diff --git a/docs/misc/changelog.rst b/docs/misc/changelog.rst index 40a88bb..86c1e5b 100644 --- a/docs/misc/changelog.rst +++ b/docs/misc/changelog.rst @@ -3,12 +3,13 @@ Changelog ========== -Pre-Release 0.6.0a6 (WIP) +Pre-Release 0.6.0a7 (WIP) ------------------------------ Breaking Changes: ^^^^^^^^^^^^^^^^^ +- Remove State-Dependent Exploration (SDE) support for ``TD3`` New Features: ^^^^^^^^^^^^^ diff --git a/stable_baselines3/common/base_class.py b/stable_baselines3/common/base_class.py index 12320a8..6ab7477 100644 --- a/stable_baselines3/common/base_class.py +++ b/stable_baselines3/common/base_class.py @@ -686,6 +686,7 @@ class OffPolicyRLModel(BaseRLModel): Default: -1 (only sample at the beginning of the rollout) :param use_sde_at_warmup: (bool) Whether to use SDE instead of uniform sampling during the warm up phase (before learning starts) + :param sde_support: (bool) Whether the model support SDE or not """ def __init__(self, @@ -705,7 +706,8 @@ class OffPolicyRLModel(BaseRLModel): seed: Optional[int] = None, use_sde: bool = False, sde_sample_freq: int = -1, - use_sde_at_warmup: bool = False): + use_sde_at_warmup: bool = False, + sde_support: bool = True): super(OffPolicyRLModel, self).__init__(policy, env, policy_base, learning_rate, policy_kwargs, verbose, @@ -717,11 +719,10 @@ class OffPolicyRLModel(BaseRLModel): self.actor = None self.replay_buffer = None # type: Optional[ReplayBuffer] # Update policy keyword arguments - self.policy_kwargs['use_sde'] = self.use_sde + if sde_support: + self.policy_kwargs['use_sde'] = self.use_sde self.policy_kwargs['device'] = self.device # For SDE only - self.rollout_data = None - self.on_policy_exploration = False self.use_sde_at_warmup = use_sde_at_warmup def _setup_model(self): @@ -786,12 +787,8 @@ class OffPolicyRLModel(BaseRLModel): assert isinstance(env, VecEnv), "You must pass a VecEnv" assert env.num_envs == 1, "OffPolicyRLModel only support single environment" - self.rollout_data = None if self.use_sde: self.actor.reset_noise() - # Reset rollout data - if self.on_policy_exploration: - self.rollout_data = {key: [] for key in ['observations', 'actions', 'rewards', 'dones', 'values']} callback.on_rollout_start() continue_training = True @@ -816,23 +813,25 @@ class OffPolicyRLModel(BaseRLModel): unscaled_action, _ = self.predict(self._last_obs, deterministic=False) # Rescale the action from [low, high] to [-1, 1] - scaled_action = self.policy.scale_action(unscaled_action) + if isinstance(self.action_space, gym.spaces.Box): + scaled_action = self.policy.scale_action(unscaled_action) - if self.use_sde: - # When using SDE, the action can be out of bounds - # TODO: fix with squashing and account for that in the proba distribution - clipped_action = np.clip(scaled_action, -1, 1) + # Add noise to the action (improve exploration) + if action_noise is not None: + # NOTE: in the original implementation of TD3, the noise was applied to the unscaled action + # Update(October 2019): Not anymore + clipped_action = np.clip(scaled_action + action_noise(), -1, 1) + + # We store the scaled action in the buffer + buffer_action = clipped_action + action = self.policy.unscale_action(clipped_action) else: - clipped_action = scaled_action - - # Add noise to the action (improve exploration) - if action_noise is not None: - # NOTE: in the original implementation of TD3, the noise was applied to the unscaled action - # Update(October 2019): Not anymore - clipped_action = np.clip(clipped_action + action_noise(), -1, 1) + # Discrete case, no need to normalize or clip + buffer_action = unscaled_action + action = buffer_action # Rescale and perform action - new_obs, reward, done, infos = env.step(self.policy.unscale_action(clipped_action)) + new_obs, reward, done, infos = env.step(action) # Only stop training if return value is False, not when it is None. if callback.on_step() is False: @@ -853,16 +852,7 @@ class OffPolicyRLModel(BaseRLModel): # Avoid