def merge_sim_episode_config(
sim_config: Config, episode: Type[Episode]
) -> Any:
sim_config.defrost()
# here's where the scene update happens, extract the scene name out of the path
sim_config.SCENE = episode.scene_id
sim_config.freeze()
if (
episode.start_position is not None
and episode.start_rotation is not None
):
agent_name = sim_config.AGENTS[sim_config.DEFAULT_AGENT_ID]
agent_cfg = getattr(sim_config, agent_name)
agent_cfg.defrost()
agent_cfg.START_POSITION = episode.start_position
agent_cfg.START_ROTATION = episode.start_rotation
agent_cfg.GOAL_POSITION = episode.goals[0].position
agent_cfg.SOUND_ID = episode.sound_id
agent_cfg.DISTRACTOR_SOUND_ID = episode.distractor_sound_id
agent_cfg.DISTRACTOR_POSITION_INDEX = episode.distractor_position_index
agent_cfg.OFFSET = episode.offset
agent_cfg.DURATION = episode.duration
agent_cfg.IS_SET_START_STATE = True
agent_cfg.freeze()
return sim_config
def get_default_config():
c = Config()
c.INPUT_TYPE = "blind"
c.MODEL_PATH = "data/checkpoints/blind.pth"
c.RESOLUTION = 256
c.HIDDEN_SIZE = 512
c.RANDOM_SEED = 7
c.PTH_GPU_ID = 0
c.GOAL_SENSOR_UUID = "pointgoal"
return c
def get_default_config() -> Config:
c = Config()
c.INPUT_TYPE = "rgb"
c.MODEL_PATH = "data/checkpoints/gibson-rgb-best.pth"
c.RESOLUTION = 256
c.HIDDEN_SIZE = 512
c.RANDOM_SEED = 7
c.PTH_GPU_ID = 0
c.GOAL_SENSOR_UUID = "pointgoal_with_gps_compass"
return c
def merge_sim_episode_config(sim_config: Config, episode: Episode) -> Any:
sim_config.defrost()
sim_config.SCENE = episode.scene_id
sim_config.freeze()
if (episode.start_position is not None
and episode.start_rotation is not None):
agent_name = sim_config.AGENTS[sim_config.DEFAULT_AGENT_ID]
agent_cfg = getattr(sim_config, agent_name)
agent_cfg.defrost()
agent_cfg.START_POSITION = episode.start_position
agent_cfg.START_ROTATION = episode.start_rotation
agent_cfg.IS_SET_START_STATE = True
agent_cfg.freeze()
return sim_config
def __init__(self, config: Config) -> None:
"""Constructor
:param config: config for the environment. Should contain id for
simulator and ``task_name`` which are passed into ``make_sim`` and
``make_task``.
:param dataset: reference to dataset for task instance level
information. Can be defined as :py:`None` in which case
``_episodes`` should be populated from outside.
"""
assert config.is_frozen(), ("Freeze the config before creating the "
"environment, use config.freeze().")
self._config = config
self._current_episode_index = None
self._current_episode = None
self._scenes = config.DATASET.CONTENT_SCENES
self._swap_building_every = config.ENVIRONMENT.ITERATOR_OPTIONS.MAX_SCENE_REPEAT_EPISODES
self._current_scene_episode_idx = 0
self._current_scene_idx = 0
self._config.defrost()
self._config.SIMULATOR.SCENE = os.path.join(
config.DATASET.SCENES_DIR,
'mp3d/{}/{}.glb'.format(self._scenes, self._scenes))
self._config.freeze()
self._sim = make_sim(id_sim=self._config.SIMULATOR.TYPE,
config=self._config.SIMULATOR)
def __init__(self,
config: Config,
dataset: Optional[Dataset] = None) -> None:
assert config.is_frozen(), ("Freeze the config before creating the "
"environment, use config.freeze()")
self._config = config
self._dataset = dataset
self._current_episode_index = None
if self._dataset is None and config.DATASET.TYPE:
self._dataset = make_dataset(id_dataset=config.DATASET.TYPE,
config=config.DATASET)
self._episodes = self._dataset.episodes if self._dataset else []
self._sim = make_sim(id_sim=self._config.SIMULATOR.TYPE,
config=self._config.SIMULATOR)
self._task = make_task(
self._config.TASK.TYPE,
task_config=self._config.TASK,
sim=self._sim,
dataset=dataset,
)
self.observation_space = SpaceDict({
**self._sim.sensor_suite.observation_spaces.spaces,
**self._task.sensor_suite.observation_spaces.spaces,
})
self.action_space = self._sim.action_space
self._max_episode_seconds = (
self._config.ENVIRONMENT.MAX_EPISODE_SECONDS)
self._max_episode_steps = self._config.ENVIRONMENT.MAX_EPISODE_STEPS
self._elapsed_steps = 0
self._episode_start_time: Optional[float] = None
self._episode_over = False
def get_scenes_to_load(cls, config: Config) -> List[str]:
r"""Return list of scene ids for which dataset has separate files with
episodes.
