parameters() self._save_to_file(save_path, data=data, params=params_to_save) def update_weights(self, buffer): obses_t, actions, rewards, obses_tp1, dones = buffer.sample( self.batch_size) weights, batch_idxes = np.ones_like(rewards), None _, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights, sess=self.sess) def learn_off_policy(self, total_timesteps, buffer): print("Off-policy training") for i in tqdm(range(total_timesteps)): self.update_weights(buffer)
class PacmanPWDQN(Agent): """ Creates the Wide Deep Q-Network Agent that iterates with the environment In addition, this agent can be set up to purely Linear or DQN Agent """ def __init__(self, args): # Load parameters from user-given arguments self.params = json_to_dict(args["path"]) os.environ["CUDA_VISIBLE_DEVICES"] = str(self.params["GPU"]) self.params['width'] = args['width'] self.params['height'] = args['height'] self.params['num_training'] = args['numTraining'] self.params['num_games'] = args['numGames'] self.path_extra = "" self.params["seed"] = args['seed'] self.random = np.random.RandomState(self.params["seed"]) self.beta_schedule = None # time started self.general_record_time = time.strftime("%a_%d_%b_%Y_%H_%M_%S", time.localtime()) self.start_time = time.time() self.rank_sort = None if self.params["prioritized"]: # For using PrioritizedReplayBuffer if self.params["ranked"]: N_list = [self.params["batch_size"]] + [int(x) for x in np.linspace(100, self.params["mem_size"], 5)] save_quantiles(N_list=N_list, k=self.params["batch_size"], alpha=self.params["prioritized_replay_alpha"], name=self.params["save_file"]) self.replay_buffer = RankBasedReplay(self.params["mem_size"], self.params["prioritized_replay_alpha"], name=self.params["save_file"]) if self.params["sort_rank"] == None: # For sorting rankbased buffer self.rank_sort = int(self.params["mem_size"] * 0.01) else: self.rank_sort = self.params["sort_rank"] else: self.replay_buffer = PrioritizedReplayBuffer(self.params["mem_size"], self.params["prioritized_replay_alpha"]) if self.params["prioritized_replay_beta_iters"] is None: prioritized_replay_beta_iters = self.params['num_training'] else: prioritized_replay_beta_iters = self.params['prioritized_replay_beta_iters'] self.beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=self.params['prioritized_replay_beta0'], final_p=1.0) else: self.replay_buffer = ReplayBuffer(self.params["mem_size"]) self.beta_schedule = None if self.params["only_dqn"]: print("Initialise DQN Agent") elif self.params["only_lin"]: print("Initialise Linear Approximative Agent") else: print("Initialise WDQN Agent") print(self.params["save_file"]) if self.params["prioritized"]: if self.params["ranked"]: print("Using Rank-Based Experience Replay Buffer") else: print("Using Prioritized Experience Replay Buffer") if self.params["model_shift"]: print("Using Model Shift") print("seed", self.params["seed"]) print("Starting time:", self.general_record_time) # Start Tensorflow session tf.reset_default_graph() tf.set_random_seed(self.params["seed"]) self.qnet = WDQN(self.params, "model") # Q-network self.tnet = WDQN(self.params, "target_model") # Q-target-network self.saver = tf.train.Saver() self.sess = tf.Session() self.qnet.set_session(self.sess) self.tnet.set_session(self.sess) self.sess.run(tf.global_variables_initializer()) # Q and cost self.Q_global = [] # Stats self.cnt = self.qnet.sess.run(self.qnet.global_step_dqn) self.local_cnt = 0 self.wins = 0 self.best_int = self.params["shift_best"] self.numeps = 0 self.model_eps = 0 self.episodeStartTime = time.time() self.last_steps = 0 self.get_direction = lambda k: ['North', 'South', 'East', 'West', 'Stop'][k] self.get_value = {'North': 0, 'South': 1, 'East': 2, 'West': 3, 'Stop': 4} self.lastWindowAccumRewards = 0.0 self.Q_accumulative = 0.0 self.accumTrainRewards = 0.0 self.sub_dir = str(self.params["save_interval"]) def registerInitialState(self, state): """Inspects the starting state""" # Reset reward self.last_score = 0 self.last_reward = 0. # Reset state self.last_state = None self.current_state = state # Reset actions self.last_action = None # Reset vars self.terminal = None self.won = True self.Q_global = [] # Shift Model between WDQN and DQN during training if self.params["model_shift"] and (self.numeps + 1) <= self.params['num_training']: if (self.numeps +1) >= self.params["start_shift"] and (self.numeps +1) % self.params["val_shift"] == 0: if self.params["only_dqn"]: self.params["only_dqn"] = False print("Using WDQN Agent starting from eps", (self.numeps + 1)) else: self.params["only_dqn"] = True print("Using DQN Agent starting from eps", (self.numeps + 1)) if self.params["model_shift"] and (self.numeps + 1) == self.params['num_training'] and not self.params["only_dqn"]: # Back to WDQN at the end self.params["only_dqn"] = False print("Back to WDQN Agent for testing") # Load model self.load_mod() # Next self.numeps += 1 def getQvalues(self, model, dropout): """Access Q Values by using the model prediction of WDQN.py""" if self.params["only_dqn"]: return model.predict_dqn(map_state_mat(self.current_state), dropout)[0] elif self.params["only_lin"]: return model.predict_lin(mat_features(self.current_state, ftrs=self.params["feat_val"]), dropout)[0] else: return model.predict_wdqn(map_state_mat(self.current_state), mat_features(self.current_state, ftrs=self.params["feat_val"]), dropout)[0] def getPolicy(self, model, dropout=1.0): """Pick up the policy """ qValues = self.getQvalues(model, dropout) qVal = {self.get_value[l]: qValues[self.get_value[l]] for l in self.current_state.getLegalActions(0) if not l == "Stop"} maxValue = max(qVal.values()) self.Q_global.append(maxValue) return self.get_direction(self.random.choice([k for k in qVal.keys() if qVal[k] == maxValue])) def getAction(self, state): """Exploit / Explore""" if self.random.rand() > self.params['eps']: # Exploit action move = self.getPolicy(self.qnet) # dropout deactivated else: legal = [v for v in state.getLegalActions(0) if not v == "Stop"] move = self.random.choice(legal) # Save last_action self.last_action = self.get_value[move] return move def observationFunction(self, state): """Do observation""" self.terminal = False self.observation_step(state) return state def observation_step(self, state): """ Realize the observation step Rewards are balanced in this part The training occurs in this section """ if self.last_action is not None: # Process current experience state self.last_state = self.current_state.deepCopy() self.current_state = state # Process current experience reward reward = state.getScore() - self.last_score self.last_score = state.getScore() # Reward system if reward > 20: self.last_reward = 50 # 0.1 # Eat ghost elif reward > 0: self.last_reward = 10 # 0.02 # Eat food elif reward < -10: self.last_reward = -500. # -1 # Get eaten self.won = False elif reward < 0: self.last_reward = -1 # -0.002 # Punish time if (self.terminal and self.won): self.last_reward = 100 # 0.2 # Won if self.isInTraining(): # Copy values to target network if self.local_cnt % self.params["target_update_network"] == 0 \ and self.local_cnt > self.params['train_start']: self.tnet.rep_network(self.qnet) print("Copied model parameters to target network. total_t = %s, period = %s" % ( self.local_cnt, self.params["target_update_network"])) # Store last experience into memory if self.params["prioritized"] and self.params["ranked"]: self.replay_buffer.add((self.last_state, self.last_action, float(self.last_reward), self.current_state, self.terminal)) else: self.replay_buffer.add(self.last_state, self.last_action, float(self.last_reward), self.current_state, self.terminal) # Train self.train() # Next self.local_cnt += 1 if self.local_cnt == self.params['train_start']: print("") print("Memory Replay populated") print("") self.model_eps = self.numeps # with open('data/lin_rb.pickle', 'wb') as handle: # pickle.dump(self.replay_buffer, handle) # print("Pickle Saved") # print(10 + "n") self.params['eps'] = max(self.params['eps_final'], 1.00 - float(self.cnt) / float(self.params['eps_step'])) def train(self): """Train different agents: WDQN, DQN and Linear""" if self.local_cnt > self.params['train_start']: if self.params["only_dqn"]: batch_s_dqn, batch_a, batch_t, qt_dqn, batch_r, batch_idxes, weights = extract_batches_per(self.params, self.tnet, self.replay_buffer, self.beta_schedule, (self.numeps-self.model_eps)) self.cnt, td_errors = self.qnet.train(batch_s_dqn, None, batch_a, batch_t, qt_dqn, None, None, batch_r, self.params["dropout"], self.params["only_dqn"], self.params["only_lin"], weights) elif self.params["only_lin"]: batch_s_lin, batch_a, batch_t, qt_lin, batch_r, batch_idxes, weights = extract_batches_per(self.params, self.tnet, self.replay_buffer, self.beta_schedule, (self.numeps-self.model_eps)) self.cnt, td_errors = self.qnet.train(None, batch_s_lin, batch_a, batch_t, None, qt_lin, None, batch_r, self.params["dropout"], self.params["only_dqn"], self.params["only_lin"], weights) else: batch_s_dqn, batch_s_lin, batch_a, batch_t, qt_lin, qt_dqn, qt_wdqn, batch_r, batch_idxes, weights = extract_batches_per( self.params, self.tnet, self.replay_buffer, self.beta_schedule, (self.numeps-self.model_eps)) self.cnt, td_errors = self.qnet.train(batch_s_dqn, batch_s_lin, batch_a, batch_t, qt_dqn, qt_lin, qt_wdqn, batch_r, self.params["dropout"], self.params["only_dqn"], self.params["only_lin"], weights) if self.params["prioritized"]: new_priorities = np.abs(td_errors) + self.params["prioritized_replay_eps"] self.replay_buffer.update_priorities(batch_idxes, new_priorities) if self.params["ranked"] and self.cnt % self.rank_sort == 0: self.replay_buffer.sort() def final(self, state): """Inspects the last state""" # Do observation self.terminal = True self.observation_step(state) NUM_EPS_UPDATE = 100 self.lastWindowAccumRewards += state.getScore() # self.accumTrainRewards += state.getScore() self.Q_accumulative += max(self.Q_global, default=float('nan')) self.wins += self.won if self.numeps % NUM_EPS_UPDATE == 0: # Print stats eps_time = time.time() - self.episodeStartTime print('Reinforcement Learning Status:') if self.numeps <= self.params['num_training']: trainAvg = self.accumTrainRewards / float(self.numeps) print('\tCompleted %d out of %d training episodes' % ( self.numeps, self.params['num_training'])) print('\tAverage Rewards over all training: %.2f' % ( trainAvg)) windowAvg = self.lastWindowAccumRewards / float(NUM_EPS_UPDATE) windowQavg = self.Q_accumulative / float(NUM_EPS_UPDATE) window_steps = (self.cnt - self.last_steps) / float(NUM_EPS_UPDATE) print('\tAverage Rewards for last %d episodes: %.2f' % ( NUM_EPS_UPDATE, windowAvg)) print('\tEpisode took %.2f seconds' % (eps_time)) print('\tEpisilon is %.8f' % self.params["eps"]) print('\tLinear Decay learning Rate is %.8f' % self.sess.run(self.qnet.lr_lin)) if self.params["save_logs"]: log_file = open('logs/' + self.params["save_file"] + "-" + str(self.general_record_time) + '-l-' + str( (self.params["num_training"])) + '.log', 'a') log_file.write("# %4d | s: %8d | t: %.2f | r: %12f | Q: %10f | won: %r \n" % (self.numeps, window_steps, eps_time, windowAvg, windowQavg, self.wins)) # Save Best Model if windowAvg >= self.params["best_thr"]: self.params["best_thr"] = windowAvg self.last_steps = self.cnt self.save_mod(best_mod=True) print("Saving the model with:", self.params["best_thr"]) self.params["best_thr"] = self.params["best_thr"] + self.best_int sys.stdout.flush() self.lastWindowAccumRewards = 0 self.Q_accumulative = 0 self.last_steps = self.cnt self.wins = 0 self.episodeStartTime = time.time() if self.numeps >= self.params['num_training']: eps_time = time.time() - self.episodeStartTime if self.numeps == self.params['num_training']: print("Starting Date of Training:", self.general_record_time) print("Ending Date of Training:", time.strftime("%a_%d_%b_%Y_%H_%M_%S", time.localtime())) print("Training time duration in minutes:", (time.time() - self.start_time) / 60) print('Training Done (turning off epsilon)') self.params["eps"] = 0.0 # no exploration log_file = open( 'logs/testedModels/' + self.params["save_file"] + '-s-' + str(self.params["seed"]) + '-n-' + str( self.params['num_games'] - self.params["num_training"]) + '.log', 'a') log_file.write("# %4d | s: %8d | t: %.2f | r: %12f | Q: %10f | won: %r \n" % (self.numeps, self.cnt - self.last_steps, eps_time, state.getScore(), max(self.Q_global, default=float('nan')) , int(self.won))) self.last_steps = self.cnt # save model self.save_mod(best_mod=False) def isInTraining(self): """Check is if agent is in training""" return self.numeps < self.params["num_training"] def isInTesting(self): """Check is if agent is in testing""" return not self.isInTraining() def load_mod(self): """ Load data and model""" if self.params["load"]: try: print("".join([self.path_extra, "model/", self.params["save_file"], "-", self.params["load_file"]])) self.saver.restore(self.sess, "".join([self.path_extra, "model/", self.params["save_file"], "-", self.params["load_file"]])) if not self.params["load_file"].lower() == "best": print("Model Restored") else: print("Best Model Restored") try: load_path = "".join( [self.path_extra, "parameters/", "params_", self.params["save_file"], "-", self.params["load_file"].lower(), ".npy"]) # Parameters to be preserved for params when charging save, save_interval, num_tr, load_data, best_thr, eps_final, dropout, decay_lr, decay_lr_val, only_dqn, only_lin, load_file, num_games, seed, save_logs = \ self.params["save"], self.params["save_interval"], \ self.params["num_training"], self.params["load_data"], \ self.params["best_thr"], self.params["eps_final"], self.params["dropout"], self.params[ "dcy_lrl"], \ self.params["dcy_lrl_val"], self.params["only_dqn"], self.params["only_lin"], self.params[ "load_file"], self.params["num_games"], self.params["seed"], self.params["save_logs"] # Load saved parameters and hyperparameters of the new starting point self.last_steps, self.accumTrainRewards, self.numeps, self.params, self.local_cnt, self.cnt, self.sub_dir = np.load( load_path) # orig_num_training = self.params["num_training"] # Load newest Parameters to params orig_save_int = self.params["save_interval"] self.params["save"], self.params["save_interval"], self.params["num_training"], \ self.params["load_data"], self.params["best_thr"], self.params["eps_final"], self.params["dropout"], \ self.params["dcy_lrl"], self.params["dcy_lrl_val"], self.params["only_dqn"], self.params[ "only_lin"], self.params["load_file"], \ self.params["num_games"], self.params["seed"], self.params["save_logs"] = save, save_interval, \ num_tr, load_data, best_thr, eps_final, \ dropout, decay_lr, decay_lr_val, only_dqn, \ only_lin, load_file, num_games, seed, save_logs if self.sub_dir == "best": print("Best Parameters Restored") else: print("Parameters Restored") if self.params["load_data"]: # Load data and starts with correct data self.params["num_training"] += orig_num_training if not self.params["load_file"].lower() == "best": src = "".join( [self.path_extra, "data/mem_rep_", self.params["save_file"], "/", self.params["save_file"], "-", str(self.sub_dir), ".pickle"]) print("Interval Data to be Restored") else: src = "".join( [self.path_extra, "data/mem_rep_", self.params["save_file"], "/", self.params["save_file"], "-", self.sub_dir, ".pickle"]) print("Best Data to be Restored") with open(src, 'rb') as handle: self.replay_buffer = pickle.load(handle) print("Data Restored") except Exception as e: print(e) print("Parameters don't exist or could not be properly loaded") except: print("Model don't exist or could not be properly loaded") self.params["load"] = False def save_mod(self, best_mod=False): """ Saving model and parameters Possibility of saving the best model """ if (self.numeps % self.params["save_interval"] == 0 and self.params["save"]) or ( best_mod and self.params["save"]): self.params["global_step"] = self.cnt save_files = [self.last_steps, self.accumTrainRewards, self.numeps, self.params, self.local_cnt, self.cnt] try: if best_mod: self.saver.save(self.sess, "".join([self.path_extra, "model/", self.params["save_file"], "-", "best"])) print("Best Model Saved") elif not self.sub_dir == "best": self.saver.save(self.sess, "".join( [self.path_extra, "model/", self.params["save_file"], "-", str(self.numeps)])) print("Model Saved") except Exception as e: print("Model could not be saved") print("Error", e) try: if str(self.numeps) == self.sub_dir: # Save memory replay and parameters dic = "".join( [self.path_extra, "data/mem_rep_", self.params["save_file"], "/"]) f_name = "".join([self.params["save_file"], "-", str(self.sub_dir), ".pickle"]) # 1 save_files.append(self.sub_dir) np.save("".join( [self.path_extra, "parameters/", "params_", self.params["save_file"], "-", str(self.numeps)]), save_files) print("Pameters Saved") save_rep_buf(self.replay_buffer, dic, f_name) self.sub_dir = str(self.numeps + self.params["save_interval"]) print("Memory Replay Saved") elif best_mod: # Save memory replay of best model in directory "best" and parameters dic = "".join( [self.path_extra, "data/mem_rep_", self.params["save_file"], "/"]) f_name = "".join([self.params["save_file"], "-", "best", ".pickle"]) save_files.append("best") np.save("".join([self.path_extra, "parameters/", "params_", self.params["save_file"], "-", "best"]), save_files) print("Best Pameters Saved") save_rep_buf(self.replay_buffer, dic, f_name) print("Best Memory Replay Saved") except Exception as e: print("Parameters could not be saved") print("Error", e)