lightning/pl_examples/domain_templates/reinforce_learn_ppo.py

# Copyright The PyTorch Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
PyTorch Lightning implementation of Proximal Policy Optimization (PPO)
<https://arxiv.org/abs/1707.06347>
Paper authors: John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov

The example implements PPO compatible to work with any continous or discrete action-space environments via OpenAI Gym.

To run the template, just run:
`python reinforce_learn_ppo.py`

References
----------
[1] https://github.com/openai/baselines/blob/master/baselines/ppo2/ppo2.py
[2] https://github.com/openai/spinningup
[3] https://github.com/sid-sundrani/ppo_lightning
"""
import argparse
from typing import Callable, Iterable, List, Tuple

import gym
import torch
from torch import nn
from torch.distributions import Categorical, Normal
from torch.optim.optimizer import Optimizer
from torch.utils.data import DataLoader, IterableDataset

import pytorch_lightning as pl
from pl_examples import cli_lightning_logo


def create_mlp(input_shape: Tuple[int], n_actions: int, hidden_size: int = 128):
    """
    Simple Multi-Layer Perceptron network
    """
    network = nn.Sequential(
        nn.Linear(input_shape[0], hidden_size),
        nn.ReLU(),
        nn.Linear(hidden_size, hidden_size),
        nn.ReLU(),
        nn.Linear(hidden_size, n_actions),
    )

    return network


class ActorCategorical(nn.Module):
    """
    Policy network, for discrete action spaces, which returns a distribution
    and an action given an observation
    """

    def __init__(self, actor_net):
        """
        Args:
            input_shape: observation shape of the environment
            n_actions: number of discrete actions available in the environment
        """
        super().__init__()

        self.actor_net = actor_net

    def forward(self, states):
        logits = self.actor_net(states)
        pi = Categorical(logits=logits)
        actions = pi.sample()

        return pi, actions

    def get_log_prob(self, pi: Categorical, actions: torch.Tensor):
        """
        Takes in a distribution and actions and returns log prob of actions under the distribution

        Args:
            pi: torch distribution
            actions: actions taken by distribution

        Returns:
            log probability of the acition under pi
        """
        return pi.log_prob(actions)


class ActorContinous(nn.Module):
    """
    Policy network, for continous action spaces, which returns a distribution
    and an action given an observation
    """

    def __init__(self, actor_net, act_dim):
        """
        Args:
            input_shape: observation shape of the environment
            n_actions: number of discrete actions available in the environment
        """
        super().__init__()
        self.actor_net = actor_net
        log_std = -0.5 * torch.ones(act_dim, dtype=torch.float)
        self.log_std = nn.Parameter(log_std)

    def forward(self, states):
        mu = self.actor_net(states)
        std = torch.exp(self.log_std)
        pi = Normal(loc=mu, scale=std)
        actions = pi.sample()

        return pi, actions

    def get_log_prob(self, pi: Normal, actions: torch.Tensor):
        """
        Takes in a distribution and actions and returns log prob of actions under the distribution

        Args:
            pi: torch distribution
            actions: actions taken by distribution

        Returns:
            log probability of the acition under pi
        """
        return pi.log_prob(actions).sum(axis=-1)


class ExperienceSourceDataset(IterableDataset):
    """
    Implementation from PyTorch Lightning Bolts:
    https://github.com/PyTorchLightning/lightning-bolts/blob/master/pl_bolts/datamodules/experience_source.py

    Basic experience source dataset. Takes a generate_batch function that returns an iterator.
    The logic for the experience source and how the batch is generated is defined the Lightning model itself
    """

    def __init__(self, generate_batch: Callable):
        self.generate_batch = generate_batch

    def __iter__(self) -> Iterable:
        iterator = self.generate_batch()
        return iterator


class PPOLightning(pl.LightningModule):
    """
    PyTorch Lightning implementation of PPO.

