sac_discrete.py

from typing import Any, Dict, List, Optional, Tuple, Type, TypeVar, Union

import numpy as np
import torch as th
from gym import spaces
# import gymnasium
from torch.nn import functional as F

from stable_baselines3.common.buffers import DictReplayBuffer, ReplayBuffer
from stable_baselines3.common.callbacks import BaseCallback
from stable_baselines3.common.noise import ActionNoise, VectorizedActionNoise
from stable_baselines3.common.off_policy_algorithm import OffPolicyAlgorithm
from stable_baselines3.common.policies import BasePolicy
from stable_baselines3.common.type_aliases import GymEnv, MaybeCallback, RolloutReturn, Schedule, TrainFreq, TrainFrequencyUnit
from stable_baselines3.common.utils import get_parameters_by_name, polyak_update, safe_mean, should_collect_more_steps
from stable_baselines3.common.vec_env import VecEnv

from sac_discrete_policies import SACDiscretePolicy, CnnPolicy, MlpPolicy, MultiInputPolicy


SelfSACDiscrete = TypeVar("SelfSACDiscrete", bound="SACDiscrete")

class SACDiscrete(OffPolicyAlgorithm):
    """
    Soft Actor-Critic (SAC) with discrete actions
    Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor, for discrete action spaces
    Paper: https://arxiv.org/abs/1910.07207
    Code: https://github.com/p-christ/Deep-Reinforcement-Learning-Algorithms-with-PyTorch
    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: The policy model to use (MlpPolicy, CnnPolicy, ...)
    :param env: The environment to learn from (if registered in Gym, can be str)
    :param learning_rate: 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: size of the replay buffer
    :param learning_starts: how many steps of the model to collect transitions for before learning starts
    :param batch_size: Minibatch size for each gradient update
    :param tau: the soft update coefficient ("Polyak update", between 0 and 1)
    :param gamma: the discount factor
    :param train_freq: Update the model every ``train_freq`` steps. Alternatively pass a tuple of frequency and unit
        like ``(5, "step")`` or ``(2, "episode")``.
    :param gradient_steps: How many gradient steps to do after each rollout (see ``train_freq``)
        Set to ``-1`` means to do as many gradient steps as steps done in the environment
        during the rollout.
    :param action_noise: the action noise type (None by default), this can help
        for hard exploration problem. Cf common.noise for the different action noise type.
    :param replay_buffer_class: Replay buffer class to use (for instance ``HerReplayBuffer``).
        If ``None``, it will be automatically selected.
    :param replay_buffer_kwargs: Keyword arguments to pass to the replay buffer on creation.
    :param optimize_memory_usage: Enable a memory efficient variant of the replay buffer
        at a cost of more complexity.
        See https://github.com/DLR-RM/stable-baselines3/issues/37#issuecomment-637501195
    :param ent_coef: 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: update the target network every ``target_network_update_freq``
        gradient steps.
    :param target_entropy: target entropy when learning ``ent_coef`` (``ent_coef = 'auto'``)
    :param use_sde: Whether to use generalized State Dependent Exploration (gSDE)
        instead of action noise exploration (default: False)
    :param sde_sample_freq: Sample a new noise matrix every n steps when using gSDE
        Default: -1 (only sample at the beginning of the rollout)
    :param use_sde_at_warmup: Whether to use gSDE instead of uniform sampling
        during the warm up phase (before learning starts)
    :param policy_kwargs: additional arguments to be passed to the policy on creation
    :param verbose: Verbosity level: 0 for no output, 1 for info messages (such as device or wrappers used), 2 for
        debug messages
    :param seed: Seed for the pseudo random generators
    :param 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: Whether or not to build the network at the creation of the instance
    """

    policy_aliases: Dict[str, Type[BasePolicy]] = {
        "MlpPolicy": MlpPolicy,
        "CnnPolicy": CnnPolicy,
        "MultiInputPolicy": MultiInputPolicy,
    }

    def __init__(
        self,
        policy: Union[str, Type[SACDiscretePolicy]],
        env: Union[GymEnv, str],
        learning_rate: Union[float, Schedule] = 3e-4,
        buffer_size: int = 1_000_000,  # 1e6
        learning_starts: int = 100,
        batch_size: int = 256,
        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,
        replay_buffer_class: Optional[Type[ReplayBuffer]] = None,
        replay_buffer_kwargs: Optional[Dict[str, Any]] = None,
        optimize_memory_usage: bool = False,
        ent_coef: Union[str, float] = "auto",
        target_update_interval: int = 1,
        target_entropy: Union[str, float] = "auto",
        target_entropy_mult: float = 0.0,
        reward_scale: float = 1.0,
        use_sde: bool = False,
        sde_sample_freq: int = -1,
        use_sde_at_warmup: bool = False,
        tensorboard_log: Optional[str] = None,
        policy_kwargs: Optional[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,
            learning_rate,
            buffer_size,
            learning_starts,
            batch_size,
            tau,
            gamma,
            train_freq,
            gradient_steps,
            action_noise,
            replay_buffer_class=replay_buffer_class,
            replay_buffer_kwargs=replay_buffer_kwargs,
            policy_kwargs=policy_kwargs,
            tensorboard_log=tensorboard_log,
            verbose=verbose,
            device=device,
            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=(spaces.Discrete, gymnasium.spaces.Discrete),
            supported_action_spaces=(spaces.Discrete,),
            support_multi_env=True,
        )

