metastable-baselines/sb3_trl/trl_pg/trl_pg.py

import warnings
from typing import Any, Dict, Optional, Type, Union

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

from stable_baselines3.common.on_policy_algorithm import OnPolicyAlgorithm
from stable_baselines3.common.policies import ActorCriticCnnPolicy, ActorCriticPolicy, BasePolicy, MultiInputActorCriticPolicy
from stable_baselines3.common.type_aliases import GymEnv, MaybeCallback, Schedule
from stable_baselines3.common.utils import explained_variance, get_schedule_fn
from stable_baselines3.common.vec_env import VecEnv
from stable_baselines3.common.buffers import RolloutBuffer
from stable_baselines3.common.callbacks import BaseCallback
from stable_baselines3.common.utils import obs_as_tensor

from ..projections.base_projection_layer import BaseProjectionLayer
from ..projections.frob_projection_layer import FrobeniusProjectionLayer


class TRL_PG(OnPolicyAlgorithm):
    """
    Differential Trust Region Layer (TRL) for Policy Gradient (PG)

    Paper: https://arxiv.org/abs/2101.09207
    Code: This implementation borrows (/steals most) code from SB3's PPO implementation https://github.com/DLR-RM/stable-baselines3/blob/master/stable_baselines3/ppo/ppo.py
    The implementation of the TRL-specific parts borrows from https://github.com/boschresearch/trust-region-layers/blob/main/trust_region_projections/algorithms/pg/pg.py

    :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: The learning rate, it can be a function
        of the current progress remaining (from 1 to 0)
    :param n_steps: The number of steps to run for each environment per update
        (i.e. rollout buffer size is n_steps * n_envs where n_envs is number of environment copies running in parallel)
        NOTE: n_steps * n_envs must be greater than 1 (because of the advantage normalization)
        See https://github.com/pytorch/pytorch/issues/29372
    :param batch_size: Minibatch size
    :param n_epochs: Number of epoch when optimizing the surrogate loss
    :param gamma: Discount factor
    :param gae_lambda: Factor for trade-off of bias vs variance for Generalized Advantage Estimator
    :param clip_range: Clipping parameter, it can be a function of the current progress
        remaining (from 1 to 0).
    :param clip_range_vf: Clipping parameter for the value function,
        it can be a function of the current progress remaining (from 1 to 0).
        This is a parameter specific to the OpenAI implementation. If None is passed (default),
        no clipping will be done on the value function.
        IMPORTANT: this clipping depends on the reward scaling.
    :param normalize_advantage: Whether to normalize or not the advantage
    :param ent_coef: Entropy coefficient for the loss calculation
    :param vf_coef: Value function coefficient for the loss calculation
    :param max_grad_norm: The maximum value for the gradient clipping
    :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 target_kl: Limit the KL divergence between updates,
        because the clipping is not enough to prevent large update
        see issue #213 (cf https://github.com/hill-a/stable-baselines/issues/213)
        By default, there is no limit on the kl div.
    :param tensorboard_log: the log location for tensorboard (if None, no logging)
    :param create_eval_env: 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: additional arguments to be passed to the policy on creation
    :param verbose: the verbosity level: 0 no output, 1 info, 2 debug
    :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 projection: What kind of Projection to use
    :param _init_setup_model: Whether or not to build the network at the creation of the instance
    """

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

    def __init__(
        self,
        policy: Union[str, Type[ActorCriticPolicy]],
        env: Union[GymEnv, str],
        learning_rate: Union[float, Schedule] = 3e-4,
        n_steps: int = 2048,
        batch_size: int = 64,
        n_epochs: int = 10,
        gamma: float = 0.99,
        gae_lambda: float = 0.95,
        clip_range: Union[float, Schedule] = 0.2,
        clip_range_vf: Union[None, float, Schedule] = None,
        normalize_advantage: bool = True,
        ent_coef: float = 0.0,
        vf_coef: float = 0.5,
        max_grad_norm: float = 0.5,
        use_sde: bool = False,
        sde_sample_freq: int = -1,
        target_kl: Optional[float] = None,
        tensorboard_log: Optional[str] = None,
        create_eval_env: bool = False,
        policy_kwargs: Optional[Dict[str, Any]] = None,
        verbose: int = 0,
        seed: Optional[int] = None,
        device: Union[th.device, str] = "auto",

