info["episode"]["r"], global_step) break # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation` real_next_obs = next_obs.copy() for idx, d in enumerate(dones): if d: real_next_obs[idx] = infos[idx]["terminal_observation"] rb.add(obs, real_next_obs, actions, rewards, dones, infos) # TRY NOT TO MODIFY: CRUCIAL step easy to overlook obs = next_obs # ALGO LOGIC: training. if global_step > args.learning_starts: data = rb.sample(args.batch_size) with torch.no_grad(): clipped_noise = (torch.randn_like(torch.Tensor(actions[0])) * args.policy_noise).clamp( -args.noise_clip, args.noise_clip) next_state_actions = ( target_actor.forward(data.next_observations) + clipped_noise.to(device)).clamp( envs.single_action_space.low[0], envs.single_action_space.high[0]) qf1_next_target = qf1_target.forward(data.next_observations, next_state_actions) qf2_next_target = qf2_target.forward(data.next_observations, next_state_actions) min_qf_next_target = torch.min(qf1_next_target,
class AWAC(OffPolicyAlgorithm): """ Soft Actor-Critic (SAC) Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor, This implementation borrows code from original implementation (https://github.com/haarnoja/sac) from OpenAI Spinning Up (https://github.com/openai/spinningup), from the softlearning repo (https://github.com/rail-berkeley/softlearning/) and from Stable Baselines (https://github.com/hill-a/stable-baselines) Paper: https://arxiv.org/abs/1801.01290 Introduction to SAC: https://spinningup.openai.com/en/latest/algorithms/sac.html Note: we use double q target and not value target as discussed in https://github.com/hill-a/stable-baselines/issues/270 :param policy: (SACPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, ...) :param env: (GymEnv or str) The environment to learn from (if registered in Gym, can be str) :param learning_rate: (float or callable) learning rate for adam optimizer, the same learning rate will be used for all networks (Q-Values, Actor and Value function) it can be a function of the current progress remaining (from 1 to 0) :param buffer_size: (int) size of the replay buffer :param learning_starts: (int) how many steps of the model to collect transitions for before learning starts :param batch_size: (int) Minibatch size for each gradient update :param tau: (float) the soft update coefficient ("Polyak update", between 0 and 1) :param gamma: (float) the discount factor :param train_freq: (int) Update the model every ``train_freq`` steps. :param gradient_steps: (int) How many gradient update after each step :param n_episodes_rollout: (int) Update the model every ``n_episodes_rollout`` episodes. Note that this cannot be used at the same time as ``train_freq`` :param action_noise: (ActionNoise) the action noise type (None by default), this can help for hard exploration problem. Cf common.noise for the different action noise type. :param ent_coef: (str or float) Entropy regularization coefficient. (Equivalent to inverse of reward scale in the original SAC paper.) Controlling exploration/exploitation trade-off. Set it to 'auto' to learn it automatically (and 'auto_0.1' for using 0.1 as initial value) :param target_update_interval: (int) update the target network every ``target_network_update_freq`` steps. :param target_entropy: (str or float) target entropy when learning ``ent_coef`` (``ent_coef = 'auto'``) :param create_eval_env: (bool) Whether to create a second environment that will be used for evaluating the agent periodically. (Only available when passing string for the environment) :param policy_kwargs: (dict) additional arguments to be passed to the policy on creation :param verbose: (int) the verbosity level: 0 no output, 1 info, 2 debug :param seed: (int) Seed for the pseudo random generators :param device: (str or th.device) Device (cpu, cuda, ...) on which the code should be run. Setting