benchmarl/algorithms/iddpg.py

#  Copyright (c) Meta Platforms, Inc. and affiliates.
#
#  This source code is licensed under the license found in the
#  LICENSE file in the root directory of this source tree.
#

from dataclasses import dataclass, MISSING
from typing import Dict, Iterable, Tuple, Type

import torch
from tensordict import TensorDictBase
from tensordict.nn import TensorDictModule, TensorDictSequential
from torchrl.data import CompositeSpec, UnboundedContinuousTensorSpec
from torchrl.modules import AdditiveGaussianWrapper, ProbabilisticActor, TanhDelta
from torchrl.objectives import DDPGLoss, LossModule, ValueEstimators

from benchmarl.algorithms.common import Algorithm, AlgorithmConfig
from benchmarl.models.common import ModelConfig


class Iddpg(Algorithm):
    """Same as :class:`~benchmarkl.algorithms.Maddpg` (from `https://arxiv.org/abs/1706.02275 <https://arxiv.org/abs/1706.02275>`__) but with decentralized critics.

    Args:
        share_param_critic (bool): Whether to share the parameters of the critics withing agent groups
        loss_function (str): loss function for the value discrepancy. Can be one of "l1", "l2" or "smooth_l1".
        delay_value (bool): whether to separate the target value networks from the value networks used for
            data collection.

    """

    def __init__(
        self, share_param_critic: bool, loss_function: str, delay_value: bool, **kwargs
    ):
        super().__init__(**kwargs)

        self.share_param_critic = share_param_critic
        self.delay_value = delay_value
        self.loss_function = loss_function

    #############################
    # Overridden abstract methods
    #############################

    def _get_loss(
        self, group: str, policy_for_loss: TensorDictModule, continuous: bool
    ) -> Tuple[LossModule, bool]:
        if continuous:
            # Loss
            loss_module = DDPGLoss(
                actor_network=policy_for_loss,
                value_network=self.get_value_module(group),
                delay_value=self.delay_value,
                loss_function=self.loss_function,
            )
            loss_module.set_keys(
                state_action_value=(group, "state_action_value"),
                reward=(group, "reward"),
                priority=(group, "td_error"),
                done=(group, "done"),
                terminated=(group, "terminated"),
            )

            loss_module.make_value_estimator(
                ValueEstimators.TD0, gamma=self.experiment_config.gamma
            )
            return loss_module, True
        else:
            raise NotImplementedError(
                "Iddpg is not compatible with discrete actions yet"
            )

    def _get_parameters(self, group: str, loss: LossModule) -> Dict[str, Iterable]:
        return {
            "loss_actor": list(loss.actor_network_params.flatten_keys().values()),
            "loss_value": list(loss.value_network_params.flatten_keys().values()),
        }

    def _get_policy_for_loss(
        self, group: str, model_config: ModelConfig, continuous: bool
    ) -> TensorDictModule:
        if continuous:
            n_agents = len(self.group_map[group])
            logits_shape = list(self.action_spec[group, "action"].shape)
            actor_input_spec = CompositeSpec(
                {
                    group: CompositeSpec(
                        {
                            "observation": self.observation_spec[group]["observation"]
                            .clone()
                            .to(self.device)
                        },
                        shape=(n_agents,),
                    )
                }
            )
            actor_output_spec = CompositeSpec(
                {
                    group: CompositeSpec(
                        {"param": UnboundedContinuousTensorSpec(shape=logits_shape)},
                        shape=(n_agents,),
                    )
                }
            )
            actor_module = model_config.get_model(
                input_spec=actor_input_spec,
                output_spec=actor_output_spec,
                agent_group=group,
                input_has_agent_dim=True,
                n_agents=n_agents,
                centralised=False,
                share_params=self.experiment_config.share_policy_params,
                device=self.device,
                action_spec=self.action_spec,
            )