"""
assert cls.check_config_paths_exist(config)
dataset_dir = os.path.dirname(
config.DATA_PATH.format(split=config.SPLIT))
cfg = config.clone()
cfg.defrost()
cfg.CONTENT_SCENES = []
dataset = cls(cfg)
has_individual_scene_files = os.path.exists(
dataset.content_scenes_path.split("{scene}")[0].format(
data_path=dataset_dir))
if has_individual_scene_files:
return cls._get_scenes_from_folder(
content_scenes_path=dataset.content_scenes_path,
dataset_dir=dataset_dir,
)
else:
# Load the full dataset, things are not split into separate files
cfg.CONTENT_SCENES = [ALL_SCENES_MASK]
dataset = cls(cfg)
return list(map(cls.scene_from_scene_path, dataset.scene_ids))
def merge_sim_episode_config(sim_config: Config,
episode: Type[Episode]) -> Any:
sim_config.defrost()
# here's where the scene update happens, extract the scene name out of the path
sim_config.SCENE = episode.scene_id
sim_config.freeze()
if (episode.start_position is not None
and episode.start_rotation is not None):
agent_name = sim_config.AGENTS[sim_config.DEFAULT_AGENT_ID]
agent_cfg = getattr(sim_config, agent_name)
agent_cfg.defrost()
agent_cfg.START_POSITION = episode.start_position
agent_cfg.START_ROTATION = episode.start_rotation
agent_cfg.GOAL_POSITION = episode.goals[0].position
agent_cfg.SOUND = episode.info['sound']
agent_cfg.IS_SET_START_STATE = True
agent_cfg.freeze()
return sim_config
def __init__(self,
config: Config,
dataset: Optional[Dataset] = None) -> None:
"""Constructor
:param config: config for the environment. Should contain id for
simulator and ``task_name`` which are passed into ``make_sim`` and
``make_task``.
:param dataset: reference to dataset for task instance level
information. Can be defined as :py:`None` in which case
``_episodes`` should be populated from outside.
"""
assert config.is_frozen(), ("Freeze the config before creating the "
"environment, use config.freeze().")
self._config = config
self._dataset = dataset
self._current_episode_index = None
if self._dataset is None and config.DATASET.TYPE:
self._dataset = make_dataset(id_dataset=config.DATASET.TYPE,
config=config.DATASET)
self._episodes = self._dataset.episodes if self._dataset else []
self._current_episode = None
iter_option_dict = {
k.lower(): v
for k, v in config.ENVIRONMENT.ITERATOR_OPTIONS.items()
}
self._episode_iterator = self._dataset.get_episode_iterator(
**iter_option_dict)
# load the first scene if dataset is present
if self._dataset:
assert (len(self._dataset.episodes) >
0), "dataset should have non-empty episodes list"
self._config.defrost()
self._config.SIMULATOR.SCENE = self._dataset.episodes[0].scene_id
self._config.freeze()
self._sim = make_sim(id_sim=self._config.SIMULATOR.TYPE,
config=self._config.SIMULATOR)
self._task = make_task(
self._config.TASK.TYPE,
config=self._config.TASK,
sim=self._sim,
dataset=self._dataset,
)
self.observation_space = SpaceDict({
**self._sim.sensor_suite.observation_spaces.spaces,
**self._task.sensor_suite.observation_spaces.spaces,
})
self.action_space = self._task.action_space
self._max_episode_seconds = (
self._config.ENVIRONMENT.MAX_EPISODE_SECONDS)
self._max_episode_steps = self._config.ENVIRONMENT.MAX_EPISODE_STEPS
self._elapsed_steps = 0
self._episode_start_time: Optional[float] = None
self._episode_over = False
def get_scenes_to_load(config: Config) -> List[str]:
r"""Return list of scene ids for which dataset has separate files with
episodes.