it to auto, the code will be run on the GPU if possible. :param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance """ def __init__(self, policy: Union[str, Type[AWACPolicy]], env: Union[GymEnv, str], learning_rate: Union[float, Callable] = 3e-4, buffer_size: int = int(1e6), learning_starts: int = 100, batch_size: int = 256, tau: float = 0.005, gamma: float = 0.99, train_freq: int = 1, gradient_steps: int = 1, n_episodes_rollout: int = -1, action_noise: Optional[ActionNoise] = None, ent_coef: Union[str, float] = 'auto', target_update_interval: int = 1, target_entropy: Union[str, float] = 'auto', awr_use_mle_for_vf: bool = True, beta: int = 50, tensorboard_log: Optional[str] = None, create_eval_env: bool = False, policy_kwargs: Dict[str, Any] = None, verbose: int = 0, seed: Optional[int] = None, device: Union[th.device, str] = 'auto', _init_setup_model: bool = True): super().__init__(policy, env, AWACPolicy, learning_rate, buffer_size, learning_starts, batch_size, policy_kwargs, tensorboard_log, verbose, device, create_eval_env=create_eval_env, seed=seed, use_sde=False, sde_sample_freq=-1, use_sde_at_warmup=False) self.target_entropy = target_entropy self.log_ent_coef = None # type: Optional[th.Tensor] self.target_update_interval = target_update_interval self.tau = tau # Entropy coefficient / Entropy temperature # Inverse of the reward scale self.ent_coef = ent_coef self.target_update_interval = target_update_interval self.train_freq = train_freq self.gradient_steps = gradient_steps self.n_episodes_rollout = n_episodes_rollout self.action_noise = action_noise self.gamma = gamma self.ent_coef_optimizer = None self.awr_use_mle_for_vf = awr_use_mle_for_vf self.beta = beta self.bc_buffer = None if _init_setup_model: self._setup_model() def _setup_model(self) -> None: super()._setup_model() self._create_aliases() # Target entropy is used when learning the entropy coefficient if self.target_entropy == 'auto': # automatically set target entropy if needed self.target_entropy = -np.prod(self.env.action_space.shape).astype( np.float32) else: # Force conversion # this will also throw an error for unexpected string self.target_entropy = float(self.target_entropy) # The entropy coefficient or entropy can be learned automatically # see Automating Entropy Adjustment for Maximum Entropy RL section # of https://arxiv.org/abs/1812.05905 if isinstance(self.ent_coef, str) and self.ent_coef.startswith('auto'): # Default initial value of ent_coef when learned init_value = 1.0 if '_' in self.ent_coef: init_value = float(self.ent_coef.split('_')[1]) assert init_value > 0., "The initial value of ent_coef must be greater than 0" # Note: we optimize the log of the entropy coeff which is slightly different from the paper # as discussed in https://github.com/rail-berkeley/softlearning/issues/37 self.log_ent_coef = th.log( th.ones(1, device=self.device) * init_value).requires_grad_(True) self.ent_coef_optimizer = th.optim.Adam([self.log_ent_coef], lr=self.lr_schedule(1)) else: # Force conversion to float # this will throw an error if a malformed string (different from 'auto') # is passed self.ent_coef_tensor = th.tensor(float(self.ent_coef)).to( self.device) self.bc_buffer = ReplayBuffer(int(1e4), self.observation_space, self.action_space, self.device) def _create_aliases(self) -> None: self.actor = self.policy.actor self.critic = self.policy.critic self.critic_target = self.policy.critic_target def pretrain_bc(self, gradient_steps: int, batch_size: int = 64): statistics = [] with trange(gradient_steps) as t: for gradient_step in t: replay_data = self.bc_buffer.sample( batch_size, env=self._vec_normalize_env) dist = self.actor(replay_data.observations) actions_pi, log_prob = dist.log_prob_and_rsample() actor_loss = -log_prob.mean() actor_mse_loss = F.mse_loss(actions_pi.detach(), replay_data.actions) self.actor.optimizer.zero_grad() actor_loss.backward() self.actor.optimizer.step() statistics.append((actor_loss.item(), actor_mse_loss.item())) t.set_postfix(mse_loss=actor_mse_loss.item(), policy_loss=actor_loss.item()) actor_losses, mse_losses = tuple(zip(*statistics)) logger.record("pretrain/n_updates", self._n_updates, exclude='tensorboard') logger.record("pretrain/actor_loss", np.mean(actor_losses)) logger.record("pretrain/actor_mse_loss", np.mean(mse_losses)) def pretrain_rl(self, gradient_steps: int, batch_size: int = 64) -> None: statistics = [] with trange(gradient_steps) as t: for gradient_step in t: replay_data = self.replay_buffer.sample( batch_size, env=self._vec_normalize_env) stats = self.train_batch(replay_data) statistics.append(stats) self._n_updates += 1 t.set_postfix(qf_loss=stats[1], policy_loss=stats[0]) actor_losses, critic_losses, ent_coef_losses, ent_coefs = tuple( zip(*statistics)) logger.record("pretrain/n_updates", self._n_updates, exclude='tensorboard') logger.record("pretrain/ent_coef", np.mean(ent_coefs)) logger.record("pretrain/actor_loss", np.mean(actor_losses)) logger.record("pretrain/critic_loss", np.mean(critic_losses)) logger.record("pretrain/ent_coef_loss", np.mean(ent_coef_losses)) def train(self, gradient_steps: int, batch_size: int = 64) -> None: statistics = [] for gradient_step in range(gradient_steps): replay_data = self.replay_buffer.sample( batch_size, env=self._vec_normalize_env) stats = self.train_batch(replay_data) statistics.append(stats) self._n_updates += 1 actor_losses, critic_losses, ent_coef_losses, ent_coefs = tuple( zip(*statistics)) logger.record("train/n_updates", self._n_updates, exclude='tensorboard') logger.record("train/ent_coef", np.mean(ent_coefs)) logger.record("train/actor_loss", np.mean(actor_losses)) logger.record("train/critic_loss", np.mean(critic_losses)) logger.record("train/ent_coef_loss", np.mean(ent_coef_losses)) def train_batch(self, replay_data): # Action by the current actor for the sampled state dist = self.actor(replay_data.observations) actions_pi, log_prob = dist.log_prob_and_rsample() actor_mle = dist.mean """ent_coeff loss""" ent_coef_loss = None if self.ent_coef_optimizer is not None: # Important: detach the variable from the graph # so we don't change it with other losses # see https://github.com/rail-berkeley/softlearning/issues/60 ent_coef = th.exp(self.log_ent_coef.detach()) ent_coef_loss = -( self.log_ent_coef * (log_prob + self.target_entropy).detach()).mean() else: ent_coef = self.ent_coef_tensor """q loss""" with th.no_grad(): # Select action according to policy next_dist = self.actor(replay_data.next_observations) next_actions, next_log_prob = next_dist.log_prob_and_rsample() # Compute the target Q value target_q1, target_q2 = self.critic_target( replay_data.next_observations, next_actions) target_q = th.min(target_q1, target_q2) target_q = replay_data.rewards + ( 1 - replay_data.dones) * self.gamma * target_q # td error + entropy term # q_backup = target_q - ent_coef * next_log_prob q_backup = target_q # Get current Q estimates # using action from the replay buffer current_q1, current_q2 = self.critic(replay_data.observations, replay_data.actions) # Compute critic loss critic_loss = 0.5 * (F.mse_loss(current_q1, q_backup) + F.mse_loss(current_q2, q_backup)) """action loss""" # Advantage-weighted regression # if self.awr_use_mle_for_vf: # v1_pi,v2_pi = self.critic(replay_data.observations, actor_mle) # v_pi = th.min(v1_pi, v2_pi) # else: # v1_pi,v2_pi = self.critic(replay_data.observations, actions_pi) # v_pi = th.min(v1_pi, v2_pi) q_adv = th.min(current_q1, current_q2) v1_pi, v2_pi = self.critic(replay_data.observations, actor_mle) v_pi = th.min(v1_pi, v2_pi) # q_adv = th.min(*self.critic(replay_data.observations, actions_pi)) score = q_adv - v_pi weights = F.softmax(score / self.beta, dim=0) # actor_loss = ent_coef * log_prob.mean() actor_logpp = dist.log_prob(replay_data.actions) actor_loss = (-actor_logpp * len(weights) * weights.detach()).mean() """Updates""" # Optimize entropy coefficient, also called # entropy temperature or alpha in the paper if ent_coef_loss is not None: self.ent_coef_optimizer.zero_grad() ent_coef_loss.backward() self.ent_coef_optimizer.step() # Optimize the critic self.critic.optimizer.zero_grad() critic_loss.backward() self.critic.optimizer.step() # Optimize the actor self.actor.optimizer.zero_grad() actor_loss.backward() self.actor.optimizer.step() # Update target networks if self._n_updates % self.target_update_interval == 0: for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) if ent_coef_loss is None: ent_coef_loss = th.tensor([0]) return actor_loss.item(), critic_loss.item(), ent_coef_loss.item( ), ent_coef.item() def learn(self, total_timesteps: int, callback: MaybeCallback = None, log_interval: int = 4, eval_env: Optional[GymEnv] = None, eval_freq: int = -1, n_eval_episodes: int = 5, tb_log_name: str = "AWAC", eval_log_path: Optional[str] = None, reset_num_timesteps: bool = True) -> OffPolicyAlgorithm: total_timesteps, callback = self._setup_learn( total_timesteps, eval_env, callback, eval_freq, n_eval_episodes, eval_log_path, reset_num_timesteps, tb_log_name) callback.on_training_start(locals(), globals()) self.pretrain_bc(int(1e3), batch_size=self.batch_size) observations, actions, next_observations, rewards, dones = self.bc_buffer.observations, self.bc_buffer.actions, self.bc_buffer.next_observations, self.bc_buffer.rewards, self.bc_buffer.dones for data in zip(observations, next_observations, actions, rewards, dones): self.replay_buffer.add(*data) self.pretrain_rl(int(1e4), batch_size=self.batch_size) while self.num_timesteps < total_timesteps: rollout = self.collect_rollouts( self.env, n_episodes=self.n_episodes_rollout, n_steps=self.train_freq, action_noise=self.action_noise, callback=callback, learning_starts=self.learning_starts, replay_buffer=self.replay_buffer, log_interval=log_interval) if rollout.continue_training is False: break self._update_current_progress_remaining(self.num_timesteps, total_timesteps) if self.num_timesteps > 0 and self.num_timesteps > self.learning_starts: gradient_steps = self.gradient_steps if self.gradient_steps > 0 else rollout.episode_timesteps self.train(gradient_steps, batch_size=self.batch_size) callback.on_training_end() return self def excluded_save_params(self) -> List[str]: """ Returns the names of the parameters that should be excluded by default when saving the model. :return: (List[str]) List of parameters that should be excluded from save """ # Exclude aliases return super().excluded_save_params() + [ "actor", "critic", "critic_target" ] def get_torch_variables(self) -> Tuple[List[str], List[str]]: """ cf base class """ state_dicts = ["policy", "actor.optimizer", "critic.optimizer"] saved_tensors = ['log_ent_coef'] if self.ent_coef_optimizer is not None: state_dicts.append('ent_coef_optimizer') else: saved_tensors.append('ent_coef_tensor') return state_dicts, saved_tensors
class CMVCVaRSAC(OffPolicyAlgorithm): def __init__( self, policy: Union[str, Type[CMVC51SACPolicy]], env: Union[GymEnv, str], min_v: float = -25, max_v: float = +25, support_dim: int = 200, learning_rate: Union[float, Schedule] = 3e-4, buffer_size: int = int(5e4), learning_starts: int = 100, batch_size: int = 64, tau: float = 0.005, gamma: float = 0.99, train_freq: Union[int, Tuple[int, str]] = 1, gradient_steps: int = 1, action_noise: Optional[ActionNoise] = None, optimize_memory_usage: bool = False, ent_coef: Union[str, float] = "auto", target_update_interval: int = 1, target_entropy: Union[str, float] = "auto", use_sde: bool = False, sde_sample_freq: int = -1, use_sde_at_warmup: bool = False, tensorboard_log: Optional[str] = None, create_eval_env: bool = False, policy_kwargs: Dict[str, Any] = None, verbose: int = 0, seed: Optional[int] = None, device: Union[th.device, str] = "auto", _init_setup_model: bool = True, cvar_alpha=0.3, cmv_beta=1, ): super(CMVCVaRSAC, self).