def _setup_actor_critic_agent(self, ppo_cfg: Config, ans_cfg: Config) -> None:
        r"""Sets up actor critic and agent for PPO.

        Args:
            ppo_cfg: config node with relevant params
            ans_cfg: config node for ActiveNeuralSLAM model

        Returns:
            None
        """
        try:
            os.mkdir(self.config.TENSORBOARD_DIR)
        except:
            pass
        logger.add_filehandler(os.path.join(self.config.TENSORBOARD_DIR, "run.log"))

        occ_cfg = ans_cfg.OCCUPANCY_ANTICIPATOR
        mapper_cfg = ans_cfg.MAPPER
        # Create occupancy anticipation model
        occupancy_model = OccupancyAnticipator(occ_cfg)
        occupancy_model = OccupancyAnticipationWrapper(
            occupancy_model, mapper_cfg.map_size, (128, 128)
        )
        # Create ANS model
        self.ans_net = ActiveNeuralSLAMNavigator(ans_cfg, occupancy_model)
        self.mapper = self.ans_net.mapper
        self.local_actor_critic = self.ans_net.local_policy
        # Create depth projection model to estimate visible occupancy
        self.depth_projection_net = DepthProjectionNet(
            ans_cfg.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION
        )
        # Set to device
        self.mapper.to(self.device)
        self.local_actor_critic.to(self.device)
        self.depth_projection_net.to(self.device)
示例#2
0
    def __init__(self, config, projection_unit):
        super().__init__()
        self.config = config
        self.map_config = {"size": config.map_size, "scale": config.map_scale}
        V = self.map_config["size"]
        s = self.map_config["scale"]
        self.img_mean_t = rearrange(
            torch.Tensor(self.config.NORMALIZATION.img_mean), "c -> () c () ()"
        )
        self.img_std_t = rearrange(
            torch.Tensor(self.config.NORMALIZATION.img_std), "c -> () c () ()"
        )
        self.pose_estimator = PoseEstimator(
            V,
            self.config.pose_predictor_inputs,
            n_pose_layers=self.config.n_pose_layers,
            n_ensemble_layers=self.config.n_ensemble_layers,
            input_shape=self.config.image_scale_hw,
        )
        self.projection_unit = projection_unit
        if self.config.freeze_projection_unit:
            for p in self.projection_unit.parameters():
                p.requires_grad = False

        self.depth_projection_net = DepthProjectionNet(
            self.projection_unit.main.config.EGO_PROJECTION
        )
        self.gt_map = None
        self.hat_map = None

        # Cache to store pre-computed information
        self._cache = {}
class OccAntNavTrainer(BaseRLTrainer):
    r"""Trainer class for Occupancy Anticipated based navigation.
    This only evaluates the transfer performance of a pre-trained model.
    """
    supported_tasks = ["Nav-v0"]

    def __init__(self, config=None):
        if config is not None:
            self._synchronize_configs(config)
        super().__init__(config)

        # Set pytorch random seed for initialization
        random.seed(config.PYT_RANDOM_SEED)
        np.random.seed(config.PYT_RANDOM_SEED)
        torch.manual_seed(config.PYT_RANDOM_SEED)
        if torch.cuda.is_available():
            torch.backends.cudnn.deterministic = True
            torch.backends.cudnn.benchmark = False

        self.mapper = None
        self.local_actor_critic = None
        self.ans_net = None
        self.planner = None
        self.envs = None
        if config is not None:
            logger.info(f"config: {config}")

    def _synchronize_configs(self, config):
        r"""Matches configs for different parts of the model as well as the simulator.
        """
        config.defrost()
        config.RL.ANS.PLANNER.nplanners = config.NUM_PROCESSES
        config.RL.ANS.MAPPER.thresh_explored = config.RL.ANS.thresh_explored
        config.RL.ANS.pyt_random_seed = config.PYT_RANDOM_SEED
        config.RL.ANS.OCCUPANCY_ANTICIPATOR.pyt_random_seed = config.PYT_RANDOM_SEED
        # Compute the EGO_PROJECTION options based on the
        # depth sensor information and agent parameters.
        map_size = config.RL.ANS.MAPPER.map_size
        map_scale = config.RL.ANS.MAPPER.map_scale
        min_depth = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.MIN_DEPTH
        max_depth = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.MAX_DEPTH
        hfov = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.HFOV
        width = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.WIDTH
        height = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.HEIGHT
        hfov_rad = np.radians(float(hfov))
        vfov_rad = 2 * np.arctan((height / width) * np.tan(hfov_rad / 2.0))
        vfov = np.degrees(vfov_rad).item()
        camera_height = config.TASK_CONFIG.SIMULATOR.DEPTH_SENSOR.POSITION[1]
        height_thresholds = [0.2, 1.5]
        # Set the EGO_PROJECTION options
        ego_proj_config = config.RL.ANS.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION
        ego_proj_config.local_map_shape = (2, map_size, map_size)
        ego_proj_config.map_scale = map_scale
        ego_proj_config.min_depth = min_depth
        ego_proj_config.max_depth = max_depth
        ego_proj_config.hfov = hfov
        ego_proj_config.vfov = vfov
        ego_proj_config.camera_height = camera_height
        ego_proj_config.height_thresholds = height_thresholds
        config.RL.ANS.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION = ego_proj_config
        # Set the correct image scaling values
        config.RL.ANS.MAPPER.image_scale_hw = config.RL.ANS.image_scale_hw
        config.RL.ANS.LOCAL_POLICY.image_scale_hw = config.RL.ANS.image_scale_hw
        # Set the agent dynamics for the local policy
        config.RL.ANS.LOCAL_POLICY.AGENT_DYNAMICS.forward_step = (
            config.TASK_CONFIG.SIMULATOR.FORWARD_STEP_SIZE
        )
        config.RL.ANS.LOCAL_POLICY.AGENT_DYNAMICS.turn_angle = (
            config.TASK_CONFIG.SIMULATOR.TURN_ANGLE
        )
        config.freeze()

    def _setup_actor_critic_agent(self, ppo_cfg: Config, ans_cfg: Config) -> None:
        r"""Sets up actor critic and agent for PPO.