changing the original ones self._last_original_obs, new_obs_, reward_ = self._last_obs, new_obs, reward - replay_buffer.add(self._last_original_obs, new_obs_, clipped_action, reward_, done) - - if self.rollout_data is not None: - # Assume only one env - self.rollout_data['observations'].append(self._last_obs[0].copy()) - self.rollout_data['actions'].append(scaled_action[0].copy()) - self.rollout_data['rewards'].append(reward[0].copy()) - self.rollout_data['dones'].append(done[0].copy()) - obs_tensor = th.FloatTensor(self._last_obs).to(self.device) - self.rollout_data['values'].append(self.vf_net(obs_tensor)[0].cpu().detach().numpy()) + replay_buffer.add(self._last_original_obs, new_obs_, buffer_action, reward_, done) self._last_obs = new_obs # Save the unnormalized observation @@ -880,6 +870,7 @@ class OffPolicyRLModel(BaseRLModel): self._episode_num += 1 episode_rewards.append(episode_reward) total_timesteps.append(episode_timesteps) + if action_noise is not None: action_noise.reset() @@ -890,7 +881,6 @@ class OffPolicyRLModel(BaseRLModel): if len(self.ep_info_buffer) > 0 and len(self.ep_info_buffer[0]) > 0: logger.logkv('ep_rew_mean', self.safe_mean([ep_info['r'] for ep_info in self.ep_info_buffer])) logger.logkv('ep_len_mean', self.safe_mean([ep_info['l'] for ep_info in self.ep_info_buffer])) - # logger.logkv("n_updates", n_updates) logger.logkv("fps", fps) logger.logkv('time_elapsed', int(time.time() - self.start_time)) logger.logkv("total timesteps", self.num_timesteps) @@ -903,27 +893,6 @@ class OffPolicyRLModel(BaseRLModel): mean_reward = np.mean(episode_rewards) if total_episodes > 0 else 0.0 - # Post processing - if self.rollout_data is not None: - for key in ['observations', 'actions', 'rewards', 'dones', 'values']: - self.rollout_data[key] = th.FloatTensor(np.array(self.rollout_data[key])).to(self.device) - - self.rollout_data['returns'] = self.rollout_data['rewards'].clone() - self.rollout_data['advantage'] = self.rollout_data['rewards'].clone() - - # Compute return and advantage - last_return = 0.0 - for step in reversed(range(len(self.rollout_data['rewards']))): - if step == len(self.rollout_data['rewards']) - 1: - next_non_terminal = 1.0 - done[0] - next_value = self.vf_net(th.FloatTensor(self._last_obs).to(self.device))[0].detach() - last_return = self.rollout_data['rewards'][step] + next_non_terminal * next_value - else: - next_non_terminal = 1.0 - self.rollout_data['dones'][step + 1] - last_return = self.rollout_data['rewards'][step] + self.gamma * last_return * next_non_terminal - self.rollout_data['returns'][step] = last_return - self.rollout_data['advantage'] = self.rollout_data['returns'] - self.rollout_data['values'] - callback.on_rollout_end() return RolloutReturn(mean_reward, total_steps, total_episodes, continue_training) diff --git a/stable_baselines3/td3/policies.py b/stable_baselines3/td3/policies.py index 5965334..a90324a 100644 --- a/stable_baselines3/td3/policies.py +++ b/stable_baselines3/td3/policies.py @@ -6,9 +6,7 @@ import torch.nn as nn from stable_baselines3.common.preprocessing import get_action_dim from stable_baselines3.common.policies import (BasePolicy, register_policy, create_mlp, - create_sde_features_extractor, NatureCNN, - BaseFeaturesExtractor, FlattenExtractor) -from stable_baselines3.common.distributions import StateDependentNoiseDistribution + NatureCNN, BaseFeaturesExtractor, FlattenExtractor) class Actor(BasePolicy): @@ -22,18 +20,6 @@ class Actor(BasePolicy): (a CNN when using images, a nn.Flatten() layer otherwise) :param features_dim: (int) Number of features :param activation_fn: (Type[nn.Module]) Activation function - :param use_sde: (bool) Whether to use State Dependent Exploration or not - :param log_std_init: (float) Initial value for the log standard deviation - :param clip_noise: (float) Clip the magnitude of the noise - :param lr_sde: (float) Learning rate for the standard deviation of the noise - :param