"""
assert PointNavDatasetV1.check_config_paths_exist(config)
dataset_dir = os.path.dirname(
config.POINTNAVV1.DATA_PATH.format(split=config.SPLIT))
cfg = config.clone()
cfg.defrost()
cfg.POINTNAVV1.CONTENT_SCENES = []
dataset = PointNavDatasetV1(cfg)
return PointNavDatasetV1._get_scenes_from_folder(
content_scenes_path=dataset.content_scenes_path,
dataset_dir=dataset_dir,
)
def __init__(self,
config: Config,
dataset: Optional[Dataset] = None) -> None:
assert config.is_frozen(), ("Freeze the config before creating the "
"environment, use config.freeze().")
self._config = config
self._dataset = dataset
self._current_episode_index = None
if self._dataset is None and config.DATASET.TYPE:
self._dataset = make_dataset(id_dataset=config.DATASET.TYPE,
config=config.DATASET)
self._episodes = self._dataset.episodes if self._dataset else []
self._current_episode = None
iter_option_dict = {
k.lower(): v
for k, v in config.ENVIRONMENT.ITERATOR_OPTIONS.items()
}
self._episode_iterator = self._dataset.get_episode_iterator(
**iter_option_dict)
# load the first scene if dataset is present
if self._dataset:
assert (len(self._dataset.episodes) >
0), "dataset should have non-empty episodes list"
self._config.defrost()
self._config.SIMULATOR.SCENE = self._dataset.episodes[0].scene_id
self._config.freeze()
self._sim = make_sim(id_sim=self._config.SIMULATOR.TYPE,
config=self._config.SIMULATOR)
self._task = make_task(
self._config.TASK.TYPE,
task_config=self._config.TASK,
sim=self._sim,
dataset=self._dataset,
)
self.observation_space = SpaceDict({
**self._sim.sensor_suite.observation_spaces.spaces,
**self._task.sensor_suite.observation_spaces.spaces,
})
self.action_space = self._sim.action_space
self._max_episode_seconds = (
self._config.ENVIRONMENT.MAX_EPISODE_SECONDS)
self._max_episode_steps = self._config.ENVIRONMENT.MAX_EPISODE_STEPS
self._elapsed_steps = 0
self._episode_start_time: Optional[float] = None
self._episode_over = False
def get_scenes_to_load(config: Config) -> List[str]:
r"""Return list of scene ids for which dataset has separate files with
episodes.