    Example:
        model = PPOLightning("CartPole-v0")
    Train:
        trainer = Trainer()
        trainer.fit(model)
    """

    def __init__(
        self,
        env: str,
        gamma: float = 0.99,
        lam: float = 0.95,
        lr_actor: float = 3e-4,
        lr_critic: float = 1e-3,
        max_episode_len: float = 200,
        batch_size: int = 512,
        steps_per_epoch: int = 2048,
        nb_optim_iters: int = 4,
        clip_ratio: float = 0.2,
        **kwargs,
    ) -> None:
        """
        Args:
            env: gym environment tag
            gamma: discount factor
            lam: advantage discount factor (lambda in the paper)
            lr_actor: learning rate of actor network
            lr_critic: learning rate of critic network
            max_episode_len: maximum number interactions (actions) in an episode
            batch_size:  batch_size when training network- can simulate number of policy updates performed per epoch
            steps_per_epoch: how many action-state pairs to rollout for trajectory collection per epoch
            nb_optim_iters: how many steps of gradient descent to perform on each batch
            clip_ratio: hyperparameter for clipping in the policy objective
        """
        super().__init__()

        # Hyperparameters
        self.lr_actor = lr_actor
        self.lr_critic = lr_critic
        self.steps_per_epoch = steps_per_epoch
        self.nb_optim_iters = nb_optim_iters
        self.batch_size = batch_size
        self.gamma = gamma
        self.lam = lam
        self.max_episode_len = max_episode_len
        self.clip_ratio = clip_ratio
        self.save_hyperparameters()

        self.env = gym.make(env)
        # value network
        self.critic = create_mlp(self.env.observation_space.shape, 1)
        # policy network (agent)
        if isinstance(self.env.action_space, gym.spaces.box.Box):
            act_dim = self.env.action_space.shape[0]
            actor_mlp = create_mlp(self.env.observation_space.shape, act_dim)
            self.actor = ActorContinous(actor_mlp, act_dim)
        elif isinstance(self.env.action_space, gym.spaces.discrete.Discrete):
            actor_mlp = create_mlp(self.env.observation_space.shape, self.env.action_space.n)
            self.actor = ActorCategorical(actor_mlp)
        else:
            raise NotImplementedError(
                'Env action space should be of type Box (continous) or Discrete (categorical).'
                f' Got type: {type(self.env.action_space)}'
            )

        self.batch_states = []
        self.batch_actions = []
        self.batch_adv = []
        self.batch_qvals = []
        self.batch_logp = []

        self.ep_rewards = []
        self.ep_values = []
        self.epoch_rewards = []

        self.episode_step = 0
        self.avg_ep_reward = 0
        self.avg_ep_len = 0
        self.avg_reward = 0

        self.state = torch.FloatTensor(self.env.reset())

    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Passes in a state x through the network and returns the policy and a sampled action

        Args:
            x: environment state

        Returns:
            Tuple of policy and action
        """
        pi, action = self.actor(x)
        value = self.critic(x)

        return pi, action, value

    def discount_rewards(self, rewards: List[float], discount: float) -> List[float]:
        """Calculate the discounted rewards of all rewards in list

        Args:
            rewards: list of rewards/advantages

        Returns:
            list of discounted rewards/advantages
        """
        assert isinstance(rewards[0], float)

        cumul_reward = []
        sum_r = 0.0

        for r in reversed(rewards):
            sum_r = (sum_r * discount) + r
            cumul_reward.append(sum_r)

        return list(reversed(cumul_reward))

    def calc_advantage(self, rewards: List[float], values: List[float], last_value: float) -> List[float]:
        """Calculate the advantage given rewards, state values, and the last value of episode

        Args:
            rewards: list of episode rewards
            values: list of state values from critic
            last_value: value of last state of episode