full_std: (bool) Whether to use (n_features x n_actions) parameters - for the std instead of only (n_features,) when using SDE. - :param sde_net_arch: ([int]) Network architecture for extracting features - when using SDE. If None, the latent features from the policy will be used. - Pass an empty list to use the states as features. - :param use_expln: (bool) Use ``expln()`` function instead of ``exp()`` when using SDE to ensure - a positive standard deviation (cf paper). It allows to keep variance - above zero and prevent it from growing too fast. In practice, ``exp()`` is usually enough. :param normalize_images: (bool) Whether to normalize images or not, dividing by 255.0 (True by default) :param device: (Union[th.device, str]) Device on which the code should run. @@ -45,65 +31,25 @@ class Actor(BasePolicy): features_extractor: nn.Module, features_dim: int, activation_fn: Type[nn.Module] = nn.ReLU, - use_sde: bool = False, - log_std_init: float = -3, - clip_noise: Optional[float] = None, - lr_sde: float = 3e-4, - full_std: bool = False, - sde_net_arch: Optional[List[int]] = None, - use_expln: bool = False, normalize_images: bool = True, device: Union[th.device, str] = 'auto'): super(Actor, self).__init__(observation_space, action_space, features_extractor=features_extractor, normalize_images=normalize_images, device=device, - squash_output=not use_sde) + squash_output=True) + - self.latent_pi, self.log_std = None, None - self.weights_dist, self.exploration_mat = None, None - self.use_sde, self.sde_optimizer = use_sde, None - self.full_std = full_std - self.sde_features_extractor = None self.features_extractor = features_extractor self.normalize_images = normalize_images self.net_arch = net_arch self.features_dim = features_dim self.activation_fn = activation_fn - self.clip_noise = clip_noise - self.lr_sde = lr_sde - self.log_std_init = log_std_init - self.sde_net_arch = sde_net_arch - self.use_expln = use_expln - self.full_std = full_std action_dim = get_action_dim(self.action_space) - - if use_sde: - latent_pi_net = create_mlp(features_dim, -1, net_arch, activation_fn, squash_output=False) - self.latent_pi = nn.Sequential(*latent_pi_net) - latent_sde_dim = net_arch[-1] - learn_features = sde_net_arch is not None - - # Separate feature extractor for SDE - if sde_net_arch is not None: - self.sde_features_extractor, latent_sde_dim = create_sde_features_extractor(features_dim, sde_net_arch, - activation_fn) - - # Create state dependent noise matrix (SDE) - self.action_dist = StateDependentNoiseDistribution(action_dim, full_std=full_std, use_expln=use_expln, - squash_output=False, learn_features=learn_features) - - action_net, self.log_std = self.action_dist.proba_distribution_net(latent_dim=net_arch[-1], - latent_sde_dim=latent_sde_dim, - log_std_init=log_std_init) - # Squash output - self.mu = nn.Sequential(action_net, nn.Tanh()) - self.sde_optimizer = th.optim.Adam([self.log_std], lr=lr_sde) - self.reset_noise() - else: - actor_net = create_mlp(features_dim, action_dim, net_arch, activation_fn, squash_output=True) - self.mu = nn.Sequential(*actor_net) + actor_net = create_mlp(features_dim, action_dim, net_arch, activation_fn, squash_output=True) + # Deterministic action + self.mu = nn.Sequential(*actor_net) def _get_data(self) -> Dict[str, Any]: data = super()._get_data() @@ -112,73 +58,14 @@ class Actor(BasePolicy): net_arch=self.net_arch, features_dim=self.features_dim, activation_fn=self.activation_fn, - use_sde=self.use_sde, - log_std_init=self.log_std_init, - clip_noise=self.clip_noise, - lr_sde=self.lr_sde, - full_std=self.full_std, - sde_net_arch=self.sde_net_arch, - use_expln=self.use_expln, features_extractor=self.features_extractor )) return data - def get_std(self) -> th.Tensor: - """ - Retrieve the standard deviation of the action distribution. - Only useful when using SDE. - It corresponds to ``th.exp(log_std)`` in the normal case, - but is slightly different when