        # Different from PPO:
        projection: BaseProjectionLayer = BaseProjectionLayer,

        _init_setup_model: bool = True,
    ):

        super().__init__(
            policy,
            env,
            learning_rate=learning_rate,
            n_steps=n_steps,
            gamma=gamma,
            gae_lambda=gae_lambda,
            ent_coef=ent_coef,
            vf_coef=vf_coef,
            max_grad_norm=max_grad_norm,
            use_sde=use_sde,
            sde_sample_freq=sde_sample_freq,
            tensorboard_log=tensorboard_log,
            policy_kwargs=policy_kwargs,
            verbose=verbose,
            device=device,
            create_eval_env=create_eval_env,
            seed=seed,
            _init_setup_model=False,
            supported_action_spaces=(
                spaces.Box,
                spaces.Discrete,
                spaces.MultiDiscrete,
                spaces.MultiBinary,
            ),
        )

        # Sanity check, otherwise it will lead to noisy gradient and NaN
        # because of the advantage normalization
        if normalize_advantage:
            assert (
                batch_size > 1
            ), "`batch_size` must be greater than 1. See https://github.com/DLR-RM/stable-baselines3/issues/440"

        if self.env is not None:
            # Check that `n_steps * n_envs > 1` to avoid NaN
            # when doing advantage normalization
            buffer_size = self.env.num_envs * self.n_steps
            assert (
                buffer_size > 1
            ), f"`n_steps * n_envs` must be greater than 1. Currently n_steps={self.n_steps} and n_envs={self.env.num_envs}"
            # Check that the rollout buffer size is a multiple of the mini-batch size
            untruncated_batches = buffer_size // batch_size
            if buffer_size % batch_size > 0:
                warnings.warn(
                    f"You have specified a mini-batch size of {batch_size},"
                    f" but because the `RolloutBuffer` is of size `n_steps * n_envs = {buffer_size}`,"
                    f" after every {untruncated_batches} untruncated mini-batches,"
                    f" there will be a truncated mini-batch of size {buffer_size % batch_size}\n"
                    f"We recommend using a `batch_size` that is a factor of `n_steps * n_envs`.\n"
                    f"Info: (n_steps={self.n_steps} and n_envs={self.env.num_envs})"
                )
        self.batch_size = batch_size
        self.n_epochs = n_epochs
        self.clip_range = clip_range
        self.clip_range_vf = clip_range_vf
        self.normalize_advantage = normalize_advantage
        self.target_kl = target_kl

        # Different from PPO:
        self.projection = projection
        self._global_steps = 0

        if _init_setup_model:
            self._setup_model()

    def _setup_model(self) -> None:
        super()._setup_model()

        # Initialize schedules for policy/value clipping
        self.clip_range = get_schedule_fn(self.clip_range)
        if self.clip_range_vf is not None:
            if isinstance(self.clip_range_vf, (float, int)):
                assert self.clip_range_vf > 0, "`clip_range_vf` must be positive, " "pass `None` to deactivate vf clipping"

            self.clip_range_vf = get_schedule_fn(self.clip_range_vf)

    def train(self) -> None:
        """
        Update policy using the currently gathered rollout buffer.
        """
        # Switch to train mode (this affects batch norm / dropout)
        self.policy.set_training_mode(True)
        # Update optimizer learning rate
        self._update_learning_rate(self.policy.optimizer)
        # Compute current clip range
        clip_range = self.clip_range(self._current_progress_remaining)
        # Optional: clip range for the value function
        if self.clip_range_vf is not None:
            clip_range_vf = self.clip_range_vf(
                self._current_progress_remaining)

        surrogate_losses = []
        entropy_losses = []
        trust_region_losses = []
        pg_losses, value_losses = [], []
        clip_fractions = []

        continue_training = True

        # train for n_epochs epochs
        for epoch in range(self.n_epochs):
            approx_kl_divs = []
            # Do a complete pass on the rollout buffer
            for rollout_data in self.rollout_buffer.get(self.batch_size):
                # This is new compared to PPO.
                # Calculating the TR-Projections we need to know the step number
                self._global_steps += 1

                actions = rollout_data.actions
                if isinstance(self.action_space, spaces.Discrete):
                    # Convert discrete action from float to long
                    actions = rollout_data.actions.long().flatten()

                # Re-sample the noise matrix because the log_std has changed
                if self.use_sde:
                    self.policy.reset_noise(self.batch_size)

                # old code for PPO
                # values, log_prob, entropy = self.policy.evaluate_actions(rollout_data.observations, actions)

                # src in TRL reference code:
                # Stolen from Fabian's Code (Public Version):
                # p = self.policy(rollout_data.observations)
                # proj_p = self.projection(self.policy, p, b_q = (b_old_mean, b_old_std), self._global_step)
                # new_logpacs = self.policy.log_probability(proj_p, b_actions)
                # log_prob == new_pogpacs (i think)