"""
assert AudioNavDataset.check_config_paths_exist(config), \
(config.DATA_PATH.format(version=config.VERSION, split=config.SPLIT), config.SCENES_DIR)
dataset_dir = os.path.dirname(
config.DATA_PATH.format(version=config.VERSION, split=config.SPLIT)
)
cfg = config.clone()
cfg.defrost()
cfg.CONTENT_SCENES = []
dataset = AudioNavDataset(cfg)
return AudioNavDataset._get_scenes_from_folder(
content_scenes_path=dataset.content_scenes_path,
dataset_dir=dataset_dir,
)
# Copyright (c) Facebook, Inc. and its affiliates. # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union from habitat.config import Config as CN # type: ignore DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # Config definition # ----------------------------------------------------------------------------- _C = CN() _C.SEED = 100 # ----------------------------------------------------------------------------- # ENVIRONMENT # ----------------------------------------------------------------------------- _C.ENVIRONMENT = CN() _C.ENVIRONMENT.MAX_EPISODE_STEPS = 1000 _C.ENVIRONMENT.MAX_EPISODE_SECONDS = 10000000 _C.ENVIRONMENT.ITERATOR_OPTIONS = CN() _C.ENVIRONMENT.ITERATOR_OPTIONS.CYCLE = True _C.ENVIRONMENT.ITERATOR_OPTIONS.SHUFFLE = True _C.ENVIRONMENT.ITERATOR_OPTIONS.GROUP_BY_SCENE = True _C.ENVIRONMENT.ITERATOR_OPTIONS.NUM_EPISODE_SAMPLE = -1 _C.ENVIRONMENT.ITERATOR_OPTIONS.MAX_SCENE_REPEAT_EPISODES = -1 _C.ENVIRONMENT.ITERATOR_OPTIONS.MAX_SCENE_REPEAT_STEPS = int(1e4) _C.ENVIRONMENT.ITERATOR_OPTIONS.STEP_REPETITION_RANGE = 0.2
def __init__(self,
config: Config,
dataset: Optional[Dataset] = None) -> None:
"""Constructor. Mimics the :ref:`Env` constructor.
Args:
:param config: config for the environment. Should contain id for
simulator and ``task_name`` for each task, which are passed into ``make_sim`` and
``make_task``.
:param dataset: reference to dataset used for first task instance level
information. Can be defined as :py:`None` in which case
dataset will be built using ``make_dataset`` and ``config``.
"""
# let superclass instantiate current task by merging first task in TASKS to TASK
if len(config.MULTI_TASK.TASKS):
logger.info(
"Overwriting config.TASK ({}) with first entry in config.MULTI_TASK.TASKS ({})."
.format(config.TASK.TYPE, config.MULTI_TASK.TASKS[0].TYPE))
config.defrost()
config.TASK.merge_from_other_cfg(config.MULTI_TASK.TASKS[0])
config.freeze()
# TASKS[0] dataset has higher priority over default one (if specified), instatiate it before
if dataset is None and config.MULTI_TASK.TASKS[0].DATASET.TYPE:
dataset = make_dataset(
id_dataset=config.MULTI_TASK.TASKS[0].DATASET.TYPE,
config=config.MULTI_TASK.TASKS[0].DATASET,
)
# initialize first task leveraging Env
super().__init__(config, dataset=dataset)
# instatiate other tasks
self._tasks = [self._task]
self._curr_task_idx = 0
# keep each tasks episode iterator to avoid re-creation
self._episode_iterators = [self._episode_iterator]
for task in config.MULTI_TASK.TASKS[1:]:
self._tasks.append(
make_task(
task.TYPE,
config=task,
sim=self._sim,
# each task gets its dataset
dataset=make_dataset( # TODO: lazy make_dataset support
id_dataset=task.DATASET.TYPE,
config=task.DATASET),
))
# get task episode iterator
iter_option_dict = {
k.lower(): v
for k, v in task.EPISODE_ITERATOR_OPTIONS.items()
}
iter_option_dict["seed"] = self._config.SEED
task_ep_iterator: EpisodeIterator = self._tasks[
-1]._dataset.get_episode_iterator(**iter_option_dict)
self._episode_iterators.append(task_ep_iterator)
# episode counter
self._eps_counter = -1
self._cumulative_steps_counter = 0
# when and how to change task is defined here
self._task_sampling_behavior = (
config.MULTI_TASK.TASK_ITERATOR.TASK_SAMPLING)