using ``expln`` function - (cf StateDependentNoiseDistribution doc). - - :return: (th.Tensor) - """ - return self.action_dist.get_std(self.log_std) - - def _get_latent(self, obs: th.Tensor) -> Tuple[th.Tensor, th.Tensor]: - features = self.extract_features(obs) - latent_pi = self.latent_pi(features) - latent_sde = self.sde_features_extractor(features) if self.sde_features_extractor is not None else latent_pi - return latent_pi, latent_sde - - def evaluate_actions(self, obs: th.Tensor, actions: th.Tensor) -> Tuple[th.Tensor, th.Tensor]: - """ - Evaluate actions according to the current policy, - given the observations. Only useful when using SDE. - - :param obs: (th.Tensor) - :param actions: (th.Tensor) - :return: (th.Tensor, th.Tensor) log likelihood of taking those actions - and entropy of the action distribution. - """ - latent_pi, latent_sde = self._get_latent(obs) - mean_actions = self.mu(latent_pi) - distribution = self.action_dist.proba_distribution(mean_actions, self.log_std, latent_sde) - log_prob = distribution.log_prob(actions) - return log_prob, distribution.entropy() - - def reset_noise(self) -> None: - """ - Sample new weights for the exploration matrix, when using SDE. - """ - self.action_dist.sample_weights(self.log_std) - def forward(self, obs: th.Tensor, deterministic: bool = True) -> th.Tensor: - if self.use_sde: - latent_pi, latent_sde = self._get_latent(obs) - if deterministic: - return self.mu(latent_pi) - - noise = self.action_dist.get_noise(latent_sde) - if self.clip_noise is not None: - noise = th.clamp(noise, -self.clip_noise, self.clip_noise) - # TODO: Replace with squashing -> need to account for that in the sde update - # -> set squash_output=True in the action_dist? - # NOTE: the clipping is done in the rollout for now - return self.mu(latent_pi) + noise - else: - features = self.extract_features(obs) - return self.mu(features) + # assert deterministic, 'The TD3 actor only outputs deterministic actions' + features = self.extract_features(obs) + return self.mu(features) def _predict(self, observation: th.Tensor, deterministic: bool = False) -> th.Tensor: return self.forward(observation, deterministic=deterministic) @@ -234,43 +121,6 @@ class Critic(BasePolicy): return self.q1_net(th.cat([features, actions], dim=1)) -class ValueFunction(BasePolicy): - """ - Value function for TD3 when doing on-policy exploration with SDE. - - :param observation_space: (gym.spaces.Space) Obervation space - :param action_space: (gym.spaces.Space) Action space - :param features_extractor: (nn.Module) Network to extract features - (a CNN when using images, a nn.Flatten() layer otherwise) - :param features_dim: (int) Number of features - :param net_arch: (Optional[List[int]]) Network architecture - :param activation_fn: (Type[nn.Module]) Activation function - :param normalize_images: (bool) Whether to normalize images or not, - dividing by 255.0 (True by default) - """ - def __init__(self, observation_space: gym.spaces.Space, - action_space: gym.spaces.Space, - features_extractor: nn.Module, - features_dim: int, - net_arch: Optional[List[int]] = None, - activation_fn: Type[nn.Module] = nn.Tanh, - normalize_images: bool = True): - super(ValueFunction, self).__init__(observation_space, action_space, - features_extractor=features_extractor, - normalize_images=normalize_images) - - if net_arch is None: - net_arch = [64, 64] - - vf_net = create_mlp(features_dim, 1, net_arch, activation_fn) - self.vf_net = nn.Sequential(*vf_net) - - def forward(self, obs: th.Tensor) -> th.Tensor: - with th.no_grad(): - features = self.extract_features(obs) - return self.vf_net(features) - - class TD3Policy(BasePolicy): """ Policy class (with both actor and critic) for TD3. @@ -281,14 +131,6 @@ class TD3Policy(BasePolicy): :param net_arch: (Optional[List[int]]) The specification of the policy and value networks. :param device: (Union[th.device, str]) Device on which the code should run. :param