def construct_envs(config: Config, env_class: Type[Union[Env,
RLEnv]]) -> VectorEnv:
r"""Create VectorEnv object with specified config and env class type.
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
Args:
config: configs that contain num_processes as well as information
necessary to create individual environments.
env_class: class type of the envs to be created.
Returns:
VectorEnv object created according to specification.
"""
num_processes = config.NUM_PROCESSES
configs = []
env_classes = [env_class for _ in range(num_processes)]
dataset = make_dataset(config.TASK_CONFIG.DATASET.TYPE)
scenes = dataset.get_scenes_to_load(config.TASK_CONFIG.DATASET)
if len(scenes) > 0:
random.shuffle(scenes)
assert len(scenes) >= num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_splits = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
task_config = config.TASK_CONFIG.clone()
task_config.defrost()
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
config.SIMULATOR_GPU_ID)
task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
task_config.freeze()
config.defrost()
config.TASK_CONFIG = task_config
config.freeze()
configs.append(config.clone())
envs = habitat.VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(zip(configs, env_classes, range(num_processes)))),
)
return envs
def construct_env_configs_mp3d(config: Config) -> List[Config]:
r"""Create list of Habitat Configs for training on multiple processes
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
Args:
config: configs that contain num_processes as well as information
necessary to create individual environments.
Returns:
List of Configs, one for each process
"""
config.freeze()
num_processes = config.NUM_PROCESSES
configs = []
# dataset = habitat.make_dataset(config.DATASET.TYPE)
# scenes = dataset.get_scenes_to_load(config.DATASET)
if num_processes == 1:
scene_splits = [["pRbA3pwrgk9"]]
else:
small = [
"rPc6DW4iMge",
"e9zR4mvMWw7",
"uNb9QFRL6hY",
"qoiz87JEwZ2",
"sKLMLpTHeUy",
"s8pcmisQ38h",
"759xd9YjKW5",
"XcA2TqTSSAj",
"SN83YJsR3w2",
"8WUmhLawc2A",
"JeFG25nYj2p",
"17DRP5sb8fy",
"Uxmj2M2itWa",
"XcA2TqTSSAj",
"SN83YJsR3w2",
"8WUmhLawc2A",
"JeFG25nYj2p",
"17DRP5sb8fy",
"Uxmj2M2itWa",
"D7N2EKCX4Sj",
"b8cTxDM8gDG",
"sT4fr6TAbpF",
"S9hNv5qa7GM",
"82sE5b5pLXE",
"pRbA3pwrgk9",
"aayBHfsNo7d",
"cV4RVeZvu5T",
"i5noydFURQK",
"YmJkqBEsHnH",
"jh4fc5c5qoQ",
"VVfe2KiqLaN",
"29hnd4uzFmX",
"Pm6F8kyY3z2",
"JF19kD82Mey",
"GdvgFV5R1Z5",
"HxpKQynjfin",
"vyrNrziPKCB",
]
med = [
"V2XKFyX4ASd",
"VFuaQ6m2Qom",
"ZMojNkEp431",
"5LpN3gDmAk7",
"r47D5H71a5s",
"ULsKaCPVFJR",
"E9uDoFAP3SH",
"kEZ7cmS4wCh",
"ac26ZMwG7aT",
"dhjEzFoUFzH",
"mJXqzFtmKg4",
"p5wJjkQkbXX",
"Vvot9Ly1tCj",
"EDJbREhghzL",
"VzqfbhrpDEA",
"7y3sRwLe3Va",
]
scene_splits = [[] for _ in range(config.NUM_PROCESSES)]
distribute(
small,
scene_splits,
num_gpus=8,
procs_per_gpu=3,
proc_offset=1,
scenes_per_process=2,
)
distribute(
med,
scene_splits,
num_gpus=8,
procs_per_gpu=3,
proc_offset=0,
scenes_per_process=1,
)