self._change_task_behavior = (
config.MULTI_TASK.TASK_ITERATOR.TASK_CHANGE_TIMESTEP)
# custom task label can be specified
self._curr_task_label = self._task._config.get("TASK_LABEL",
self._curr_task_idx)
# add task_idx to observation space
self.observation_space = spaces.Dict({
**self._sim.sensor_suite.observation_spaces.spaces,
**self._task.sensor_suite.observation_spaces.spaces,
"task_idx":
spaces.Discrete(len(self._tasks)),
})
# Copyright (c) Facebook, Inc. and its affiliates. # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union from habitat.config import Config as CN # type: ignore DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # Config definition # ----------------------------------------------------------------------------- _C = CN() _C.SEED = 100 # ----------------------------------------------------------------------------- # ENVIRONMENT # ----------------------------------------------------------------------------- _C.ENVIRONMENT = CN() _C.ENVIRONMENT.MAX_EPISODE_STEPS = 1000 _C.ENVIRONMENT.MAX_EPISODE_SECONDS = 10000000 # ----------------------------------------------------------------------------- # TASK # ----------------------------------------------------------------------------- _C.TASK = CN() _C.TASK.TYPE = "Nav-v0" _C.TASK.SUCCESS_DISTANCE = 0.2 _C.TASK.SENSORS = [] _C.TASK.MEASUREMENTS = []
import os import logging import shutil import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN import habitat DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.SEED = 0 _C.BASE_TASK_CONFIG_PATH = "configs/tasks/pointgoal.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "AVNavTrainer" _C.ENV_NAME = "AudioNavRLEnv" _C.SIMULATOR_GPU_ID = 0 _C.TORCH_GPU_ID = 0 _C.VIDEO_OPTION = ["disk", "tensorboard"] _C.VISUALIZATION_OPTION = ["top_down_map"] _C.TENSORBOARD_DIR = "tb" _C.VIDEO_DIR = "video_dir" _C.TEST_EPISODE_COUNT = 2 _C.EVAL_CKPT_PATH_DIR = "data/checkpoints" # path to ckpt or path to ckpts dir _C.NUM_PROCESSES = 16
# This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.BASE_TASK_CONFIG_PATH = "configs/tasks/vln_r2r.yaml" _C.BERT_CONFIG = "/home/erick/Research/vln/libs/habitat/habitat-api/habitat_baselines/vln/config/bert_base_6layer_6connect.json" _C.BERT_GPU = 2 _C.BERT_PRE_TRAINED_MODEL = "bert-base-uncased" _C.MAX_TOKEN_LENGTH = 128 _C.TOKENIZER = "BertTokenizer" _C.DETECTRON2_GPU = 1 _C.DETECTRON2_MODEL = 'BUA_R_101_C4_MAX' #"IS_R_101_C4_3x" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "vln" _C.ENV_NAME = "VLNRLEnv" _C.SIMULATOR_GPU_ID = 3 _C.TORCH_GPU_ID = 4 _C.VIDEO_OPTION = ["disk", "tensorboard"]
# Copyright (c) Facebook, Inc. and its affiliates. # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. import os from typing import Optional from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" # ----------------------------------------------------------------------------- # Config definition # ----------------------------------------------------------------------------- _C = CN() _C.SEED = 100 # ----------------------------------------------------------------------------- # BASELINES # ----------------------------------------------------------------------------- _C.BASELINE = CN() # ----------------------------------------------------------------------------- # REINFORCEMENT LEARNING (RL) # ----------------------------------------------------------------------------- _C.BASELINE.RL = CN() _C.BASELINE.RL.SUCCESS_REWARD = 10.0 _C.BASELINE.RL.SLACK_REWARD = -0.01 # ----------------------------------------------------------------------------- def cfg(config_file: Optional[str] = None,
# This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.MULTIPLY_SCENES = False _C.BASE_TASK_CONFIG_PATH = "configs/tasks/explore_replica.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.SHARED_DATA = [] # HACKY share stuff _C.SHARED_SIZES = [] # HACKY share stuff _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "ppo" _C.ENV_NAME = "NavRLEnv" _C.SIMULATOR_GPU_ID = 0 _C.TORCH_GPU_ID = 0 _C.VIDEO_OPTION = ["disk", "tensorboard"] _C.VIDEO_OPTION_INTERVAL = 10 _C.TENSORBOARD_DIR = "tb" _C.VIDEO_DIR = "video_dir" _C.TEST_EPISODE_COUNT = 36
def get_default_r2r_v1_config(split: str = "val"):