activation_fn: (Type[nn.Module]) Activation function - :param use_sde: (bool) Whether to use State Dependent Exploration or not - :param log_std_init: (float) Initial value for the log standard deviation - :param sde_net_arch: ([int]) Network architecture for extracting features - when using SDE. If None, the latent features from the policy will be used. - Pass an empty list to use the states as features. - :param use_expln: (bool) Use ``expln()`` function instead of ``exp()`` when using SDE to ensure - a positive standard deviation (cf paper). It allows to keep variance - above zero and prevent it from growing too fast. In practice, ``exp()`` is usually enough. :param features_extractor_class: (Type[BaseFeaturesExtractor]) Features extractor to use. :param features_extractor_kwargs: (Optional[Dict[str, Any]]) Keyword arguments to pass to the feature extractor. @@ -305,12 +147,6 @@ class TD3Policy(BasePolicy): net_arch: Optional[List[int]] = None, device: Union[th.device, str] = 'auto', activation_fn: Type[nn.Module] = nn.ReLU, - use_sde: bool = False, - log_std_init: float = -3, - clip_noise: Optional[float] = None, - lr_sde: float = 3e-4, - sde_net_arch: Optional[List[int]] = None, - use_expln: bool = False, features_extractor_class: Type[BaseFeaturesExtractor] = FlattenExtractor, features_extractor_kwargs: Optional[Dict[str, Any]] = None, normalize_images: bool = True, @@ -347,23 +183,9 @@ class TD3Policy(BasePolicy): 'normalize_images': normalize_images, 'device': device } - self.actor_kwargs = self.net_args.copy() - sde_kwargs = { - 'use_sde': use_sde, - 'log_std_init': log_std_init, - 'clip_noise': clip_noise, - 'lr_sde': lr_sde, - 'sde_net_arch': sde_net_arch, - 'use_expln': use_expln - } - self.actor_kwargs.update(sde_kwargs) - self.actor, self.actor_target = None, None self.critic, self.critic_target = None, None - # For SDE only - self.use_sde = use_sde - self.vf_net = None - self.log_std_init = log_std_init + self._build(lr_schedule) def _build(self, lr_schedule: Callable) -> None: @@ -377,11 +199,6 @@ class TD3Policy(BasePolicy): self.critic_target.load_state_dict(self.critic.state_dict()) self.critic.optimizer = self.optimizer_class(self.critic.parameters(), lr=lr_schedule(1), **self.optimizer_kwargs) - if self.use_sde: - self.vf_net = ValueFunction(self.observation_space, self.action_space, - features_extractor=self.features_extractor, - features_dim=self.features_dim) - self.actor.sde_optimizer.add_param_group({'params': self.vf_net.parameters()}) # pytype: disable=attribute-error def _get_data(self) -> Dict[str, Any]: data = super()._get_data() @@ -389,12 +206,6 @@ class TD3Policy(BasePolicy): data.update(dict( net_arch=self.net_args['net_arch'], activation_fn=self.net_args['activation_fn'], - use_sde=self.actor_kwargs['use_sde'], - log_std_init=self.actor_kwargs['log_std_init'], - clip_noise=self.actor_kwargs['clip_noise'], - lr_sde=self.actor_kwargs['lr_sde'], - sde_net_arch=self.actor_kwargs['sde_net_arch'], - use_expln=self.actor_kwargs['use_expln'], lr_schedule=self._dummy_schedule, # dummy lr schedule, not needed for loading policy alone optimizer_class=self.optimizer_class, optimizer_kwargs=self.optimizer_kwargs, @@ -403,11 +214,8 @@ class TD3Policy(BasePolicy): )) return data - def reset_noise(self) -> None: - return self.actor.reset_noise() - def make_actor(self) -> Actor: - return Actor(**self.actor_kwargs).to(self.device) + return Actor(**self.net_args).to(self.device) def make_critic(self) -> Critic: return Critic(**self.net_args).to(self.device) @@ -432,14 +240,6 @@ class CnnPolicy(TD3Policy): :param net_arch: (Optional[List[int]]) The specification of the policy and value networks. :param device: (Union[th.device, str]) Device on which the code should run. :param activation_fn: (Type[nn.Module]) Activation function - :param use_sde: (bool) Whether to use State Dependent Exploration or not - :param log_std_init: (float) Initial value for the log