ml-agents/ml-agents/mlagents/trainers/trajectory.py

from typing import List, NamedTuple
import numpy as np

from mlagents.trainers.buffer import AgentBuffer
from mlagents_envs.base_env import ActionTuple, BehaviorSpec, SensorType
from mlagents.trainers.torch.action_log_probs import LogProbsTuple


class AgentExperience(NamedTuple):
    obs: List[np.ndarray]
    reward: float
    done: bool
    action: ActionTuple
    action_probs: LogProbsTuple
    action_mask: np.ndarray
    prev_action: np.ndarray
    interrupted: bool
    memory: np.ndarray


class SplitObservations(NamedTuple):
    vector_observations: np.ndarray
    visual_observations: List[np.ndarray]
    goals: np.ndarray

    @staticmethod
    def from_observations(obs: List[np.ndarray], behavior_spec) -> "SplitObservations":
        """
        Divides a List of numpy arrays into a SplitObservations NamedTuple.
        This allows you to access the vector and visual observations directly,
        without enumerating the list over and over.
        :param obs: List of numpy arrays (observation)
        :returns: A SplitObservations object.
        """
        vis_obs_list: List[np.ndarray] = []
        vec_obs_list: List[np.ndarray] = []
        goal_list: List[np.ndarray] = []
        last_obs = None
        for observation, sensor_type in zip(obs, behavior_spec.sensor_types):
            if sensor_type == SensorType.PARAMETERIZATION:
                goal_list.append(observation)
            elif sensor_type == SensorType.OBSERVATION:
                # Obs could be batched or single
                if len(observation.shape) == 1 or len(observation.shape) == 2:
                    vec_obs_list.append(observation)
                if len(observation.shape) == 3 or len(observation.shape) == 4:
                    vis_obs_list.append(observation)
                last_obs = observation
        if last_obs is not None:
            is_batched = len(last_obs.shape) == 2 or len(last_obs.shape) == 4
            if is_batched:
                vec_obs = (
                    np.concatenate(vec_obs_list, axis=1)
                    if len(vec_obs_list) > 0
                    else np.zeros((last_obs.shape[0], 0), dtype=np.float32)
                )
                goals = (
                    np.concatenate(goal_list, axis=1)
                    if len(goal_list) > 0
                    else np.zeros((last_obs.shape[0], 0), dtype=np.float32)
                )

            else:
                vec_obs = (
                    np.concatenate(vec_obs_list, axis=0)
                    if len(vec_obs_list) > 0
                    else np.array([], dtype=np.float32)
                )
                goals = (
                    np.concatenate(goal_list, axis=0)
                    if len(goal_list) > 0
                    else np.array([], dtype=np.float32)
                )

        else:
            vec_obs = []
        return SplitObservations(
            vector_observations=vec_obs, visual_observations=vis_obs_list, goals=goals
        )


class Trajectory(NamedTuple):
    steps: List[AgentExperience]
    next_obs: List[
        np.ndarray
    ]  # Observation following the trajectory, for bootstrapping
    agent_id: str
    behavior_id: str

    def to_agentbuffer(self, behavior_spec: BehaviorSpec) -> AgentBuffer:
        """
        Converts a Trajectory to an AgentBuffer
        :param trajectory: A Trajectory
        :returns: AgentBuffer. Note that the length of the AgentBuffer will be one
        less than the trajectory, as the next observation need to be populated from the last
        step of the trajectory.
        """
        agent_buffer_trajectory = AgentBuffer()
        vec_vis_obs = SplitObservations.from_observations(
            self.steps[0].obs, behavior_spec
        )
        for step, exp in enumerate(self.steps):
            if step < len(self.steps) - 1:
                next_vec_vis_obs = SplitObservations.from_observations(
                    self.steps[step + 1].obs, behavior_spec
                )
            else:
                next_vec_vis_obs = SplitObservations.from_observations(
                    self.next_obs, behavior_spec
                )

            for i, _ in enumerate(vec_vis_obs.visual_observations):
                agent_buffer_trajectory["visual_obs%d" % i].append(
                    vec_vis_obs.visual_observations[i]
                )
                agent_buffer_trajectory["next_visual_obs%d" % i].append(
                    next_vec_vis_obs.visual_observations[i]
                )
            agent_buffer_trajectory["vector_obs"].append(
                vec_vis_obs.vector_observations
            )

            agent_buffer_trajectory["next_vector_in"].append(
                next_vec_vis_obs.vector_observations
            )
            agent_buffer_trajectory["goals"].append(vec_vis_obs.goals)
            # this shouldnt be necessary in an optimized implementation since the goal does not change
            agent_buffer_trajectory["next_goals"].append(next_vec_vis_obs.goals)

            if exp.memory is not None:
                agent_buffer_trajectory["memory"].append(exp.memory)

            agent_buffer_trajectory["masks"].append(1.0)
            agent_buffer_trajectory["done"].append(exp.done)

            # Adds the log prob and action of continuous/discrete separately
            agent_buffer_trajectory["continuous_action"].append(exp.action.continuous)
            agent_buffer_trajectory["discrete_action"].append(exp.action.discrete)
            agent_buffer_trajectory["continuous_log_probs"].append(
                exp.action_probs.continuous
            )
            agent_buffer_trajectory["discrete_log_probs"].append(
                exp.action_probs.discrete
            )

            # Store action masks if necessary. Note that 1 means active, while
            # in AgentExperience False means active.
            if exp.action_mask is not None:
                mask = 1 - np.concatenate(exp.action_mask)
                agent_buffer_trajectory["action_mask"].append(mask, padding_value=1)
            else:
                # This should never be needed unless the environment somehow doesn't supply the
                # action mask in a discrete space.

                action_shape = exp.action.discrete.shape
                agent_buffer_trajectory["action_mask"].append(
                    np.ones(action_shape, dtype=np.float32), padding_value=1
                )
            agent_buffer_trajectory["prev_action"].append(exp.prev_action)
            agent_buffer_trajectory["environment_rewards"].append(exp.reward)

            # Store the next visual obs as the current
            vec_vis_obs = next_vec_vis_obs
        return agent_buffer_trajectory

    @property
    def done_reached(self) -> bool:
        """
        Returns true if trajectory is terminated with a Done.
        """
        return self.steps[-1].done

    @property
    def interrupted(self) -> bool:
        """
        Returns true if trajectory was terminated because max steps was reached.
        """
        return self.steps[-1].interrupted