config = Config()
config.name = "VLNR2R-v1"
config.DATA_PATH = "data/datasets/vln/r2r/v1/{split}/R2R_{split}.json.gz"
config.SPLIT = split
return config
# LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import math import numpy as np from habitat_extensions import get_extended_config as get_task_config from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.PYT_RANDOM_SEED = 123 _C.BASE_TASK_CONFIG_PATH = "habitat_extensions/config/exploration_gibson.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "occant_exp" _C.ENV_NAME = "ExpRLEnv" _C.SIMULATOR_GPU_ID = 0 _C.SIMULATOR_GPU_IDS = [] # Assign specific GPUs to simulator _C.TORCH_GPU_ID = 0 _C.VIDEO_OPTION = ["disk", "tensorboard"] _C.TENSORBOARD_DIR = "tb" _C.VIDEO_DIR = "video_dir" _C.TEST_EPISODE_COUNT = -1 _C.EVAL_CKPT_PATH_DIR = "data/checkpoints" # path to ckpt or path to ckpts dir _C.EVAL_PREV_CKPT_ID = -1 # The evaluation starts at (this value + 1)th ckpt
# LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN import os DEFAULT_CONFIG_DIR = "IL_configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.VERSION = 'base' _C.AGENT_TASK = 'search' _C.BASE_TASK_CONFIG_PATH = "IL_configs/tasks/pointnav.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "ppo" _C.ENV_NAME = "NavRLEnv" _C.SIMULATOR_GPU_ID = 0 _C.TORCH_GPU_ID = 0 _C.VIDEO_OPTION = ["disk", "tensorboard"] _C.TENSORBOARD_DIR = "logs/" _C.VIDEO_DIR = "data/video_dir" _C.TEST_EPISODE_COUNT = 2 _C.EVAL_CKPT_PATH_DIR = "data/eval_checkpoints" # path to ckpt or path to ckpts dir _C.NUM_PROCESSES = 16
# This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # Config definition # ----------------------------------------------------------------------------- _C = CN() _C.BASE_TASK_CONFIG_PATH = "configs/tasks/pointnav.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = "" # store command line options" # ----------------------------------------------------------------------------- # TRAINER ALGORITHMS # ----------------------------------------------------------------------------- _C.TRAINER = CN() _C.TRAINER.TRAINER_NAME = "ppo" # ----------------------------------------------------------------------------- # REINFORCEMENT LEARNING (RL) # ----------------------------------------------------------------------------- _C.TRAINER.RL = CN() _C.TRAINER.RL.SUCCESS_REWARD = 10.0 _C.TRAINER.RL.SLACK_REWARD = -0.01 # -----------------------------------------------------------------------------
# This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union import numpy as np from habitat import get_config as get_task_config from habitat.config import Config as CN DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # EXPERIMENT CONFIG # ----------------------------------------------------------------------------- _C = CN() _C.BASE_TASK_CONFIG_PATH = "configs/tasks/pointnav.yaml" _C.TASK_CONFIG = CN() # task_config will be stored as a config node _C.CMD_TRAILING_OPTS = [] # store command line options as list of strings _C.TRAINER_NAME = "ppo" _C.ENV_NAME = "NavRLEnv" _C.SIMULATOR_GPU_ID = 0 _C.TORCH_GPU_ID = 0 _C.VIDEO_OPTION = ["disk", "tensorboard"] _C.TENSORBOARD_DIR = "tb" _C.VIDEO_DIR = "video_dir" _C.TEST_EPISODE_COUNT = 2 _C.EVAL_CKPT_PATH_DIR = "data/checkpoints" # path to ckpt or path to ckpts dir _C.NUM_PROCESSES = 16 _C.SENSORS = ["RGB_SENSOR", "DEPTH_SENSOR"] _C.CHECKPOINT_FOLDER = "data/checkpoints"
# Copyright (c) Facebook, Inc. and its affiliates. # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from typing import List, Optional, Union from habitat.config import Config as CN # type: ignore DEFAULT_CONFIG_DIR = "configs/" CONFIG_FILE_SEPARATOR = "," # ----------------------------------------------------------------------------- # Config definition # ----------------------------------------------------------------------------- _C = CN() _C.SEED = 100 # ----------------------------------------------------------------------------- # ENVIRONMENT # ----------------------------------------------------------------------------- _C.ENVIRONMENT = CN() _C.ENVIRONMENT.MAX_EPISODE_STEPS = 1000 _C.ENVIRONMENT.MAX_EPISODE_SECONDS = 10000000 _C.ENVIRONMENT.ITERATOR_OPTIONS = CN() _C.ENVIRONMENT.ITERATOR_OPTIONS.CYCLE = True _C.ENVIRONMENT.ITERATOR_OPTIONS.SHUFFLE = True _C.ENVIRONMENT.ITERATOR_OPTIONS.GROUP_BY_SCENE = False _C.ENVIRONMENT.ITERATOR_OPTIONS.NUM_EPISODE_SAMPLE = -1 _C.ENVIRONMENT.ITERATOR_OPTIONS.MAX_SCENE_REPEAT = -1 _C.ENVIRONMENT.OVERRIDE_RAND_GOAL = CN()
def get_default_mp3d_v1_config(split: str = "val"):
config = Config()
config.name = "MP3DEQA-v1"
config.DATA_PATH = "data/datasets/eqa/mp3d/v1/{split}.json.gz"
config.SPLIT = split
return config
def __init__(self, config: Config) -> None:
"""Constructor
:param config: config for the environment. Should contain id for
simulator and ``task_name`` which are passed into ``make_sim`` and
``make_task``.
:param dataset: reference to dataset for task instance level
information. Can be defined as :py:`None` in which case
``_episodes`` should be populated from outside.
"""
assert config.is_frozen(), ("Freeze the config before creating the "
"environment, use config.freeze().")
self._config = config
self._current_episode_index = None
self._current_episode = None
iter_option_dict = {
k.lower(): v
for k, v in config.ENVIRONMENT.ITERATOR_OPTIONS.items()
}
self._scenes = config.DATASET.CONTENT_SCENES
self._swap_building_every = config.ENVIRONMENT.ITERATOR_OPTIONS.MAX_SCENE_REPEAT_EPISODES
print('[HabitatEnv] Total {} scenes : '.format(len(self._scenes)),
self._scenes)
print('[HabitatEnv] swap building every', self._swap_building_every)
self._current_scene_episode_idx = 0
self._current_scene_idx = 0
self._config.defrost()
if 'mp3d' in config.DATASET.DATA_PATH:
self._config.SIMULATOR.SCENE = os.path.join(
config.DATASET.SCENES_DIR,
'mp3d/{}/{}.glb'.format(self._scenes[0], self._scenes[0]))
else:
self._config.SIMULATOR.SCENE = os.path.join(
config.DATASET.SCENES_DIR,
'gibson_habitat/{}.glb'.format(self._scenes[0]))
if not os.path.exists(self._config.SIMULATOR.SCENE):
self._config.SIMULATOR.SCENE = os.path.join(
self._config.DATASET.SCENES_DIR,
'gibson_more/{}.glb'.format(self._scenes[0]))
self._config.freeze()
self._sim = make_sim(id_sim=self._config.SIMULATOR.TYPE,
config=self._config.SIMULATOR)
self._task = make_task(self._config.TASK.TYPE,
config=self._config.TASK,
sim=self._sim)
self.observation_space = SpaceDict({
**self._sim.sensor_suite.observation_spaces.spaces,
**self._task.sensor_suite.observation_spaces.spaces,
})
self.action_space = self._task.action_space
self._max_episode_seconds = (
self._config.ENVIRONMENT.MAX_EPISODE_SECONDS)
self._max_episode_steps = self._config.ENVIRONMENT.MAX_EPISODE_STEPS
self._elapsed_steps = 0
self._episode_start_time: Optional[float] = None
self._episode_over = False
#TODO listup demonstration data
self._episode_dataset = {}