        Args:
            ppo_cfg: config node with relevant params
            ans_cfg: config node for ActiveNeuralSLAM model

        Returns:
            None
        """
        try:
            os.mkdir(self.config.TENSORBOARD_DIR)
        except:
            pass
        logger.add_filehandler(os.path.join(self.config.TENSORBOARD_DIR, "run.log"))

        occ_cfg = ans_cfg.OCCUPANCY_ANTICIPATOR
        mapper_cfg = ans_cfg.MAPPER
        # Create occupancy anticipation model
        occupancy_model = OccupancyAnticipator(occ_cfg)
        occupancy_model = OccupancyAnticipationWrapper(
            occupancy_model, mapper_cfg.map_size, (128, 128)
        )
        # Create ANS model
        self.ans_net = ActiveNeuralSLAMNavigator(ans_cfg, occupancy_model)
        self.mapper = self.ans_net.mapper
        self.local_actor_critic = self.ans_net.local_policy
        # Create depth projection model to estimate visible occupancy
        self.depth_projection_net = DepthProjectionNet(
            ans_cfg.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION
        )
        # Set to device
        self.mapper.to(self.device)
        self.local_actor_critic.to(self.device)
        self.depth_projection_net.to(self.device)

    def save_checkpoint(
        self, file_name: str, extra_state: Optional[Dict] = None
    ) -> None:
        r"""Save checkpoint with specified name.

        Args:
            file_name: file name for checkpoint

        Returns:
            None
        """
        checkpoint = {
            "mapper_state_dict": self.mapper_agent.state_dict(),
            "local_state_dict": self.local_agent.state_dict(),
            "config": self.config,
        }
        if extra_state is not None:
            checkpoint["extra_state"] = extra_state

        torch.save(checkpoint, os.path.join(self.config.CHECKPOINT_FOLDER, file_name))

    def load_checkpoint(self, checkpoint_path: str, *args, **kwargs) -> Dict:
        r"""Load checkpoint of specified path as a dict.

        Args:
            checkpoint_path: path of target checkpoint
            *args: additional positional args
            **kwargs: additional keyword args

        Returns:
            dict containing checkpoint info
        """
        return torch.load(checkpoint_path, *args, **kwargs)

    def _convert_actions_to_delta(self, actions):
        """actions -> torch Tensor
        """
        sim_cfg = self.config.TASK_CONFIG.SIMULATOR
        delta_xyt = torch.zeros(self.envs.num_envs, 3, device=self.device)
        # Forward step
        act_mask = actions.squeeze(1) == 0
        delta_xyt[act_mask, 0] = sim_cfg.FORWARD_STEP_SIZE
        # Turn left
        act_mask = actions.squeeze(1) == 1
        delta_xyt[act_mask, 2] = math.radians(-sim_cfg.TURN_ANGLE)
        # Turn right
        act_mask = actions.squeeze(1) == 2
        delta_xyt[act_mask, 2] = math.radians(sim_cfg.TURN_ANGLE)
        return delta_xyt

    def _remap_actions(self, actions: torch.Tensor) -> torch.Tensor:
        """Converts actions of exploration agent to actions for navigation.
        Remapping:
            0 -> 1 (forward)
            1 -> 2 (turn left)
            2 -> 3 (turn right)
            3 -> 0 (stop)
        """
        actions_rmp = torch.remainder(actions + 1, 4).long()
        return actions_rmp

    def _prepare_batch(self, observations, device=None, actions=None):
        imH, imW = self.config.RL.ANS.image_scale_hw
        device = self.device if device is None else device
        batch = batch_obs(observations, device=device)
        if batch["rgb"].size(1) != imH or batch["rgb"].size(2) != imW:
            rgb = rearrange(batch["rgb"], "b h w c -> b c h w")
            rgb = F.interpolate(rgb, (imH, imW), mode="bilinear")
            batch["rgb"] = rearrange(rgb, "b c h w -> b h w c")
        if batch["depth"].size(1) != imH or batch["depth"].size(2) != imW:
            depth = rearrange(batch["depth"], "b h w c -> b c h w")
            depth = F.interpolate(depth, (imH, imW), mode="nearest")
            batch["depth"] = rearrange(depth, "b c h w -> b h w c")
        # Compute ego_map_gt from depth
        ego_map_gt_b = self.depth_projection_net(
            rearrange(batch["depth"], "b h w c -> b c h w")
        )
        batch["ego_map_gt"] = rearrange(ego_map_gt_b, "b c h w -> b h w c")
        # Add previous action to batch as well
        batch["prev_actions"] = self.prev_actions
        # Add a rough pose estimate if GT pose is not available
        if "pose" not in batch:
            if self.prev_batch is None:
                # Set initial pose estimate to zero
                batch["pose"] = torch.zeros(self.envs.num_envs, 3).to(self.device)
            else:
                actions_delta = self._convert_actions_to_delta(self.prev_actions)
                batch["pose"] = add_pose(self.prev_batch["pose"], actions_delta)

        return batch

    def train(self) -> None:
        r"""Main method for training PPO.

        Returns:
            None
        """
        raise NotImplementedError

    def _eval_checkpoint(
        self,
        checkpoint_path: str,
        writer: TensorboardWriter,
        checkpoint_index: int = 0,
    ) -> None:
        r"""Evaluates a single checkpoint.

        Args:
            checkpoint_path: path of checkpoint
            writer: tensorboard writer object for logging to tensorboard
            checkpoint_index: index of cur checkpoint for logging

        Returns:
            None
        """
        # Map location CPU is almost always better than mapping to a CUDA device.
        ckpt_dict = self.load_checkpoint(checkpoint_path, map_location="cpu")

        if self.config.EVAL.USE_CKPT_CONFIG:
            config = self._setup_eval_config(ckpt_dict["config"])
        else:
            config = self.config.clone()

        ppo_cfg = config.RL.PPO
        ans_cfg = config.RL.ANS

        config.defrost()
        config.TASK_CONFIG.DATASET.SPLIT = config.EVAL.SPLIT
        config.freeze()

        self.envs = construct_envs(config, get_env_class(config.ENV_NAME))
        self._setup_actor_critic_agent(ppo_cfg, ans_cfg)