        Returns:
            list of advantages
        """
        rews = rewards + [last_value]
        vals = values + [last_value]
        # GAE
        delta = [rews[i] + self.gamma * vals[i + 1] - vals[i] for i in range(len(rews) - 1)]
        adv = self.discount_rewards(delta, self.gamma * self.lam)

        return adv

    def generate_trajectory_samples(self) -> Tuple[List[torch.Tensor], List[torch.Tensor], List[torch.Tensor]]:
        """
        Contains the logic for generating trajectory data to train policy and value network
        Yield:
           Tuple of Lists containing tensors for states, actions, log probs, qvals and advantage
        """

        for step in range(self.steps_per_epoch):
            self.state = self.state.to(device=self.device)

            with torch.no_grad():
                pi, action, value = self(self.state)
                log_prob = self.actor.get_log_prob(pi, action)

            next_state, reward, done, _ = self.env.step(action.cpu().numpy())

            self.episode_step += 1

            self.batch_states.append(self.state)
            self.batch_actions.append(action)
            self.batch_logp.append(log_prob)

            self.ep_rewards.append(reward)
            self.ep_values.append(value.item())

            self.state = torch.FloatTensor(next_state)

            epoch_end = step == (self.steps_per_epoch - 1)
            terminal = len(self.ep_rewards) == self.max_episode_len

            if epoch_end or done or terminal:
                # if trajectory ends abtruptly, boostrap value of next state
                if (terminal or epoch_end) and not done:
                    self.state = self.state.to(device=self.device)
                    with torch.no_grad():
                        _, _, value = self(self.state)
                        last_value = value.item()
                        steps_before_cutoff = self.episode_step
                else:
                    last_value = 0
                    steps_before_cutoff = 0

                # discounted cumulative reward
                self.batch_qvals += self.discount_rewards(self.ep_rewards + [last_value], self.gamma)[:-1]
                # advantage
                self.batch_adv += self.calc_advantage(self.ep_rewards, self.ep_values, last_value)
                # logs
                self.epoch_rewards.append(sum(self.ep_rewards))
                # reset params
                self.ep_rewards = []
                self.ep_values = []
                self.episode_step = 0
                self.state = torch.FloatTensor(self.env.reset())

            if epoch_end:
                train_data = zip(
                    self.batch_states, self.batch_actions, self.batch_logp, self.batch_qvals, self.batch_adv
                )

                for state, action, logp_old, qval, adv in train_data:
                    yield state, action, logp_old, qval, adv

                self.batch_states.clear()
                self.batch_actions.clear()
                self.batch_adv.clear()
                self.batch_logp.clear()
                self.batch_qvals.clear()

                # logging
                self.avg_reward = sum(self.epoch_rewards) / self.steps_per_epoch

                # if epoch ended abruptly, exlude last cut-short episode to prevent stats skewness
                epoch_rewards = self.epoch_rewards
                if not done:
                    epoch_rewards = epoch_rewards[:-1]

                total_epoch_reward = sum(epoch_rewards)
                nb_episodes = len(epoch_rewards)

                self.avg_ep_reward = total_epoch_reward / nb_episodes
                self.avg_ep_len = (self.steps_per_epoch - steps_before_cutoff) / nb_episodes

                self.epoch_rewards.clear()

    def actor_loss(self, state, action, logp_old, qval, adv) -> torch.Tensor:
        pi, _ = self.actor(state)
        logp = self.actor.get_log_prob(pi, action)
        ratio = torch.exp(logp - logp_old)
        clip_adv = torch.clamp(ratio, 1 - self.clip_ratio, 1 + self.clip_ratio) * adv
        loss_actor = -(torch.min(ratio * adv, clip_adv)).mean()
        return loss_actor

    def critic_loss(self, state, action, logp_old, qval, adv) -> torch.Tensor:
        value = self.critic(state)
        loss_critic = (qval - value).pow(2).mean()
        return loss_critic

    def training_step(self, batch: Tuple[torch.Tensor, torch.Tensor], batch_idx, optimizer_idx):
        """
        Carries out a single update to actor and critic network from a batch of replay buffer.