        self.target_entropy = target_entropy
        self.target_entropy_mult = target_entropy_mult
        self.reward_scale = reward_scale
        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

        if _init_setup_model:
            self._setup_model()

    def _setup_model(self) -> None:
        super()._setup_model()
        self._create_aliases()
        # Running mean and running var
        self.batch_norm_stats = get_parameters_by_name(self.critic, ["running_"])
        self.batch_norm_stats_target = get_parameters_by_name(self.critic_target, ["running_"])
        # 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)
            # why does this have a negative sign by default???
            # possibly differential entropy issues?
            # this is the minimum entropy we expect out of the policy
            # so we're driving the policy to deterministic...
            # same as p-christ implementation, almost max-entropy
            # entropy target as 0.9 on one and 0.1 on rest
            # num_actions = self.env.action_space.n
            # self.target_entropy = 0.9 * np.log(0.9) - (0.1 / (num_actions - 1)) * np.log(0.1 / (num_actions - 1))
            assert self.target_entropy_mult >= 0.0 and self.target_entropy_mult <= 1.0
            if self.target_entropy_mult > 0.0:
                self.target_entropy = self.target_entropy_mult * np.log(self.env.action_space.n)
            else:
                self.target_entropy = -np.log(self.env.action_space.n)
        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), device=self.device)
            self.log_ent_coef = th.log(self.ent_coef_tensor)

    def _create_aliases(self) -> None:
        self.actor = self.policy.actor
        self.critic = self.policy.critic
        self.critic_target = self.policy.critic_target

    def train(self, gradient_steps: int, batch_size: int = 64) -> None:
        # Switch to train mode (this affects batch norm / dropout)
        self.policy.set_training_mode(True)
        # 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 = [], []

        for gradient_step in range(gradient_steps):
            # Sample replay buffer
            replay_data = self.replay_buffer.sample(batch_size, env=self._vec_normalize_env)

            # 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)
            # change to get whole action distribution
            log_probs = self.actor.get_action_distrib(replay_data.observations)
            probs = log_probs.exp()

            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()
                # change with modification to distrib
                # entropy loss over all actions
                # ent_coef_vec = -self.log_ent_coef.exp() * (log_probs + self.target_entropy).detach()
                ent_coef_vec = -self.log_ent_coef * (log_probs + self.target_entropy).detach()
                ent_coef_loss = (probs.detach() * ent_coef_vec).sum(dim=1).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():
                # change to whole distribution
                # 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
                # next_q_values = th.cat(self.critic_target(replay_data.next_observations, next_actions), dim=1)
                # next_q_values, _ = th.min(next_q_values, dim=1, keepdim=True)
                next_log_probs = self.actor.get_action_distrib(replay_data.next_observations)
                next_log_probs = next_log_probs.detach()
                next_q_values = th.stack(self.critic_target(replay_data.next_observations), dim=2)
                next_q_values, _ = th.min(next_q_values, dim=2)
                # add entropy term
                # next_q_values = next_q_values - ent_coef * next_log_prob.reshape(-1, 1)
                next_q_values = next_q_values - ent_coef * next_log_probs
                next_q_values = (next_log_probs.exp() * next_q_values).sum(dim=1).reshape(-1, 1)
                # td error + entropy term
                target_q_values = replay_data.rewards + (1 - replay_data.dones) * self.gamma * next_q_values  # next_q_values should be (bsz, 1)

            # Get current Q-values estimates for each critic network
            # using action from the replay buffer
            current_q_values = self.critic(replay_data.observations)
            # select action idxs
            current_q_values = tuple(current_q.gather(1, replay_data.actions) for current_q in current_q_values)

            critic_loss = 0.5 * sum(F.mse_loss(current_q, target_q_values) for current_q in current_q_values)
            critic_losses.append(critic_loss.item())

            # Optimize the critic
            self.critic.optimizer.zero_grad()
            critic_loss.backward()
            self.critic.optimizer.step()

            # Compute actor loss
            # Alternative: actor_loss = th.mean(log_prob - qf1_pi)
            # Min over all critic networks
            q_values_pi = th.stack(self.critic(replay_data.observations), dim=2)
            min_qf_pi, _ = th.min(q_values_pi, dim=2)
            #: detaches around probs, following ray implementation
            actor_loss = (probs * (ent_coef * log_probs - min_qf_pi.detach())).sum(dim=1).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)
                # Copy running stats, see GH issue #996
                polyak_update(self.batch_norm_stats, self.batch_norm_stats_target, 1.0)

        self._n_updates += gradient_steps

        self.logger.record("train/n_updates", self._n_updates, exclude="tensorboard")
        self.logger.record("train/ent_coef", np.mean(ent_coefs))
        self.logger.record("train/actor_loss", np.mean(actor_losses))
        self.logger.record("train/critic_loss", np.mean(critic_losses))
        if len(ent_coef_losses) > 0:
            self.logger.record("train/ent_coef_loss", np.mean(ent_coef_losses))