        # Convert the state_dict of mapper_agent to mapper
        mapper_dict = {
            k.replace("mapper.", ""): v
            for k, v in ckpt_dict["mapper_state_dict"].items()
        }
        # Converting the state_dict of local_agent to just the local_policy.
        local_dict = {
            k.replace("actor_critic.", ""): v
            for k, v in ckpt_dict["local_state_dict"].items()
        }
        # Strict = False is set to ignore to handle the case where
        # pose_estimator is not required.
        self.mapper.load_state_dict(mapper_dict, strict=False)
        self.local_actor_critic.load_state_dict(local_dict)

        # Set models to evaluation
        self.mapper.eval()
        self.local_actor_critic.eval()

        number_of_eval_episodes = self.config.TEST_EPISODE_COUNT
        if number_of_eval_episodes == -1:
            number_of_eval_episodes = sum(self.envs.number_of_episodes)
        else:
            total_num_eps = sum(self.envs.number_of_episodes)
            if total_num_eps < number_of_eval_episodes:
                logger.warn(
                    f"Config specified {number_of_eval_episodes} eval episodes"
                    ", dataset only has {total_num_eps}."
                )
                logger.warn(f"Evaluating with {total_num_eps} instead.")
                number_of_eval_episodes = total_num_eps

        M = ans_cfg.overall_map_size
        V = ans_cfg.MAPPER.map_size
        s = ans_cfg.MAPPER.map_scale
        imH, imW = ans_cfg.image_scale_hw

        assert (
            self.envs.num_envs == 1
        ), "Number of environments needs to be 1 for evaluation"

        # Define metric accumulators
        # Navigation metrics
        navigation_metrics = {
            "success_rate": Metric(),
            "spl": Metric(),
            "distance_to_goal": Metric(),
            "time": Metric(),
            "softspl": Metric(),
        }
        per_difficulty_navigation_metrics = {
            "easy": {
                "success_rate": Metric(),
                "spl": Metric(),
                "distance_to_goal": Metric(),
                "time": Metric(),
                "softspl": Metric(),
            },
            "medium": {
                "success_rate": Metric(),
                "spl": Metric(),
                "distance_to_goal": Metric(),
                "time": Metric(),
                "softspl": Metric(),
            },
            "hard": {
                "success_rate": Metric(),
                "spl": Metric(),
                "distance_to_goal": Metric(),
                "time": Metric(),
                "softspl": Metric(),
            },
        }

        times_per_episode = deque()
        times_per_step = deque()
        # Define a simple function to return episode difficulty based on
        # the geodesic distance
        def classify_difficulty(gd):
            if gd < 5.0:
                return "easy"
            elif gd < 10.0:
                return "medium"
            else:
                return "hard"

        eval_start_time = time.time()
        # Reset environments only for the very first batch
        observations = self.envs.reset()
        for ep in range(number_of_eval_episodes):
            # ============================== Reset agent ==============================
            # Reset agent states
            state_estimates = {
                "pose_estimates": torch.zeros(self.envs.num_envs, 3).to(self.device),
                "map_states": torch.zeros(self.envs.num_envs, 2, M, M).to(self.device),
                "recurrent_hidden_states": torch.zeros(
                    1, self.envs.num_envs, ans_cfg.LOCAL_POLICY.hidden_size
                ).to(self.device),
            }
            # Reset ANS states
            self.ans_net.reset()
            self.not_done_masks = torch.zeros(self.envs.num_envs, 1, device=self.device)
            self.prev_actions = torch.zeros(self.envs.num_envs, 1, device=self.device)
            self.prev_batch = None
            self.ep_time = torch.zeros(self.envs.num_envs, 1, device=self.device)
            # =========================== Episode loop ================================
            ep_start_time = time.time()
            current_episodes = self.envs.current_episodes()
            for ep_step in range(self.config.T_MAX):
                step_start_time = time.time()
                # ============================ Action step ============================
                batch = self._prepare_batch(observations)
                if self.prev_batch is None:
                    self.prev_batch = copy.deepcopy(batch)

                prev_pose_estimates = state_estimates["pose_estimates"]
                with torch.no_grad():
                    (
                        _,
                        _,
                        mapper_outputs,
                        local_policy_outputs,
                        state_estimates,
                    ) = self.ans_net.act(
                        batch,
                        self.prev_batch,
                        state_estimates,
                        self.ep_time,
                        self.not_done_masks,
                        deterministic=ans_cfg.LOCAL_POLICY.deterministic_flag,
                    )
                    actions = local_policy_outputs["actions"]
                    # Make masks not done till reset (end of episode)
                    self.not_done_masks = torch.ones(
                        self.envs.num_envs, 1, device=self.device
                    )
                    self.prev_actions.copy_(actions)

                if ep_step == 0:
                    state_estimates["pose_estimates"].copy_(prev_pose_estimates)

                self.ep_time += 1
                # Update prev batch
                for k, v in batch.items():
                    self.prev_batch[k].copy_(v)

                # Remap actions from exploration to navigation agent.
                actions_rmp = self._remap_actions(actions)

                # =========================== Environment step ========================
                outputs = self.envs.step([a[0].item() for a in actions_rmp])

                observations, _, dones, infos = [list(x) for x in zip(*outputs)]

                times_per_step.append(time.time() - step_start_time)
                # ============================ Process metrics ========================
                if dones[0]:
                    times_per_episode.append(time.time() - ep_start_time)
                    mins_per_episode = np.mean(times_per_episode).item() / 60.0
                    eta_completion = mins_per_episode * (
                        number_of_eval_episodes - ep - 1
                    )
                    secs_per_step = np.mean(times_per_step).item()
                    for i in range(self.envs.num_envs):
                        episode_id = int(current_episodes[i].episode_id)
                        curr_metrics = {
                            "spl": infos[i]["spl"],
                            "softspl": infos[i]["softspl"],
                            "success_rate": infos[i]["success"],
                            "time": ep_step + 1,
                            "distance_to_goal": infos[i]["distance_to_goal"],
                        }
                        # Estimate difficulty of episode
                        episode_difficulty = classify_difficulty(
                            current_episodes[i].info["geodesic_distance"]
                        )
                        for k, v in curr_metrics.items():
                            navigation_metrics[k].update(v, 1.0)
                            per_difficulty_navigation_metrics[episode_difficulty][
                                k
                            ].update(v, 1.0)