        Args:
            batch: batch of replay buffer/trajectory data
            batch_idx: not used
            optimizer_idx: idx that controls optimizing actor or critic network

        Returns:
            loss
        """
        state, action, old_logp, qval, adv = batch

        # normalize advantages
        adv = (adv - adv.mean()) / adv.std()

        self.log("avg_ep_len", self.avg_ep_len, prog_bar=True, on_step=False, on_epoch=True)
        self.log("avg_ep_reward", self.avg_ep_reward, prog_bar=True, on_step=False, on_epoch=True)
        self.log("avg_reward", self.avg_reward, prog_bar=True, on_step=False, on_epoch=True)

        if optimizer_idx == 0:
            loss_actor = self.actor_loss(state, action, old_logp, qval, adv)
            self.log('loss_actor', loss_actor, on_step=False, on_epoch=True, prog_bar=True, logger=True)

            return loss_actor

        elif optimizer_idx == 1:
            loss_critic = self.critic_loss(state, action, old_logp, qval, adv)
            self.log('loss_critic', loss_critic, on_step=False, on_epoch=True, prog_bar=False, logger=True)

            return loss_critic

    def configure_optimizers(self) -> List[Optimizer]:
        """ Initialize Adam optimizer"""
        optimizer_actor = torch.optim.Adam(self.actor.parameters(), lr=self.lr_actor)
        optimizer_critic = torch.optim.Adam(self.critic.parameters(), lr=self.lr_critic)

        return optimizer_actor, optimizer_critic

    def optimizer_step(self, *args, **kwargs):
        """
        Run 'nb_optim_iters' number of iterations of gradient descent on actor and critic
        for each data sample.
        """
        for _ in range(self.nb_optim_iters):
            super().optimizer_step(*args, **kwargs)

    def _dataloader(self) -> DataLoader:
        """Initialize the Replay Buffer dataset used for retrieving experiences"""
        dataset = ExperienceSourceDataset(self.generate_trajectory_samples)
        dataloader = DataLoader(dataset=dataset, batch_size=self.batch_size)
        return dataloader

    def train_dataloader(self) -> DataLoader:
        """Get train loader"""
        return self._dataloader()

    @staticmethod
    def add_model_specific_args(parent_parser):  # pragma: no-cover
        parser = parent_parser.add_argument_group("PPOLightning")
        parser.add_argument("--env", type=str, default="CartPole-v0")
        parser.add_argument("--gamma", type=float, default=0.99, help="discount factor")
        parser.add_argument("--lam", type=float, default=0.95, help="advantage discount factor")
        parser.add_argument("--lr_actor", type=float, default=3e-4, help="learning rate of actor network")
        parser.add_argument("--lr_critic", type=float, default=1e-3, help="learning rate of critic network")
        parser.add_argument("--max_episode_len", type=int, default=1000, help="capacity of the replay buffer")
        parser.add_argument("--batch_size", type=int, default=512, help="batch_size when training network")
        parser.add_argument(
            "--steps_per_epoch",
            type=int,
            default=2048,
            help="how many action-state pairs to rollout for trajectory collection per epoch"
        )
        parser.add_argument(
            "--nb_optim_iters", type=int, default=4, help="how many steps of gradient descent to perform on each batch"
        )
        parser.add_argument(
            "--clip_ratio", type=float, default=0.2, help="hyperparameter for clipping in the policy objective"
        )

        return parent_parser


def main(args) -> None:
    model = PPOLightning(**vars(args))

    trainer = pl.Trainer.from_argparse_args(args)
    trainer.fit(model)


if __name__ == '__main__':
    cli_lightning_logo()
    pl.seed_everything(0)

    parent_parser = argparse.ArgumentParser(add_help=False)
    parent_parser = pl.Trainer.add_argparse_args(parent_parser)

    parser = PPOLightning.add_model_specific_args(parent_parser)
    args = parser.parse_args()

    main(args)