# gpu0 = [['pRbA3pwrgk9', '82sE5b5pLXE', 'S9hNv5qa7GM'],
# ['Uxmj2M2itWa', '17DRP5sb8fy', 'JeFG25nYj2p'],
# ['5q7pvUzZiYa', '759xd9YjKW5', 's8pcmisQ38h'],
# ['e9zR4mvMWw7', 'rPc6DW4iMge', 'vyrNrziPKCB']]
# gpu1 = [['sT4fr6TAbpF', 'b8cTxDM8gDG', 'D7N2EKCX4Sj'],
# ['8WUmhLawc2A', 'SN83YJsR3w2', 'XcA2TqTSSAj'],
# ['sKLMLpTHeUy', 'qoiz87JEwZ2', 'uNb9QFRL6hY'],
# ['V2XKFyX4ASd', 'VFuaQ6m2Qom', 'ZMojNkEp431']]
# gpu2 = [['5LpN3gDmAk7', 'r47D5H71a5s', 'ULsKaCPVFJR', 'E9uDoFAP3SH'],
# ['VVfe2KiqLaN', 'jh4fc5c5qoQ', 'YmJkqBEsHnH'], # small
# ['i5noydFURQK', 'cV4RVeZvu5T', 'aayBHfsNo7d']] # small
# gpu3 = [['kEZ7cmS4wCh', 'ac26ZMwG7aT', 'dhjEzFoUFzH'],
# ['mJXqzFtmKg4', 'p5wJjkQkbXX', 'Vvot9Ly1tCj']]
# gpu4 = [['EDJbREhghzL', 'VzqfbhrpDEA', '7y3sRwLe3Va'],
# ['ur6pFq6Qu1A', 'PX4nDJXEHrG', 'PuKPg4mmafe']]
# gpu5 = [['r1Q1Z4BcV1o', 'gTV8FGcVJC9', '1pXnuDYAj8r'],
# ['JF19kD82Mey', 'Pm6F8kyY3z2', '29hnd4uzFmX']] # small
# gpu6 = [['VLzqgDo317F', '1LXtFkjw3qL'],
# ['HxpKQynjfin', 'gZ6f7yhEvPG', 'GdvgFV5R1Z5']] # small
# gpu7 = [['D7G3Y4RVNrH', 'B6ByNegPMKs']]
#
# scene_splits = gpu0 + gpu1 + gpu2 + gpu3 + gpu4 + gpu5 + gpu6 + gpu7
for i in range(num_processes):
task_config = config.clone()
task_config.defrost()
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = config.SIMULATOR_GPU_IDS[
i % len(config.SIMULATOR_GPU_IDS)
]
task_config.freeze()
configs.append(task_config.clone())
return configs
def construct_env_configs(
config: Config, allow_scene_repeat: bool = False,
) -> List[Config]:
"""Create list of Habitat Configs for training on multiple processes To
allow better performance, dataset are split into small ones for each
individual env, grouped by scenes.
# Parameters
config : configs that contain num_processes as well as information
necessary to create individual environments.
allow_scene_repeat: if `True` and the number of distinct scenes
in the dataset is less than the total number of processes this will
result in scenes being repeated across processes. If `False`, then
if the total number of processes is greater than the number of scenes,
this will result in a RuntimeError exception being raised.
# Returns
List of Configs, one for each process.
"""
config.freeze()
num_processes = config.NUM_PROCESSES
configs = []
dataset = habitat.make_dataset(config.DATASET.TYPE)
scenes = dataset.get_scenes_to_load(config.DATASET)
if len(scenes) > 0:
if len(scenes) < num_processes:
if not allow_scene_repeat:
raise RuntimeError(
"reduce the number of processes as there aren't enough number of scenes."
)
else:
scenes = (scenes * (1 + (num_processes // len(scenes))))[:num_processes]
scene_splits: List[List] = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
task_config = config.clone()
task_config.defrost()
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
if len(config.SIMULATOR_GPU_IDS) == 0:
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = -1
else:
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = config.SIMULATOR_GPU_IDS[
i % len(config.SIMULATOR_GPU_IDS)
]
task_config.freeze()
configs.append(task_config.clone())
return configs
def construct_envs(
config: Config,
env_class: Union[Type[Env], Type[RLEnv]],
workers_ignore_signals: bool = False,
) -> VectorEnv:
r"""Create VectorEnv object with specified config and env class type.