                        logger.info(f"====> {ep}/{number_of_eval_episodes} done")
                        for k, v in curr_metrics.items():
                            logger.info(f"{k:25s} : {v:10.3f}")
                        logger.info("{:25s} : {:10d}".format("episode_id", episode_id))
                        logger.info(f"Time per episode: {mins_per_episode:.3f} mins")
                        logger.info(f"Time per step: {secs_per_step:.3f} secs")
                        logger.info(f"ETA: {eta_completion:.3f} mins")

                    # For navigation, terminate episode loop when dones is called
                    break
            # done-for

        if checkpoint_index == 0:
            try:
                eval_ckpt_idx = self.config.EVAL_CKPT_PATH_DIR.split("/")[-1].split(
                    "."
                )[1]
                logger.add_filehandler(
                    f"{self.config.TENSORBOARD_DIR}/navigation_results_ckpt_final_{eval_ckpt_idx}.txt"
                )
            except:
                logger.add_filehandler(
                    f"{self.config.TENSORBOARD_DIR}/navigation_results_ckpt_{checkpoint_index}.txt"
                )
        else:
            logger.add_filehandler(
                f"{self.config.TENSORBOARD_DIR}/navigation_results_ckpt_{checkpoint_index}.txt"
            )

        logger.info(
            f"======= Evaluating over {number_of_eval_episodes} episodes ============="
        )

        logger.info(f"=======> Navigation metrics")
        for k, v in navigation_metrics.items():
            logger.info(f"{k}: {v.get_metric():.3f}")
            writer.add_scalar(f"navigation/{k}", v.get_metric(), checkpoint_index)

        for diff, diff_metrics in per_difficulty_navigation_metrics.items():
            logger.info(f"=============== {diff:^10s} metrics ==============")
            for k, v in diff_metrics.items():
                logger.info(f"{k}: {v.get_metric():.3f}")
                writer.add_scalar(
                    f"{diff}_navigation/{k}", v.get_metric(), checkpoint_index
                )

        total_eval_time = (time.time() - eval_start_time) / 60.0
        logger.info(f"Total evaluation time: {total_eval_time:.3f} mins")
        self.envs.close()
示例#4
0
    def _setup_actor_critic_agent(self, ppo_cfg: Config,
                                  ans_cfg: Config) -> None:
        r"""Sets up actor critic and agent for PPO.

        Args:
            ppo_cfg: config node with relevant params
            ans_cfg: config node for ActiveNeuralSLAM model

        Returns:
            None
        """

        try:
            os.mkdir('video_dir')
            os.mkdir('tb')
            os.mkdir(self.config.TENSORBOARD_DIR)
        except:
            pass
        logger.add_filehandler(
            os.path.join(self.config.TENSORBOARD_DIR, "run.log"))

        occ_cfg = ans_cfg.OCCUPANCY_ANTICIPATOR
        mapper_cfg = ans_cfg.MAPPER
        # Create occupancy anticipation model
        occupancy_model = OccupancyAnticipator(occ_cfg)
        occupancy_model = OccupancyAnticipationWrapper(occupancy_model,
                                                       mapper_cfg.map_size,
                                                       (128, 128))
        # Create ANS model
        self.ans_net = ActiveNeuralSLAMExplorer(ans_cfg, occupancy_model)
        self.mapper = self.ans_net.mapper
        self.local_actor_critic = self.ans_net.local_policy
        self.global_actor_critic = self.ans_net.global_policy
        # Create depth projection model to estimate visible occupancy
        self.depth_projection_net = DepthProjectionNet(
            ans_cfg.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION)
        # Set to device
        self.mapper.to(self.device)
        self.local_actor_critic.to(self.device)
        self.global_actor_critic.to(self.device)
        self.depth_projection_net.to(self.device)

        if ans_cfg.use_ddp:
            self.ans_net.to_ddp()

        # ============================== Create agents ================================
        # Mapper agent
        self.mapper_agent = MapUpdate(
            self.mapper,
            lr=mapper_cfg.lr,
            eps=mapper_cfg.eps,
            label_id=mapper_cfg.label_id,
            max_grad_norm=mapper_cfg.max_grad_norm,
            pose_loss_coef=mapper_cfg.pose_loss_coef,
            occupancy_anticipator_type=ans_cfg.OCCUPANCY_ANTICIPATOR.type,
            freeze_projection_unit=mapper_cfg.freeze_projection_unit,
            num_update_batches=mapper_cfg.num_update_batches,
            batch_size=mapper_cfg.map_batch_size,
            mapper_rollouts=self.mapper_rollouts,
        )
        # Local policy
        if ans_cfg.LOCAL_POLICY.use_heuristic_policy:
            self.local_agent = None
        elif ans_cfg.LOCAL_POLICY.learning_algorithm == "rl":
            self.local_agent = PPO(
                actor_critic=self.local_actor_critic,
                clip_param=ppo_cfg.clip_param,
                ppo_epoch=ppo_cfg.ppo_epoch,
                num_mini_batch=ppo_cfg.num_mini_batch,
                value_loss_coef=ppo_cfg.value_loss_coef,
                entropy_coef=ppo_cfg.local_entropy_coef,
                lr=ppo_cfg.local_policy_lr,
                eps=ppo_cfg.eps,
                max_grad_norm=ppo_cfg.max_grad_norm,
            )
        else:
            self.local_agent = Imitation(
                actor_critic=self.local_actor_critic,
                lr=ppo_cfg.local_policy_lr,
                eps=ppo_cfg.eps,
                max_grad_norm=ppo_cfg.max_grad_norm,
            )
        # Global policy
        self.global_agent = PPO(
            actor_critic=self.global_actor_critic,
            clip_param=ppo_cfg.clip_param,
            ppo_epoch=ppo_cfg.ppo_epoch,
            num_mini_batch=ppo_cfg.num_mini_batch,
            value_loss_coef=ppo_cfg.value_loss_coef,
            entropy_coef=ppo_cfg.entropy_coef,
            lr=ppo_cfg.lr,
            eps=ppo_cfg.eps,
            max_grad_norm=ppo_cfg.max_grad_norm,
        )
        if ans_cfg.model_path != "":
            self.resume_checkpoint(ans_cfg.model_path)
示例#5
0
class LocobotExplorer(BaseRLTrainer):
    def __init__(self, config):
        super().__init__(config)