                # src of evaluate_actions:
                # features = self.extract_features(obs)
                # latent_pi, latent_vf = self.mlp_extractor(features)
                # distribution = self._get_action_dist_from_latent(latent_pi)
                # log_prob = distribution.log_prob(actions)
                # values = self.value_net(latent_vf)
                # return values, log_prob, distribution.entropy()

                # here we go:
                pol = self.policy
                features = pol.extract_features(rollout_data.observations)
                latent_pi, latent_vf = pol.mlp_extractor(features)
                p = pol._get_action_dist_from_latent(latent_pi)
                b_q = rollout_data.mean, rollout_data.std
                proj_p = self.projection(pol, p, b_q, self._global_step)
                log_prob = proj_p.log_prob(actions)
                # or log_prob = pol.log_probability(proj_p, actions)
                values = self.value_net(latent_vf)
                entropy = proj_p.entropy()  # or not...

                values = values.flatten()
                # Normalize advantage
                advantages = rollout_data.advantages
                if self.normalize_advantage:
                    advantages = (advantages - advantages.mean()
                                  ) / (advantages.std() + 1e-8)

                # ratio between old and new policy, should be one at the first iteration
                ratio = th.exp(log_prob - rollout_data.old_log_prob)

                # Difference from PPO: We renamed 'policy_loss' to 'surrogate_loss'
                # clipped surrogate loss
                surrogate_loss_1 = advantages * ratio
                surrogate_loss_2 = advantages * \
                    th.clamp(ratio, 1 - clip_range, 1 + clip_range)
                surrogate_loss = - \
                    th.min(surrogate_loss_1, surrogate_loss_2).mean()

                surrogate_losses.append(surrogate_loss.item())

                clip_fraction = th.mean(
                    (th.abs(ratio - 1) > clip_range).float()).item()
                clip_fractions.append(clip_fraction)

                if self.clip_range_vf is None:
                    # No clipping
                    values_pred = values
                else:
                    # Clip the different between old and new value
                    # NOTE: this depends on the reward scaling
                    values_pred = rollout_data.old_values + th.clamp(
                        values - rollout_data.old_values, -clip_range_vf, clip_range_vf
                    )
                # Value loss using the TD(gae_lambda) target
                value_loss = F.mse_loss(rollout_data.returns, values_pred)
                value_losses.append(value_loss.item())

                # Entropy loss favor exploration
                if entropy is None:
                    # Approximate entropy when no analytical form
                    entropy_loss = -th.mean(-log_prob)
                else:
                    entropy_loss = -th.mean(entropy)

                entropy_losses.append(entropy_loss.item())

                # Difference to PPO: Added trust_region_loss; policy_loss includes entropy_loss + trust_region_loss
                trust_region_loss = self.projection.get_trust_region_loss(
                    pol, p, proj_p)
                # NOTE to future-self: policy has a different interface then in orig TRL-impl.

                trust_region_losses.append(trust_region_loss.item())

                policy_loss = surrogate_loss + self.ent_coef * entropy_loss + trust_region_loss
                pg_losses.append(policy_loss.item())

                loss = policy_loss + self.vf_coef * value_loss

                # Calculate approximate form of reverse KL Divergence for early stopping
                # see issue #417: https://github.com/DLR-RM/stable-baselines3/issues/417
                # and discussion in PR #419: https://github.com/DLR-RM/stable-baselines3/pull/419
                # and Schulman blog: http://joschu.net/blog/kl-approx.html
                with th.no_grad():
                    log_ratio = log_prob - rollout_data.old_log_prob
                    approx_kl_div = th.mean(
                        (th.exp(log_ratio) - 1) - log_ratio).cpu().numpy()
                    approx_kl_divs.append(approx_kl_div)

                if self.target_kl is not None and approx_kl_div > 1.5 * self.target_kl:
                    continue_training = False
                    if self.verbose >= 1:
                        print(
                            f"Early stopping at step {epoch} due to reaching max kl: {approx_kl_div:.2f}")
                    break

                # Optimization step
                self.policy.optimizer.zero_grad()
                loss.backward()
                # Clip grad norm
                th.nn.utils.clip_grad_norm_(
                    self.policy.parameters(), self.max_grad_norm)
                self.policy.optimizer.step()

            if not continue_training:
                break

        self._n_updates += self.n_epochs
        explained_var = explained_variance(
            self.rollout_buffer.values.flatten(), self.rollout_buffer.returns.flatten())