    def learn(
        self: SelfSACDiscrete,
        total_timesteps: int,
        callback: MaybeCallback = None,
        log_interval: int = 4,
        tb_log_name: str = "SACDiscrete",
        reset_num_timesteps: bool = True,
        progress_bar: bool = False,
    ) -> SelfSACDiscrete:

        return super().learn(
            total_timesteps=total_timesteps,
            callback=callback,
            log_interval=log_interval,
            tb_log_name=tb_log_name,
            reset_num_timesteps=reset_num_timesteps,
            progress_bar=progress_bar,
        )

    def _excluded_save_params(self) -> List[str]:
        return super()._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"]
        if self.ent_coef_optimizer is not None:
            saved_pytorch_variables = ["log_ent_coef"]
            state_dicts.append("ent_coef_optimizer")
        else:
            saved_pytorch_variables = ["ent_coef_tensor"]
        return state_dicts, saved_pytorch_variables


class TemporallyExtendedSACDiscrete(SACDiscrete):
    """
    Modification of SACDiscrete, like SB3, so that temporally extended actions
    count timesteps correctly and make comparison fair.
    """

    def collect_rollouts(
        self,
        env: VecEnv,
        callback: BaseCallback,
        train_freq: TrainFreq,
        replay_buffer: ReplayBuffer,
        action_noise: Optional[ActionNoise] = None,
        learning_starts: int = 0,
        log_interval: Optional[int] = None,
    ) -> RolloutReturn:
        """
        Collect experiences and store them into a ``ReplayBuffer``.

        :param env: The training environment
        :param callback: Callback that will be called at each step
            (and at the beginning and end of the rollout)
        :param train_freq: How much experience to collect
            by doing rollouts of current policy.
            Either ``TrainFreq(<n>, TrainFrequencyUnit.STEP)``
            or ``TrainFreq(<n>, TrainFrequencyUnit.EPISODE)``
            with ``<n>`` being an integer greater than 0.
        :param action_noise: Action noise that will be used for exploration
            Required for deterministic policy (e.g. TD3). This can also be used
            in addition to the stochastic policy for SAC.
        :param learning_starts: Number of steps before learning for the warm-up phase.
        :param replay_buffer:
        :param log_interval: Log data every ``log_interval`` episodes
        :return:
        """
        # Switch to eval mode (this affects batch norm / dropout)
        self.policy.set_training_mode(False)

        num_collected_steps, num_collected_episodes = 0, 0

        assert isinstance(env, VecEnv), "You must pass a VecEnv"
        assert train_freq.frequency > 0, "Should at least collect one step or episode."

        if env.num_envs > 1:
            assert train_freq.unit == TrainFrequencyUnit.STEP, "You must use only one env when doing episodic training."

        # Vectorize action noise if needed
        if action_noise is not None and env.num_envs > 1 and not isinstance(action_noise, VectorizedActionNoise):
            action_noise = VectorizedActionNoise(action_noise, env.num_envs)

        callback.on_rollout_start()
        continue_training = True
        while should_collect_more_steps(train_freq, num_collected_steps, num_collected_episodes):

            # Select action randomly or according to policy
            actions, buffer_actions = self._sample_action(learning_starts, action_noise, env.num_envs)

            # Rescale and perform action
            new_obs, rewards, dones, infos = env.step(actions)


            # tally primitive action timesteps passed for all environments
            # self.num_timesteps += env.num_envs
            self.num_timesteps += sum(e._primitive_steps for e in env.envs)
            num_collected_steps += 1

            # Give access to local variables
            callback.update_locals(locals())
            # Only stop training if return value is False, not when it is None.
            if callback.on_step() is False:
                return RolloutReturn(num_collected_steps * env.num_envs, num_collected_episodes, continue_training=False)

            # Retrieve reward and episode length if using Monitor wrapper
            self._update_info_buffer(infos, dones)

            # Store data in replay buffer (normalized action and unnormalized observation)
            self._store_transition(replay_buffer, buffer_actions, new_obs, rewards, dones, infos)

            self._update_current_progress_remaining(self.num_timesteps, self._total_timesteps)

            # For DQN, check if the target network should be updated
            # and update the exploration schedule
            # For SAC/TD3, the update is dones as the same time as the gradient update
            # see https://github.com/hill-a/stable-baselines/issues/900
            self._on_step()

            for idx, done in enumerate(dones):
                if done:
                    # Update stats
                    num_collected_episodes += 1
                    self._episode_num += 1

                    if action_noise is not None:
                        kwargs = dict(indices=[idx]) if env.num_envs > 1 else {}
                        action_noise.reset(**kwargs)

                    # Log training infos
                    if log_interval is not None and self._episode_num % log_interval == 0:
                        self._dump_logs()
        callback.on_rollout_end()

        return RolloutReturn(num_collected_steps * env.num_envs, num_collected_episodes, continue_training)