        # Set pytorch random seed for initialization
        torch.manual_seed(config.PYT_RANDOM_SEED)

        self.mapper = None
        self.local_actor_critic = None
        self.global_actor_critic = None
        self.ans_net = None
        self.planner = None
        self.mapper_agent = None
        self.local_agent = None
        self.global_agent = None
        self.sim = None
        self.logger = logger

        self.device = (torch.device("cuda", self.config.TORCH_GPU_ID)
                       if torch.cuda.is_available() else torch.device("cpu"))

        self.ACT_2_COMMAND = {
            0: (
                "go_to_relative",
                {
                    "xyt_position": [
                        config.RL.ANS.LOCAL_POLICY.AGENT_DYNAMICS.forward_step,
                        0, 0
                    ],
                    "use_map":
                    False,
                    "close_loop":
                    True,
                    "smooth":
                    False,
                },
            ),
            1: (
                "go_to_relative",
                {
                    "xyt_position": [
                        0, 0,
                        (config.RL.ANS.LOCAL_POLICY.AGENT_DYNAMICS.turn_angle /
                         180) * np.pi
                    ],
                    "use_map":
                    False,
                    "close_loop":
                    True,
                    "smooth":
                    False,
                },
            ),
            2: (
                "go_to_relative",
                {
                    "xyt_position": [
                        0, 0,
                        (-config.RL.ANS.LOCAL_POLICY.AGENT_DYNAMICS.turn_angle
                         / 180) * np.pi
                    ],
                    "use_map":
                    False,
                    "close_loop":
                    True,
                    "smooth":
                    False,
                },
            ),
        }

        self.ACT_2_NAME = {0: 'MOVE_FORWARD', 1: 'TURN_LEFT', 2: 'TURN_RIGHT'}

    def _eval(self):

        start_time = time.time()

        if self.config.MANUAL_COMMANDS:
            init_time = None
            manual_step_start_time = None
            total_manual_time = 0.0

        checkpoint_index = int(
            (re.findall('\d+', self.config.EVAL_CKPT_PATH_DIR))[-1])
        ckpt_dict = torch.load(self.config.EVAL_CKPT_PATH_DIR,
                               map_location="cpu")

        print(
            f'Number of steps of the ckpt: {ckpt_dict["extra_state"]["step"]}')

        config = self._setup_config(ckpt_dict)
        ppo_cfg = config.RL.PPO
        ans_cfg = config.RL.ANS

        self.mapper_rollouts = None
        self._setup_actor_critic_agent(ppo_cfg, ans_cfg)

        self.mapper_agent.load_state_dict(ckpt_dict["mapper_state_dict"])
        if self.local_agent is not None:
            self.local_agent.load_state_dict(ckpt_dict["local_state_dict"])
            self.local_actor_critic = self.local_agent.actor_critic
        else:
            self.local_actor_critic = self.ans_net.local_policy
        self.global_agent.load_state_dict(ckpt_dict["global_state_dict"])
        self.mapper = self.mapper_agent.mapper
        self.global_actor_critic = self.global_agent.actor_critic

        # Set models to evaluation
        self.mapper.eval()
        self.local_actor_critic.eval()
        self.global_actor_critic.eval()

        M = ans_cfg.overall_map_size
        V = ans_cfg.MAPPER.map_size
        s = ans_cfg.MAPPER.map_scale
        imH, imW = ans_cfg.image_scale_hw

        num_steps = self.config.T_EXP

        prev_action = torch.zeros(1, 1, device=self.device, dtype=torch.long)
        masks = torch.zeros(1, 1, device=self.device)

        try:
            self.sim = make_sim('PyRobot-v1',
                                config=self.config.TASK_CONFIG.PYROBOT)
        except (KeyboardInterrupt, SystemExit):
            sys.exit()

        pose = defaultdict()
        self.sim._robot.camera.set_tilt(math.radians(self.config.CAMERA_TILT),
                                        wait=True)
        print(
            f"\nStarting Camera State: {self.sim.get_agent_state()['camera']}")
        print(f"Starting Agent State: {self.sim.get_agent_state()['base']}")
        obs = [self.sim.reset()]

        if self.config.SAVE_OBS_IMGS:
            cv2.imwrite(f'obs/depth_dirty_s.jpg', obs[0]['depth'] * 255.0)

        obs[0]['depth'][..., 0] = self._correct_depth(obs, -1)

        if self.config.SAVE_OBS_IMGS:
            cv2.imwrite(f'obs/rgb_s.jpg', obs[0]['rgb'][:, :, ::-1])
            cv2.imwrite(f'depth_s.jpg', obs[0]['depth'] * 255.0)

        starting_agent_state = self.sim.get_agent_state()
        locobot2relative = CoordProjection(starting_agent_state['base'])
        pose['base'] = locobot2relative(starting_agent_state['base'])

        print(f"Starting Agent Pose: {pose['base']}\n")
        batch = self._prepare_batch(obs, -1, device=self.device)
        if ans_cfg.MAPPER.use_sensor_positioning:
            batch['pose'] = pose['base'].to(self.device)
            batch['pose'][0][1:] = -batch['pose'][0][1:]
        prev_batch = batch

        num_envs = self.config.NUM_PROCESSES
        agent_poses_over_time = []
        for i in range(num_envs):
            agent_poses_over_time.append(
                torch.tensor([(M - 1) / 2, (M - 1) / 2, 0]))
        state_estimates = {
            "pose_estimates":
            torch.zeros(num_envs, 3).to(self.device),
            "map_states":
            torch.zeros(num_envs, 2, M, M).to(self.device),
            "recurrent_hidden_states":
            torch.zeros(1, num_envs,
                        ans_cfg.LOCAL_POLICY.hidden_size).to(self.device),
            "visited_states":
            torch.zeros(num_envs, 1, M, M).to(self.device),
        }
        ground_truth_states = {
            "visible_occupancy": torch.zeros(num_envs, 2, M,
                                             M).to(self.device),
            "pose": torch.zeros(num_envs, 3).to(self.device),
            "environment_layout": torch.zeros(num_envs, 2, M,
                                              M).to(self.device)
        }