__init__( policy, env, CMVC51SACPolicy, learning_rate, buffer_size, learning_starts, batch_size, tau, gamma, train_freq, gradient_steps, action_noise, policy_kwargs=policy_kwargs, tensorboard_log=tensorboard_log, verbose=verbose, device=device, create_eval_env=create_eval_env, seed=seed, use_sde=use_sde, sde_sample_freq=sde_sample_freq, use_sde_at_warmup=use_sde_at_warmup, optimize_memory_usage=optimize_memory_usage, supported_action_spaces=(gym.spaces.Box), ) self.target_entropy = target_entropy self.log_ent_coef = None # type: Optional[th.Tensor] # Entropy coefficient / Entropy temperature # Inverse of the reward scale self.ent_coef = ent_coef self.target_update_interval = target_update_interval self.ent_coef_optimizer = None self.min_v = min_v self.max_v = max_v self.support_dim = support_dim self.interval = (1 / (support_dim - 1)) * (max_v - min_v) self.supports = th.from_numpy( np.array([min_v + i * self.interval for i in range(support_dim)], dtype=np.float32)).to(self.device) self._total_timesteps = None self.cvar_alpha = cvar_alpha self.cmv_beta = cmv_beta if _init_setup_model: self._setup_model() def _setup_model(self) -> None: self._setup_lr_schedule() self.set_random_seed(self.seed) self.replay_buffer = ReplayBuffer( self.buffer_size, self.observation_space, self.action_space, self.device, optimize_memory_usage=self.optimize_memory_usage, ) self.policy = self.policy_class( self.observation_space, self.action_space, self.support_dim, self.lr_schedule, **self.policy_kwargs, # pytype:disable=not-instantiable ) self.policy = self.policy.to(self.device) # Convert train freq parameter to TrainFreq object self._convert_train_freq() self._create_aliases() # Target entropy is used when learning the entropy coefficient if self.target_entropy == "auto": # automatically set target entropy if needed self.target_entropy = -np.prod(self.env.action_space.shape).astype( np.float32) else: # Force conversion # this will also throw an error for unexpected string self.target_entropy = float(self.target_entropy) # The entropy coefficient or entropy can be learned automatically # see Automating Entropy Adjustment for Maximum Entropy RL section # of https://arxiv.org/abs/1812.05905 if isinstance(self.ent_coef, str) and self.ent_coef.startswith("auto"): # Default initial value of ent_coef when learned init_value = 1.0 if "_" in self.ent_coef: init_value = float(self.ent_coef.split("_")[1]) assert init_value > 0.0, "The initial value of ent_coef must be greater than 0" # Note: we optimize the log of the entropy coeff which is slightly different from the paper # as discussed in https://github.com/rail-berkeley/softlearning/issues/37 self.log_ent_coef = th.log( th.ones(1, device=self.device) * init_value).requires_grad_(True) self.ent_coef_optimizer = th.optim.Adam([self.log_ent_coef], lr=self.lr_schedule(1)) else: # Force conversion to float # this will throw an error if a malformed string (different from 'auto') # is passed self.ent_coef_tensor = th.tensor(float(self.ent_coef)).to( self.device) def _create_aliases(self) -> None: self.actor = self.policy.actor self.critic = self.policy.critic self.critic_target = self.policy.critic_target self.cmv_net = self.policy.cmv_net self.beta_critic = self.policy.beta_critic self.beta_critic_target = self.policy.beta_critic_target def projection(self, support_rows, target_z): projected_target_z = th.zeros_like(target_z) support_rows = support_rows.clamp(self.min_v, self.max_v - 1e-3) p = ((support_rows - self.min_v) % self.interval) / self.interval idx = ((support_rows - self.min_v) // self.interval).long() projected_target_z = projected_target_z.scatter_add( 1, idx, target_z * p) projected_target_z = projected_target_z.scatter_add( 1, idx + 1, target_z * (1 - p)) return projected_target_z def train(self, gradient_steps: int, batch_size: int = 64) -> None: # Update optimizers learning rate optimizers = [self.actor.optimizer, self.critic.optimizer] if self.ent_coef_optimizer is not None: optimizers += [self.ent_coef_optimizer] # Update learning rate according to lr schedule self._update_learning_rate(optimizers) ent_coef_losses, ent_coefs = [], [] actor_losses, critic_losses, reward_losses, feature_pred_losses = [], [], [], [] cvars = [] qs = [] critic_beta_losses = [] for gradient_step in range(gradient_steps): # Sample replay buffer replay_data = self.replay_buffer.sample( batch_size, env=self._vec_normalize_env) # We need to sample because `log_std` may have changed between two gradient steps if self.use_sde: self.actor.reset_noise() # Action by the current actor for the sampled state actions_pi, log_prob = self.actor.action_log_prob( replay_data.observations) log_prob = log_prob.reshape(-1, 1) ent_coef_loss = None if self.ent_coef_optimizer is not None: # Important: detach the variable from the graph # so we don't change it with other losses # see https://github.com/rail-berkeley/softlearning/issues/60 ent_coef = th.exp(self.log_ent_coef.detach()) ent_coef_loss = -( self.log_ent_coef * (log_prob + self.target_entropy).detach()).mean() ent_coef_losses.append(ent_coef_loss.item()) else: ent_coef = self.ent_coef_tensor ent_coefs.append(ent_coef.item()) # Optimize entropy coefficient, also called # entropy temperature or alpha in the paper if ent_coef_loss is not None: self.ent_coef_optimizer.zero_grad() ent_coef_loss.backward() self.ent_coef_optimizer.step() with th.no_grad(): # Select action according to policy next_actions, next_log_prob = self.actor.action_log_prob( replay_data.next_observations) # Compute the next Q values: min over all critics targets target_zs = th.cat(self.critic_target( replay_data.next_observations, next_actions), dim=1) # add entropy term target_supports = self.supports.clone().detach() target_supports = target_supports - ent_coef * next_log_prob.reshape( -1, 1) # td error + entropy term target_supports = replay_data.rewards + ( 1 - replay_data.dones) * self.gamma * target_supports # Get current Q-values estimates for each critic network # using action from the replay buffer current_zs = th.cat(self.critic(replay_data.observations, replay_data.actions), dim=1) target_zs = self.projection(target_supports, target_zs) # Compute critic loss critic_loss = -th.mean(th.log(current_zs + 1e-12) * target_zs) critic_losses.append(critic_loss.item()) # Optimize the critic self.critic.optimizer.zero_grad() critic_loss.backward() self.critic.optimizer.step() # Compute # For the CMV Learning with th.no_grad(): next_z = self.critic_target.features_extractor( replay_data.next_observations) # predicted reward, and predicted next observation r_pred, z_pred = self.cmv_net(replay_data.observations, replay_data.actions) mse_r_pred = th.mean(th.square(r_pred - replay_data.rewards)) mse_z_pred = th.mean(th.square(z_pred - next_z)) loss_cmv = mse_r_pred + mse_z_pred reward_losses.append(mse_r_pred.item()) feature_pred_losses.append(mse_z_pred.item()) # Optimize the CMV Nets self.cmv_net.optimizer.zero_grad() loss_cmv.backward() self.cmv_net.optimizer.step() with th.no_grad(): # Select action according to policy # Compute the next Q values: min over all critics targets next_q_beta_values = th.cat(self.beta_critic_target( replay_data.next_observations, next_actions), dim=1) next_q_beta_values, _ = th.min(next_q_beta_values, dim=1, keepdim=True) # add entropy term next_q_beta_values = next_q_beta_values - ent_coef * next_log_prob.reshape( -1, 1) # td error + entropy