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
:param config: configs that contain num_environments as well as information
:param necessary to create individual environments.
:param env_class: class type of the envs to be created.
:param workers_ignore_signals: Passed to :ref:`habitat.VectorEnv`'s constructor
:return: VectorEnv object created according to specification.
"""
num_environments = config.NUM_ENVIRONMENTS
configs = []
env_classes = [env_class for _ in range(num_environments)]
dataset = make_dataset(config.TASK_CONFIG.DATASET.TYPE)
scenes = config.TASK_CONFIG.DATASET.CONTENT_SCENES
if "*" in config.TASK_CONFIG.DATASET.CONTENT_SCENES:
scenes = dataset.get_scenes_to_load(config.TASK_CONFIG.DATASET)
if num_environments > 1:
if len(scenes) == 0:
raise RuntimeError(
"No scenes to load, multiple process logic relies on being able to split scenes uniquely between processes"
)
if len(scenes) < num_environments:
raise RuntimeError(
"reduce the number of environments as there "
"aren't enough number of scenes.\n"
"num_environments: {}\tnum_scenes: {}".format(
num_environments, len(scenes)
)
)
random.shuffle(scenes)
scene_splits: List[List[str]] = [[] for _ in range(num_environments)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_environments):
proc_config = config.clone()
proc_config.defrost()
task_config = proc_config.TASK_CONFIG
task_config.SEED = task_config.SEED + i
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
config.SIMULATOR_GPU_ID
)
task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
proc_config.freeze()
configs.append(proc_config)
envs = habitat.VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(zip(configs, env_classes)),
workers_ignore_signals=workers_ignore_signals,
)
return envs
def construct_envs(config: Config, env_class: Type[Union[Env,
RLEnv]]) -> VectorEnv:
r"""Create VectorEnv object with specified config and env class type.
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
Args:
config: configs that contain num_processes as well as information
necessary to create individual environments.
env_class: class type of the envs to be created
Returns:
VectorEnv object created according to specification.
"""
num_processes = config.NUM_PROCESSES
configs = []
env_classes = [env_class for _ in range(num_processes)]
dataset = make_dataset(config.TASK_CONFIG.DATASET.TYPE)
scenes = dataset.get_scenes_to_load(config.TASK_CONFIG.DATASET)
# rearrange scenes in the order of scene size since there is a severe imbalance of data size
if "replica" in config.TASK_CONFIG.DATASET.SCENES_DIR:
scenes_new = list()
for scene in SCENES:
if scene in scenes:
scenes_new.append(scene)
scenes = scenes_new
if len(scenes) > 0:
# random.shuffle(scenes)
assert len(scenes) >= num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_splits = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
task_config = config.TASK_CONFIG.clone()
task_config.defrost()
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
logging.debug('All scenes: {}'.format(','.join(scene_splits[i])))
# overwrite the task config with top-level config file
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
config.SIMULATOR_GPU_ID)
task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
task_config.freeze()
config.defrost()
config.TASK_CONFIG = task_config
config.freeze()
configs.append(config.clone())
# use VectorEnv for the best performance and ThreadedVectorEnv for debugging
if config.USE_SYNC_VECENV:
env_launcher = SyncVectorEnv
logging.info('Using SyncVectorEnv')
elif config.USE_VECENV:
env_launcher = habitat.VectorEnv
logging.info('Using VectorEnv')
else:
env_launcher = habitat.ThreadedVectorEnv
logging.info('Using ThreadedVectorEnv')
envs = env_launcher(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(zip(configs, env_classes, range(num_processes)))),
)
return envs
def construct_envs(config: Config, env_class: Type[Union[Env,
RLEnv]]) -> VectorEnv:
r"""Create VectorEnv object with specified config and env class type.
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
Args:
config: configs that contain num_processes as well as information
necessary to create individual environments.
env_class: class type of the envs to be created.
Returns:
VectorEnv object created according to specification.