        # Reset ANS states
        self.ans_net.reset()

        # Frames for video creation
        rgb_frames = []
        if len(self.config.VIDEO_OPTION) > 0:
            os.makedirs(self.config.VIDEO_DIR, exist_ok=True)

        step_start_time = time.time()

        for i in range(num_steps):
            print(
                f"\n\n---------------------------------------------------<<< STEP {i} >>>---------------------------------------------------"
            )
            ep_time = torch.zeros(num_envs, 1, device=self.device).fill_(i)

            (
                mapper_inputs,
                local_policy_inputs,
                global_policy_inputs,
                mapper_outputs,
                local_policy_outputs,
                global_policy_outputs,
                state_estimates,
                intrinsic_rewards,
            ) = self.ans_net.act(
                batch,
                prev_batch,
                state_estimates,
                ep_time,
                masks,
                deterministic=True,
            )
            if self.config.SAVE_MAP_IMGS:
                cv2.imwrite(
                    f'maps/test_map_{i - 1}.jpg',
                    self._round_map(state_estimates['map_states']) * 255)

            action = local_policy_outputs["actions"][0][0]

            distance2ggoal = torch.norm(
                mapper_outputs['curr_map_position'] -
                self.ans_net.states["curr_global_goals"],
                dim=1) * s

            print(f"Distance to Global Goal: {distance2ggoal}")

            reached_flag = distance2ggoal < ans_cfg.goal_success_radius

            if self.config.MANUAL_COMMANDS:
                if init_time is None:
                    init_time = time.time() - start_time
                    total_manual_time = total_manual_time + init_time
                if manual_step_start_time is not None:
                    manual_step_time = time.time() - manual_step_start_time
                    total_manual_time = total_manual_time + manual_step_time
                action = torch.tensor(
                    int(input('Waiting input to start new action: ')))
                manual_step_start_time = time.time()

                if action.item() == 3:
                    reached_flag = True

            prev_action.copy_(action)

            if not reached_flag and action.item() != 3:
                print(f'Doing Env Step [{self.ACT_2_NAME[action.item()]}]...')
                action_command = self.ACT_2_COMMAND[action.item()]

                obs = self._do_action(action_command)

                if self.config.SAVE_OBS_IMGS:
                    cv2.imwrite(f'obs/depth_dirty_{i}.jpg',
                                obs[0]['depth'] * 255.0)

                # Correcting invalid depth pixels
                obs[0]['depth'][..., 0] = self._correct_depth(obs, i)

                if self.config.SAVE_OBS_IMGS:
                    cv2.imwrite(f'obs/rgb_{i}.jpg', obs[0]['rgb'][:, :, ::-1])
                    cv2.imwrite(f'obs/depth_{i}.jpg', obs[0]['depth'] * 255.0)

                agent_state = self.sim.get_agent_state()
                prev_batch = batch
                batch = self._prepare_batch(obs, i, device=self.device)

                pose = defaultdict()
                pose['base'] = locobot2relative(agent_state['base'])

                if ans_cfg.MAPPER.use_sensor_positioning:
                    batch['pose'] = pose['base'].to(self.device)
                    batch['pose'][0][1:] = -batch['pose'][0][1:]

                map_coords = convert_world2map(
                    batch['pose'], (M, M), ans_cfg.OCCUPANCY_ANTICIPATOR.
                    EGO_PROJECTION.map_scale).squeeze()
                map_coords = torch.cat(
                    (map_coords, batch['pose'][0][-1].reshape(1)))
                if self.config.COORD_DEBUG:
                    print('COORDINATES CHECK')
                    print(
                        f'Starting Agent State: {starting_agent_state["base"]}'
                    )
                    print(f'Current Agent State: {agent_state["base"]}')
                    print(
                        f'Current Sim Agent State: {self.sim.get_agent_state()["base"]}'
                    )
                    print(f'Current Global Coords: {batch["pose"]}')
                    print(f'Current Map Coords: {map_coords}')
                agent_poses_over_time.append(map_coords)

                step_time = time.time() - step_start_time
                print(f"\nStep Time: {step_time}")
                step_start_time = time.time()

            # Create new frame of the video
            if (len(self.config.VIDEO_OPTION) > 0):
                frame = observations_to_image(
                    obs[0],
                    observation_size=300,
                    collision_flag=self.config.DRAW_COLLISIONS)
                # Add ego_map_gt to frame
                ego_map_gt_i = asnumpy(batch["ego_map_gt"][0])  # (2, H, W)
                ego_map_gt_i = convert_gt2channel_to_gtrgb(ego_map_gt_i)
                ego_map_gt_i = cv2.resize(ego_map_gt_i, (300, 300))
                # frame = np.concatenate([frame], axis=1)
                # Generate ANS specific visualizations
                environment_layout = asnumpy(
                    ground_truth_states["environment_layout"][0])  # (2, H, W)
                visible_occupancy = mapper_outputs["gt_mt"][0].cpu().numpy(
                )  # (2, H, W)
                anticipated_occupancy = mapper_outputs["hat_mt"][0].cpu(
                ).numpy()  # (2, H, W)

                H = frame.shape[0]
                visible_occupancy_vis = generate_topdown_allocentric_map(
                    environment_layout,
                    visible_occupancy,
                    agent_poses_over_time,
                    thresh_explored=ans_cfg.thresh_explored,
                    thresh_obstacle=ans_cfg.thresh_obstacle,
                    zoom=False)
                visible_occupancy_vis = cv2.resize(visible_occupancy_vis,
                                                   (H, H))
                anticipated_occupancy_vis = generate_topdown_allocentric_map(
                    environment_layout,
                    anticipated_occupancy,
                    agent_poses_over_time,
                    thresh_explored=ans_cfg.thresh_explored,
                    thresh_obstacle=ans_cfg.thresh_obstacle,
                    zoom=False)
                anticipated_occupancy_vis = cv2.resize(
                    anticipated_occupancy_vis, (H, H))
                anticipated_action_map = generate_topdown_allocentric_map(
                    environment_layout,
                    anticipated_occupancy,
                    agent_poses_over_time,
                    zoom=False,
                    thresh_explored=ans_cfg.thresh_explored,
                    thresh_obstacle=ans_cfg.thresh_obstacle,
                )
                global_goals = self.ans_net.states["curr_global_goals"]
                local_goals = self.ans_net.states["curr_local_goals"]
                if global_goals is not None:
                    cX = int(global_goals[0, 0].item())
                    cY = int(global_goals[0, 1].item())
                    anticipated_action_map = cv2.circle(
                        anticipated_action_map,
                        (cX, cY),
                        10,
                        (255, 0, 0),
                        -1,
                    )
                if local_goals is not None:
                    cX = int(local_goals[0, 0].item())
                    cY = int(local_goals[0, 1].item())
                    anticipated_action_map = cv2.circle(
                        anticipated_action_map,
                        (cX, cY),
                        10,
                        (0, 255, 255),
                        -1,
                    )
                anticipated_action_map = cv2.resize(anticipated_action_map,
                                                    (H, H))