        # Logs
        self.logger.record("train/surrogate_loss", np.mean(surrogate_losses))
        self.logger.record("train/entropy_loss", np.mean(entropy_losses))
        self.logger.record("train/trust_region_loss",
                           np.mean(trust_region_losses))
        self.logger.record("train/policy_gradient_loss", np.mean(pg_losses))
        self.logger.record("train/value_loss", np.mean(value_losses))
        self.logger.record("train/approx_kl", np.mean(approx_kl_divs))
        self.logger.record("train/clip_fraction", np.mean(clip_fractions))
        self.logger.record("train/loss", loss.item())
        self.logger.record("train/explained_variance", explained_var)
        if hasattr(self.policy, "log_std"):
            self.logger.record(
                "train/std", th.exp(self.policy.log_std).mean().item())

        self.logger.record("train/n_updates",
                           self._n_updates, exclude="tensorboard")
        self.logger.record("train/clip_range", clip_range)
        if self.clip_range_vf is not None:
            self.logger.record("train/clip_range_vf", clip_range_vf)

    def learn(
        self,
        total_timesteps: int,
        callback: MaybeCallback = None,
        log_interval: int = 1,
        eval_env: Optional[GymEnv] = None,
        eval_freq: int = -1,
        n_eval_episodes: int = 5,
        tb_log_name: str = "TRL_PG",
        eval_log_path: Optional[str] = None,
        reset_num_timesteps: bool = True,
    ) -> "TRL_PG":

        return super().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,
        )

    # This is new compared to PPO.
    # TRL requires us to also save the original mean and std in our rollouts
    def collect_rollouts(
        self,
        env: VecEnv,
        callback: BaseCallback,
        rollout_buffer: RolloutBuffer,
        n_rollout_steps: int,
    ) -> bool:
        """
        Collect experiences using the current policy and fill a ``RolloutBuffer``.
        The term rollout here refers to the model-free notion and should not
        be used with the concept of rollout used in model-based RL or planning.
        :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 rollout_buffer: Buffer to fill with rollouts
        :param n_steps: Number of experiences to collect per environment
        :return: True if function returned with at least `n_rollout_steps`
            collected, False if callback terminated rollout prematurely.
        """
        assert self._last_obs is not None, "No previous observation was provided"
        # Switch to eval mode (this affects batch norm / dropout)
        self.policy.set_training_mode(False)

        n_steps = 0
        rollout_buffer.reset()
        # Sample new weights for the state dependent exploration
        if self.use_sde:
            self.policy.reset_noise(env.num_envs)

        callback.on_rollout_start()

        while n_steps < n_rollout_steps:
            if self.use_sde and self.sde_sample_freq > 0 and n_steps % self.sde_sample_freq == 0:
                # Sample a new noise matrix
                self.policy.reset_noise(env.num_envs)

            with th.no_grad():
                # Convert to pytorch tensor or to TensorDict
                obs_tensor = obs_as_tensor(self._last_obs, self.device)
                actions, values, log_probs = self.policy(obs_tensor)
            actions = actions.cpu().numpy()

            # Rescale and perform action
            clipped_actions = actions
            # Clip the actions to avoid out of bound error
            if isinstance(self.action_space, spaces.Box):
                clipped_actions = np.clip(
                    actions, self.action_space.low, self.action_space.high)

            new_obs, rewards, dones, infos = env.step(clipped_actions)

            self.num_timesteps += env.num_envs

            # Give access to local variables
            callback.update_locals(locals())
            if callback.on_step() is False:
                return False

            self._update_info_buffer(infos)
            n_steps += 1

            if isinstance(self.action_space, spaces.Discrete):
                # Reshape in case of discrete action
                actions = actions.reshape(-1, 1)

            # Handle timeout by bootstraping with value function
            # see GitHub issue #633
            for idx, done in enumerate(dones):
                if (
                    done
                    and infos[idx].get("terminal_observation") is not None
                    and infos[idx].get("TimeLimit.truncated", False)
                ):
                    terminal_obs = self.policy.obs_to_tensor(
                        infos[idx]["terminal_observation"])[0]
                    with th.no_grad():
                        terminal_value = self.policy.predict_values(terminal_obs)[
                            0]
                    rewards[idx] += self.gamma * terminal_value

            # TODO: how to calc mean + std
            rollout_buffer.add(self._last_obs, actions, rewards,
                               self._last_episode_starts, values, log_probs, mean, std)
            self._last_obs = new_obs
            self._last_episode_starts = dones

        with th.no_grad():
            # Compute value for the last timestep
            values = self.policy.predict_values(
                obs_as_tensor(new_obs, self.device))

        rollout_buffer.compute_returns_and_advantage(
            last_values=values, dones=dones)

        callback.on_rollout_end()

        return True