            policy = ProbabilisticActor(
                module=actor_module,
                spec=self.action_spec[group, "action"],
                in_keys=[(group, "param")],
                out_keys=[(group, "action")],
                distribution_class=TanhDelta,
                distribution_kwargs={
                    "min": self.action_spec[(group, "action")].space.low,
                    "max": self.action_spec[(group, "action")].space.high,
                },
                return_log_prob=False,
            )
            return policy
        else:
            raise NotImplementedError(
                "Iddpg is not compatible with discrete actions yet"
            )

    def _get_policy_for_collection(
        self, policy_for_loss: TensorDictModule, group: str, continuous: bool
    ) -> TensorDictModule:
        return AdditiveGaussianWrapper(
            policy_for_loss,
            annealing_num_steps=self.experiment_config.get_exploration_anneal_frames(
                self.on_policy
            ),
            action_key=(group, "action"),
            sigma_init=self.experiment_config.exploration_eps_init,
            sigma_end=self.experiment_config.exploration_eps_end,
        )

    def process_batch(self, group: str, batch: TensorDictBase) -> TensorDictBase:
        keys = list(batch.keys(True, True))
        group_shape = batch.get(group).shape

        nested_done_key = ("next", group, "done")
        nested_terminated_key = ("next", group, "terminated")
        nested_reward_key = ("next", group, "reward")

        if nested_done_key not in keys:
            batch.set(
                nested_done_key,
                batch.get(("next", "done")).unsqueeze(-1).expand((*group_shape, 1)),
            )
        if nested_terminated_key not in keys:
            batch.set(
                nested_terminated_key,
                batch.get(("next", "terminated"))
                .unsqueeze(-1)
                .expand((*group_shape, 1)),
            )

        if nested_reward_key not in keys:
            batch.set(
                nested_reward_key,
                batch.get(("next", "reward")).unsqueeze(-1).expand((*group_shape, 1)),
            )

        return batch

    #####################
    # Custom new methods
    #####################

    def get_value_module(self, group: str) -> TensorDictModule:
        n_agents = len(self.group_map[group])
        modules = []

        modules.append(
            TensorDictModule(
                lambda obs, action: torch.cat([obs, action], dim=-1),
                in_keys=[(group, "observation"), (group, "action")],
                out_keys=[(group, "obs_action")],
            )
        )
        critic_input_spec = CompositeSpec(
            {
                group: CompositeSpec(
                    {
                        "obs_action": UnboundedContinuousTensorSpec(
                            shape=(
                                n_agents,
                                self.observation_spec[group, "observation"].shape[-1]
                                + self.action_spec[group, "action"].shape[-1],
                            )
                        )
                    },
                    shape=(n_agents,),
                )
            }
        )
        critic_output_spec = CompositeSpec(
            {
                group: CompositeSpec(
                    {
                        "state_action_value": UnboundedContinuousTensorSpec(
                            shape=(n_agents, 1)
                        )
                    },
                    shape=(n_agents,),
                )
            }
        )

        modules.append(
            self.critic_model_config.get_model(
                input_spec=critic_input_spec,
                output_spec=critic_output_spec,
                n_agents=n_agents,
                centralised=False,
                input_has_agent_dim=True,
                agent_group=group,
                share_params=self.share_param_critic,
                device=self.device,
                action_spec=self.action_spec,
            )
        )

        return TensorDictSequential(*modules)


@dataclass
class IddpgConfig(AlgorithmConfig):
    """Configuration dataclass for :class:`~benchmarl.algorithms.Iddpg`."""

    share_param_critic: bool = MISSING
    loss_function: str = MISSING
    delay_value: bool = MISSING

    @staticmethod
    def associated_class() -> Type[Algorithm]:
        return Iddpg

    @staticmethod
    def supports_continuous_actions() -> bool:
        return True

    @staticmethod
    def supports_discrete_actions() -> bool:
        return False

    @staticmethod
    def on_policy() -> bool:
        return False