                maps_vis = np.concatenate(
                    [
                        visible_occupancy_vis, anticipated_occupancy_vis,
                        anticipated_action_map, ego_map_gt_i
                    ],
                    axis=1,
                )

                if self.config.RL.ANS.overall_map_size == 2001 or self.config.RL.ANS.overall_map_size == 961:
                    if frame.shape[1] < maps_vis.shape[1]:
                        diff = maps_vis.shape[1] - frame.shape[1]
                        npad = ((0, 0), (diff // 2, diff // 2), (0, 0))
                        frame = np.pad(frame,
                                       pad_width=npad,
                                       mode='constant',
                                       constant_values=0)
                    elif frame.shape[1] > maps_vis.shape[1]:
                        diff = frame.shape[1] - maps_vis.shape[1]
                        npad = ((0, 0), (diff // 2, diff // 2), (0, 0))
                        maps_vis = np.pad(maps_vis,
                                          pad_width=npad,
                                          mode='constant',
                                          constant_values=0)
                frame = np.concatenate([frame, maps_vis], axis=0)
                rgb_frames.append(frame)
                if self.config.SAVE_VIDEO_IMGS:
                    try:
                        os.mkdir("fig1")
                    except:
                        pass
                    print("Saved imgs for Fig. 1!")
                    cv2.imwrite(f'fig1/rgb_{step_start_time}.jpg',
                                obs[0]['rgb'][:, :, ::-1])
                    cv2.imwrite(f'fig1/depth_{step_start_time}.jpg',
                                obs[0]['depth'] * 255.0)
                    cv2.imwrite(f'fig1/aap_{step_start_time}.jpg',
                                anticipated_action_map[..., ::-1])
                    cv2.imwrite(f'fig1/em_{step_start_time}.jpg',
                                ego_map_gt_i[..., ::-1])
                if self.config.DEBUG_VIDEO_FRAME:
                    cv2.imwrite('last_frame.jpg', frame)

                if reached_flag:
                    for f in range(20):
                        rgb_frames.append(frame)

                # Video creation
                video_dict = {"t": start_time}
                if (i + 1) % 10 == 0 or reached_flag:
                    generate_video(
                        video_option=self.config.VIDEO_OPTION,
                        video_dir=self.config.VIDEO_DIR,
                        images=rgb_frames,
                        episode_id=0,
                        checkpoint_idx=checkpoint_index,
                        metrics=video_dict,
                        tb_writer=TensorboardWriter('tb/locobot'),
                    )

            if reached_flag:
                if self.config.MANUAL_COMMANDS:
                    manual_step_time = time.time() - manual_step_start_time
                    total_manual_time = total_manual_time + manual_step_time
                    print(f"Manual elapsed time: {total_manual_time}")

                print(f"Number of steps: {i + 1}")
                print(f"Elapsed time: {time.time() - start_time}")
                print(f"Final Distance to Goal: {distance2ggoal}")
                if "bump" in obs[0]:
                    print(f"Collision: {obs[0]['bump']}")
                print("Exiting...")
                break
        return

    def _correct_depth(self, obs, i):
        # Inpainting, median blur and border replaced
        mask = (obs[0]['depth'] <= 0).astype(np.uint8)
        mask_dilated = cv2.morphologyEx(mask,
                                        cv2.MORPH_DILATE,
                                        np.ones((3, 3)),
                                        iterations=3)
        corrected_depth = (cv2.inpaint(
            (obs[0]['depth'] * 255.0).astype(np.uint16), mask_dilated, 5,
            cv2.INPAINT_TELEA)).astype(np.float32) / 255.0
        median_depth = cv2.medianBlur(corrected_depth, 5)
        removed_border = median_depth[1:-1, 1:-1]
        final_depth = cv2.copyMakeBorder(removed_border, 1, 1, 1, 1,
                                         cv2.BORDER_REFLECT)
        return final_depth

    def _setup_actor_critic_agent(self, ppo_cfg: Config,
                                  ans_cfg: Config) -> None:
        r"""Sets up actor critic and agent for PPO.

        Args:
            ppo_cfg: config node with relevant params
            ans_cfg: config node for ActiveNeuralSLAM model

        Returns:
            None
        """

        try:
            os.mkdir('video_dir')
            os.mkdir('tb')
            os.mkdir(self.config.TENSORBOARD_DIR)
        except:
            pass
        logger.add_filehandler(
            os.path.join(self.config.TENSORBOARD_DIR, "run.log"))

        occ_cfg = ans_cfg.OCCUPANCY_ANTICIPATOR
        mapper_cfg = ans_cfg.MAPPER
        # Create occupancy anticipation model
        occupancy_model = OccupancyAnticipator(occ_cfg)
        occupancy_model = OccupancyAnticipationWrapper(occupancy_model,
                                                       mapper_cfg.map_size,
                                                       (128, 128))
        # Create ANS model
        self.ans_net = ActiveNeuralSLAMExplorer(ans_cfg, occupancy_model)
        self.mapper = self.ans_net.mapper
        self.local_actor_critic = self.ans_net.local_policy
        self.global_actor_critic = self.ans_net.global_policy
        # Create depth projection model to estimate visible occupancy
        self.depth_projection_net = DepthProjectionNet(
            ans_cfg.OCCUPANCY_ANTICIPATOR.EGO_PROJECTION)
        # Set to device
        self.mapper.to(self.device)
        self.local_actor_critic.to(self.device)
        self.global_actor_critic.to(self.device)
        self.depth_projection_net.to(self.device)

        if ans_cfg.use_ddp:
            self.ans_net.to_ddp()