"""
num_processes = config.NUM_PROCESSES
configs = []
env_classes = [env_class for _ in range(num_processes)]
dataset = make_dataset(config.TASK_CONFIG.DATASET.TYPE)
scenes = config.TASK_CONFIG.DATASET.CONTENT_SCENES
if "*" in config.TASK_CONFIG.DATASET.CONTENT_SCENES:
scenes = dataset.get_scenes_to_load(config.TASK_CONFIG.DATASET)
print("************************************* scenes lens:", len(scenes))
print("************************************* num_processes lens:",
num_processes)
print("************************************* env_classes:", env_class)
if num_processes > 1:
if len(scenes) == 0:
raise RuntimeError(
"No scenes to load, multiple process logic relies on being able to split scenes uniquely between processes"
)
if len(scenes) < num_processes:
raise RuntimeError("reduce the number of processes as there "
"aren't enough number of scenes")
random.shuffle(scenes)
scene_splits = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
proc_config = config.clone()
proc_config.defrost()
task_config = proc_config.TASK_CONFIG
task_config.SEED = task_config.SEED + i
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
config.SIMULATOR_GPU_ID)
task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
proc_config.freeze()
configs.append(proc_config)
envs = habitat.VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(tuple(zip(configs, env_classes))),
)
print("************************************* envs type:", type(envs))
print("************************************* count_episodes:",
(envs.count_episodes()))
# print("************************************* current_episodes:", (envs.current_episodes()))
return envs
def construct_env(config: Config) -> Env:
r"""Create VectorEnv object with specified config and env class type.
To allow better performance, dataset are split into small ones for
each individual env, grouped by scenes.
Args:
config: configs that contain num_processes as well as information
necessary to create individual environments.
env_class: class type of the envs to be created.
auto_reset_done: Whether or not to automatically reset the env on done
Returns:
VectorEnv object created according to specification.
"""
num_processes = config.NUM_PROCESSES
configs = []
dataset = make_dataset(config.TASK_CONFIG.DATASET.TYPE)
scenes = dataset.get_scenes_to_load(config.TASK_CONFIG.DATASET)
if num_processes > 1:
if len(scenes) == 0:
raise RuntimeError(
"No scenes to load, multiple process logic relies on being able to split scenes uniquely between processes"
)
if len(scenes) < num_processes:
raise RuntimeError("reduce the number of processes as there "
"aren't enough number of scenes")
random.shuffle(scenes)
scene_splits = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
new_config = config.clone()
task_config = new_config.TASK_CONFIG.clone()
task_config.defrost()
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
config.SIMULATOR_GPU_ID[i % len(config.SIMULATOR_GPU_ID)])
logger.info(f"Simulator GPU ID {config.SIMULATOR_GPU_ID}")
logger.info(
f"Simulator GPU ID {task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID}"
)
task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
task_config.freeze()
new_config.defrost()
new_config.TASK_CONFIG = task_config
new_config.freeze()
configs.append(new_config)
# for i in range(num_processes):
# proc_config = config.clone()
# proc_config.defrost()
# task_config = proc_config.TASK_CONFIG
# if len(scenes) > 0:
# task_config.DATASET.CONTENT_SCENES = scene_splits[i]
# task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = (
# config.SIMULATOR_GPU_ID[i % len(config.SIMULATOR_GPU_ID)]
# )
# # task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = config.SIMULATOR_GPU_ID
# task_config.SIMULATOR.AGENT_0.SENSORS = config.SENSORS
# proc_config.freeze()
# configs.append(proc_config)
for config in configs:
logger.info(
f"[construct_envs] Using GPU ID {config.TASK_CONFIG.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID}"
)
env = VLNCEDaggerEnv(config)
return env
def construct_env_configs(config: Config) -> List[Config]:
"""Create list of Habitat Configs for training on multiple processes To
allow better performance, dataset are split into small ones for each
individual env, grouped by scenes.
# Parameters
config : configs that contain num_processes as well as information
necessary to create individual environments.
# Returns
List of Configs, one for each process.