standard deviation - :param sde_net_arch: ([int]) Network architecture for extracting features - when using SDE. If None, the latent features from the policy will be used. - Pass an empty list to use the states as features. - :param use_expln: (bool) Use ``expln()`` function instead of ``exp()`` when using SDE to ensure - a positive standard deviation (cf paper). It allows to keep variance - above zero and prevent it from growing too fast. In practice, ``exp()`` is usually enough. :param features_extractor_class: (Type[BaseFeaturesExtractor]) Features extractor to use. :param features_extractor_kwargs: (Optional[Dict[str, Any]]) Keyword arguments to pass to the feature extractor. @@ -456,12 +256,6 @@ class CnnPolicy(TD3Policy): net_arch: Optional[List[int]] = None, device: Union[th.device, str] = 'auto', activation_fn: Type[nn.Module] = nn.ReLU, - use_sde: bool = False, - log_std_init: float = -3, - clip_noise: Optional[float] = None, - lr_sde: float = 3e-4, - sde_net_arch: Optional[List[int]] = None, - use_expln: bool = False, features_extractor_class: Type[BaseFeaturesExtractor] = NatureCNN, features_extractor_kwargs: Optional[Dict[str, Any]] = None, normalize_images: bool = True, @@ -473,12 +267,6 @@ class CnnPolicy(TD3Policy): net_arch, device, activation_fn, - use_sde, - log_std_init, - clip_noise, - lr_sde, - sde_net_arch, - use_expln, features_extractor_class, features_extractor_kwargs, normalize_images, diff --git a/stable_baselines3/td3/td3.py b/stable_baselines3/td3/td3.py index 015727d..ae9edce 100644 --- a/stable_baselines3/td3/td3.py +++ b/stable_baselines3/td3/td3.py @@ -39,15 +39,6 @@ class TD3(OffPolicyRLModel): :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 use_sde: (bool) Whether to use State Dependent Exploration (SDE) - instead of action noise exploration (default: False) - :param sde_sample_freq: (int) Sample a new noise matrix every n steps when using SDE - Default: -1 (only sample at the beginning of the rollout) - :param sde_max_grad_norm: (float) - :param sde_ent_coef: (float) - :param sde_log_std_scheduler: (callable) - :param use_sde_at_warmup: (bool) Whether to use SDE instead of uniform sampling - during the warm up phase (before learning starts) :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 @@ -73,12 +64,6 @@ class TD3(OffPolicyRLModel): policy_delay: int = 2, target_policy_noise: float = 0.2, target_noise_clip: float = 0.5, - use_sde: bool = False, - sde_sample_freq: int = -1, - sde_max_grad_norm: float = 1, - sde_ent_coef: float = 0.0, - sde_log_std_scheduler: Optional[Callable] = None, - use_sde_at_warmup: bool = False, tensorboard_log: Optional[str] = None, create_eval_env: bool = False, policy_kwargs: Dict[str, Any] = None, @@ -91,8 +76,7 @@ class TD3(OffPolicyRLModel): buffer_size, learning_starts, batch_size, policy_kwargs, verbose, 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) + sde_support=False) self.train_freq = train_freq self.gradient_steps = gradient_steps @@ -104,13 +88,6 @@ class TD3(OffPolicyRLModel): self.target_noise_clip = target_noise_clip self.target_policy_noise = target_policy_noise - # State Dependent Exploration - self.sde_max_grad_norm = sde_max_grad_norm - self.sde_ent_coef = sde_ent_coef - self.sde_log_std_scheduler = sde_log_std_scheduler - self.on_policy_exploration = True - self.sde_vf = None - if _init_setup_model: self._setup_model() @@ -123,7 +100,6 @@ class TD3(OffPolicyRLModel): self.actor_target = self.policy.actor_target self.critic = self.policy.critic self.critic_target = self.policy.critic_target - self.vf_net = self.policy.vf_net def train(self, gradient_steps: int, batch_size: int = 100, policy_delay: int = 2) -> None: @@ -178,52 +154,6 @@ class TD3(OffPolicyRLModel): self._n_updates += gradient_steps logger.logkv("n_updates", self._n_updates) - def train_sde(self) -> None: - # Update optimizer learning rate - # self._update_learning_rate(self.policy.optimizer) - - # Unpack - obs, action, advantage, returns = [self.rollout_data[key] for key in - ['observations', 'actions', 'advantage', 'returns']] - - log_prob, entropy = self.actor.evaluate_actions(obs, action) - values = self.vf_net(obs).flatten() - - # Normalize advantage - # if self.normalize_advantage: - # advantage = (advantage - advantage.mean()) / (advantage.std() + 1e-8) - - # Value loss using the TD(gae_lambda) target - value_loss = F.mse_loss(returns, values) - - # A2C loss - policy_loss = -(advantage * log_prob).mean() - - # Entropy loss favor exploration - if entropy is None: - # Approximate entropy when no analytical form - entropy_loss = -log_prob.mean() - else: - entropy_loss = -th.mean(entropy) - - vf_coef = 0.5 - loss = policy_loss + self.sde_ent_coef * entropy_loss + vf_coef * value_loss - - # Optimization step - self.actor.sde_optimizer.zero_grad() - loss.backward() - - assert not th.isnan(log_prob).any(), log_prob - assert not th.isnan(entropy).any() - assert not th.isnan(self.actor.log_std.grad).any() - assert not th.isnan(self.actor.log_std).any() - - # Clip grad norm - th.nn.utils.clip_grad_norm_([self.actor.log_std], self.sde_max_grad_norm) - self.actor.sde_optimizer.step() - - del self.rollout_data - def learn(self, total_timesteps: int, callback: MaybeCallback = None, @@ -255,16 +185,6 @@ class TD3(OffPolicyRLModel): self._update_current_progress(self.num_timesteps, total_timesteps) if self.num_timesteps > 0 and self.num_timesteps > self.learning_starts: - - if self.use_sde: - if self.sde_log_std_scheduler is not None: - # Call the scheduler - value = self.sde_log_std_scheduler(self._current_progress) - self.actor.log_std.data = th.ones_like(self.actor.log_std) * value - else: - # On-policy gradient - self.train_sde() - gradient_steps = self.gradient_steps if self.gradient_steps > 0 else rollout.episode_timesteps self.train(gradient_steps, batch_size=self.batch_size, policy_delay=self.policy_delay) @@ -280,7 +200,7 @@ class TD3(OffPolicyRLModel): :return: (List[str]) List of parameters that should be excluded from save """ # Exclude aliases - return super(TD3, self).excluded_save_params() + ["actor", "critic", "vf_net", "actor_target", "critic_target"] + return super(TD3, self).excluded_save_params() + ["actor", "critic", "actor_target", "critic_target"] def get_torch_variables(self) -> Tuple[List[str], List[str]]: """ diff --git a/stable_baselines3/version.txt b/stable_baselines3/version.txt index e2dcf9a..04b69b8 100644 --- a/stable_baselines3/version.txt +++ b/stable_baselines3/version.txt @@ -1 +1 @@ -0.6.0a6 +0.6.0a7 diff --git a/tests/test_sde.py b/tests/test_sde.py index eed012f..dbf1963 100644 --- a/tests/test_sde.py +++ b/tests/test_sde.py @@ -2,7 +2,7 @@ import pytest import torch as th from torch.distributions import Normal -from stable_baselines3 import A2C, TD3, SAC, PPO +from stable_baselines3 import A2C, SAC, PPO def test_state_dependent_exploration_grad(): @@ -54,20 +54,10 @@ def test_state_dependent_exploration_grad(): assert sigma_hat.grad.allclose(grad) -@pytest.mark.parametrize("model_class", [TD3, SAC, A2C, PPO]) +@pytest.mark.parametrize("model_class", [SAC, A2C, PPO]) @pytest.mark.parametrize("sde_net_arch", [None, [32, 16], []]) @pytest.mark.parametrize("use_expln", [False, True]) def test_state_dependent_offpolicy_noise(model_class, sde_net_arch, use_expln): model = model_class('MlpPolicy', 'Pendulum-v0', use_sde=True, seed=None, create_eval_env=True, verbose=1, policy_kwargs=dict(log_std_init=-2, sde_net_arch=sde_net_arch, use_expln=use_expln)) model.learn(total_timesteps=int(500), eval_freq=250) - - -def test_scheduler(): - def scheduler(progress): - return -2.0 * progress + 1 - - model = TD3('MlpPolicy', 'Pendulum-v0', use_sde=True, seed=None, create_eval_env=True, - verbose=1, sde_log_std_scheduler=scheduler) - model.learn(total_timesteps=int(1000), eval_freq=500) - assert th.isclose(model.actor.log_std, th.ones_like(model.actor.log_std)).all()