        # ============================== Create agents ================================
        # Mapper agent
        self.mapper_agent = MapUpdate(
            self.mapper,
            lr=mapper_cfg.lr,
            eps=mapper_cfg.eps,
            label_id=mapper_cfg.label_id,
            max_grad_norm=mapper_cfg.max_grad_norm,
            pose_loss_coef=mapper_cfg.pose_loss_coef,
            occupancy_anticipator_type=ans_cfg.OCCUPANCY_ANTICIPATOR.type,
            freeze_projection_unit=mapper_cfg.freeze_projection_unit,
            num_update_batches=mapper_cfg.num_update_batches,
            batch_size=mapper_cfg.map_batch_size,
            mapper_rollouts=self.mapper_rollouts,
        )
        # Local policy
        if ans_cfg.LOCAL_POLICY.use_heuristic_policy:
            self.local_agent = None
        elif ans_cfg.LOCAL_POLICY.learning_algorithm == "rl":
            self.local_agent = PPO(
                actor_critic=self.local_actor_critic,
                clip_param=ppo_cfg.clip_param,
                ppo_epoch=ppo_cfg.ppo_epoch,
                num_mini_batch=ppo_cfg.num_mini_batch,
                value_loss_coef=ppo_cfg.value_loss_coef,
                entropy_coef=ppo_cfg.local_entropy_coef,
                lr=ppo_cfg.local_policy_lr,
                eps=ppo_cfg.eps,
                max_grad_norm=ppo_cfg.max_grad_norm,
            )
        else:
            self.local_agent = Imitation(
                actor_critic=self.local_actor_critic,
                lr=ppo_cfg.local_policy_lr,
                eps=ppo_cfg.eps,
                max_grad_norm=ppo_cfg.max_grad_norm,
            )
        # Global policy
        self.global_agent = PPO(
            actor_critic=self.global_actor_critic,
            clip_param=ppo_cfg.clip_param,
            ppo_epoch=ppo_cfg.ppo_epoch,
            num_mini_batch=ppo_cfg.num_mini_batch,
            value_loss_coef=ppo_cfg.value_loss_coef,
            entropy_coef=ppo_cfg.entropy_coef,
            lr=ppo_cfg.lr,
            eps=ppo_cfg.eps,
            max_grad_norm=ppo_cfg.max_grad_norm,
        )
        if ans_cfg.model_path != "":
            self.resume_checkpoint(ans_cfg.model_path)

    def _setup_config(self, ckpt_dict):
        if self.config.EVAL.USE_CKPT_CONFIG:
            config = self._setup_eval_config(ckpt_dict["config"])
        else:
            config = self.config.clone()
        config.defrost()
        config.TASK_CONFIG.DATASET.SPLIT = config.EVAL.SPLIT
        config.freeze()
        if "COLLISION_SENSOR" not in config.TASK_CONFIG.TASK.SENSORS:
            config.TASK_CONFIG.TASK.SENSORS.append("COLLISION_SENSOR")
        if len(self.config.VIDEO_OPTION) > 0:
            config.defrost()
            config.TASK_CONFIG.TASK.MEASUREMENTS.append("TOP_DOWN_MAP_EXP")
            config.TASK_CONFIG.TASK.MEASUREMENTS.append("COLLISIONS")
            config.freeze()
        self.logger.info(f"env config: {config}")
        return config

    def _setup_eval_config(self, checkpoint_config: Config) -> Config:
        r"""Sets up and returns a merged config for evaluation. Config
            object saved from checkpoint is merged into config file specified
            at evaluation time with the following overwrite priority:
                  eval_opts > ckpt_opts > eval_cfg > ckpt_cfg
            If the saved config is outdated, only the eval config is returned.

        Args:
            checkpoint_config: saved config from checkpoint.

        Returns:
            Config: merged config for eval.
        """

        config = self.config.clone()
        config.defrost()

        ckpt_cmd_opts = checkpoint_config.CMD_TRAILING_OPTS
        eval_cmd_opts = config.CMD_TRAILING_OPTS

        try:
            config.merge_from_other_cfg(checkpoint_config)
            config.merge_from_other_cfg(self.config)
            config.merge_from_list(ckpt_cmd_opts)
            config.merge_from_list(eval_cmd_opts)
        except KeyError:
            logger.info("Saved config is outdated, using solely eval config")
            config = self.config.clone()
            config.merge_from_list(eval_cmd_opts)
        if config.TASK_CONFIG.DATASET.SPLIT == "train":
            config.TASK_CONFIG.defrost()
            config.TASK_CONFIG.DATASET.SPLIT = "val"

        config.TASK_CONFIG.SIMULATOR.AGENT_0.SENSORS = self.config.SENSORS
        config.freeze()

        return config

    def _prepare_batch(self, observations, i, device=None, actions=None):
        imH, imW = self.config.RL.ANS.image_scale_hw
        device = self.device if device is None else device

        batch = batch_obs(observations, device=device)

        if batch["rgb"].size(1) != imH or batch["rgb"].size(2) != imW:
            rgb = rearrange(batch["rgb"], "b h w c -> b c h w")
            rgb = F.interpolate(rgb, (imH, imW), mode="bilinear")
            batch["rgb"] = rearrange(rgb, "b c h w -> b h w c")
        if batch["depth"].size(1) != imH or batch["depth"].size(2) != imW:
            depth = rearrange(batch["depth"], "b h w c -> b c h w")
            depth = F.interpolate(depth, (imH, imW), mode="bilinear")
            batch["depth"] = rearrange(depth, "b c h w -> b h w c")

        # Compute ego_map_gt from depth
        ego_map_gt_b = self.depth_projection_net(
            rearrange(batch["depth"], "b h w c -> b c h w"))
        batch["ego_map_gt"] = rearrange(ego_map_gt_b, "b c h w -> b h w c")

        if actions is None:
            batch["prev_actions"] = torch.zeros(1, 1).to(self.device)
        else:
            batch["prev_actions"] = actions

        return batch

    def _do_action(self, action_command):
        obs = [self.sim.step(action_command[0], action_command[1])]
        return obs

    def _round_map(self, sem_map):
        new_map = sem_map.cpu().numpy()[0, 0]
        new_map[new_map >= 0.5] = 1.0
        new_map[new_map < 0.5] = 0.0
        return new_map