term target_q_beta_values = loss_cmv.detach() + ( 1 - replay_data.dones) * (self.gamma** 2) * next_q_beta_values # Get current Q-beta values estimates for each critic network # using action from the replay buffer current_q_beta_values = self.beta_critic(replay_data.observations, replay_data.actions) # Compute critic beta loss critic_beta_loss = \ 0.5 * sum( [F.mse_loss(current_q_beta, target_q_beta_values) for current_q_beta in current_q_beta_values]) critic_beta_losses.append(critic_beta_loss.item()) # Optimize the critic beta self.beta_critic.optimizer.zero_grad() critic_beta_loss.backward() self.beta_critic.optimizer.step() # Compute actor loss # Alternative: actor_loss = th.mean(log_prob - qf1_pi - qf1_beta_pi) # Mean over all critic networks z_pi = th.cat(self.critic.forward(replay_data.observations, actions_pi), dim=1) z_cdf = th.cumsum(z_pi, dim=-1) adjust_pdf = th.where(th.le(z_cdf, self.cvar_alpha), z_pi, th.zeros_like(z_pi)) adjust_pdf = th.div(adjust_pdf, th.sum(adjust_pdf, dim=-1, keepdim=True)) q_pi = adjust_pdf @ self.supports cvars.append(th.mean(q_pi).item()) qs.append(th.mean(z_pi @ self.supports).item()) q_beta_values_pi = th.cat(self.beta_critic.forward( replay_data.observations, actions_pi), dim=1) max_qf_beta_pi, _ = th.max(q_beta_values_pi, dim=1, keepdim=True) actor_loss = (ent_coef * log_prob - q_pi + self.cmv_beta * next_q_beta_values).mean() actor_losses.append(actor_loss.item()) # Optimize the actor self.actor.optimizer.zero_grad() actor_loss.backward() self.actor.optimizer.step() # Update target networks if gradient_step % self.target_update_interval == 0: polyak_update(self.critic.parameters(), self.critic_target.parameters(), self.tau) polyak_update(self.beta_critic.parameters(), self.beta_critic_target.parameters(), self.tau) self._n_updates += gradient_steps fps = int(self.num_timesteps / (time.time() - self.start_time)) remaining_steps = self._total_timesteps - self.num_timesteps eta = int(round(remaining_steps / fps)) logger.record("time/eta", timedelta(seconds=eta), exclude="tensorboard") logger.record("train/CVaR Alpha", self.cvar_alpha) logger.record("train/CMV Beta", self.cmv_beta) logger.record("train/CVaR", np.mean(cvars)) logger.record("train/Q-value", np.mean(qs)) logger.record("train/n_updates", self._n_updates, exclude="tensorboard") logger.record("train/ent_coef", np.mean(ent_coefs)) logger.record("train/actor_loss", np.mean(actor_losses)) logger.record("train/critic_loss", np.mean(critic_losses)) logger.record("train/reward error", np.mean(reward_losses)) logger.record("train/s_t+1_error", np.mean(feature_pred_losses)) logger.record("train/beta_Q_loss", np.mean(critic_beta_losses)) if len(ent_coef_losses) > 0: logger.record("train/ent_coef_loss", np.mean(ent_coef_losses)) def learn( self, total_timesteps: int, callback: MaybeCallback = None, log_interval: int = 4, eval_env: Optional[GymEnv] = None, eval_freq: int = -1, n_eval_episodes: int = 5, tb_log_name: str = "C51SAC", eval_log_path: Optional[str] = None, reset_num_timesteps: bool = True, ) -> OffPolicyAlgorithm: return super(CMVCVaRSAC, self).learn( total_timesteps=total_timesteps, callback=callback, log_interval=log_interval, eval_env=eval_env, eval_freq=eval_freq, n_eval_episodes=n_eval_episodes, tb_log_name=tb_log_name, eval_log_path=eval_log_path, reset_num_timesteps=reset_num_timesteps, ) def _excluded_save_params(self) -> List[str]: return super(CMVCVaRSAC, self)._excluded_save_params() + [ "actor", "critic", "critic_target" ] def _get_torch_save_params(self) -> Tuple[List[str], List[str]]: state_dicts = ["policy", "actor.optimizer", "critic.optimizer"] saved_pytorch_variables = ["log_ent_coef"] if self.ent_coef_optimizer is not None: state_dicts.append("ent_coef_optimizer") else: saved_pytorch_variables.append("ent_coef_tensor") return state_dicts, saved_pytorch_variables