"""
config.freeze()
num_processes = config.NUM_PROCESSES
configs = []
dataset = habitat.make_dataset(config.DATASET.TYPE)
scenes = dataset.get_scenes_to_load(config.DATASET)
# scenes = [
# "sT4fr6TAbpF",
# "HxpKQynjfin",
# "8WUmhLawc2A",
# "r47D5H71a5s",
# "Pm6F8kyY3z2",
# "17DRP5sb8fy",
# "Vvot9Ly1tCj",
# "GdvgFV5R1Z5",
# "sT4fr6TAbpF",
# "HxpKQynjfin",
# "8WUmhLawc2A",
# "r47D5H71a5s",
# "Pm6F8kyY3z2",
# "17DRP5sb8fy",
# "Vvot9Ly1tCj",
# "GdvgFV5R1Z5",
# "sT4fr6TAbpF",
# "HxpKQynjfin",
# "8WUmhLawc2A",
# "r47D5H71a5s",
# "Pm6F8kyY3z2",
# "17DRP5sb8fy",
# "Vvot9Ly1tCj",
# "GdvgFV5R1Z5",
# ]
# scenes = ['rPc6DW4iMge', 'e9zR4mvMWw7', 'uNb9QFRL6hY', 'sKLMLpTHeUy', 's8pcmisQ38h', '759xd9YjKW5',
# 'XcA2TqTSSAj', 'SN83YJsR3w2', '8WUmhLawc2A', 'JeFG25nYj2p', '17DRP5sb8fy', 'Uxmj2M2itWa',
# 'b8cTxDM8gDG', 'sT4fr6TAbpF', 'S9hNv5qa7GM', '82sE5b5pLXE', 'pRbA3pwrgk9', 'aayBHfsNo7d',
# 'cV4RVeZvu5T', 'i5noydFURQK', 'jh4fc5c5qoQ', 'VVfe2KiqLaN', '29hnd4uzFmX', 'Pm6F8kyY3z2',
# 'JF19kD82Mey', 'GdvgFV5R1Z5', 'HxpKQynjfin']
# scenes = ['rPc6DW4iMge', 'e9zR4mvMWw7', 'uNb9QFRL6hY', 'qoiz87JEwZ2', 'sKLMLpTHeUy', 's8pcmisQ38h', '759xd9YjKW5',
# '5q7pvUzZiYa', 'XcA2TqTSSAj', 'SN83YJsR3w2', '8WUmhLawc2A', 'JeFG25nYj2p', '17DRP5sb8fy', 'Uxmj2M2itWa',
# 'D7N2EKCX4Sj', 'b8cTxDM8gDG', 'sT4fr6TAbpF', 'S9hNv5qa7GM', '82sE5b5pLXE', 'pRbA3pwrgk9', 'aayBHfsNo7d',
# 'cV4RVeZvu5T', 'i5noydFURQK', 'YmJkqBEsHnH', 'jh4fc5c5qoQ', 'VVfe2KiqLaN', '29hnd4uzFmX', 'Pm6F8kyY3z2',
# 'JF19kD82Mey', 'GdvgFV5R1Z5', 'HxpKQynjfin', 'vyrNrziPKCB']
# scenes = ['29hnd4uzFmX', 'i5noydFURQK', 'cV4RVeZvu5T', '82sE5b5pLXE', 'JeFG25nYj2p', '8WUmhLawc2A', 'VFuaQ6m2Qom',
# 'rPc6DW4iMge', '29hnd4uzFmX', 'i5noydFURQK', 'cV4RVeZvu5T', '82sE5b5pLXE',
# 'JeFG25nYj2p', '8WUmhLawc2A', 'VFuaQ6m2Qom', 'rPc6DW4iMge']
if len(scenes) > 0:
# random.shuffle(scenes)
assert len(scenes) >= num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_splits: List[List] = [[] for _ in range(num_processes)]
for idx, scene in enumerate(scenes):
scene_splits[idx % len(scene_splits)].append(scene)
assert sum(map(len, scene_splits)) == len(scenes)
for i in range(num_processes):
task_config = config.clone()
task_config.defrost()
if len(scenes) > 0:
task_config.DATASET.CONTENT_SCENES = scene_splits[i]
task_config.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = config.SIMULATOR_GPU_IDS[
i % len(config.SIMULATOR_GPU_IDS)]
task_config.freeze()
configs.append(task_config.clone())
return configs