def test_dataset_splitting(split):
dataset_config = get_config(CFG_MULTI_TEST).DATASET
dataset_config.defrost()
dataset_config.SPLIT = split
if not PointNavDatasetV1.check_config_paths_exist(dataset_config):
pytest.skip("Test skipped as dataset files are missing.")
scenes = PointNavDatasetV1.get_scenes_to_load(config=dataset_config)
assert (len(scenes) >
0), "Expected dataset contains separate episode file per scene."
dataset_config.CONTENT_SCENES = scenes[:PARTIAL_LOAD_SCENES]
full_dataset = make_dataset(id_dataset=dataset_config.TYPE,
config=dataset_config)
full_episodes = {(ep.scene_id, ep.episode_id)
for ep in full_dataset.episodes}
dataset_config.CONTENT_SCENES = scenes[:PARTIAL_LOAD_SCENES // 2]
split1_dataset = make_dataset(id_dataset=dataset_config.TYPE,
config=dataset_config)
split1_episodes = {(ep.scene_id, ep.episode_id)
for ep in split1_dataset.episodes}
dataset_config.CONTENT_SCENES = scenes[PARTIAL_LOAD_SCENES //
2:PARTIAL_LOAD_SCENES]
split2_dataset = make_dataset(id_dataset=dataset_config.TYPE,
config=dataset_config)
split2_episodes = {(ep.scene_id, ep.episode_id)
for ep in split2_dataset.episodes}
assert full_episodes == split1_episodes.union(
split2_episodes), "Split dataset is not equal to full dataset"
assert (len(split1_episodes.intersection(split2_episodes)) == 0
), "Intersection of split datasets is not the empty set"
def gibson_env(hparams):
basic_config = cfg_env(config_file=hparams.task_config,
config_dir=os.path.join(dirname, 'configs'))
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
config_env = cfg_env(config_file=hparams.task_config,
config_dir=os.path.join(dirname, 'configs'))
config_env.defrost()
if len(scenes) > 0:
random.shuffle(scenes)
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = scenes
for sensor in hparams.sensors:
assert sensor in ["RGB_SENSOR", "DEPTH_SENSOR"]
config_env.SIMULATOR.AGENT_0.SENSORS = hparams.sensors
config_env.freeze()
config_baseline = cfg_baseline()
dataset = PointNavDatasetV1(config_env.DATASET)
config_env.defrost()
config_env.SIMULATOR.SCENE = dataset.episodes[0].scene_id
config_env.freeze()
env = NavRLEnv(config_env=config_env,
config_baseline=config_baseline,
dataset=dataset)
return env
def test_multiple_files_scene_path():
dataset_config = get_config(CFG_MULTI_TEST).DATASET
if not PointNavDatasetV1.check_config_paths_exist(dataset_config):
pytest.skip("Test skipped as dataset files are missing.")
scenes = PointNavDatasetV1.get_scenes_to_load(config=dataset_config)
assert (
len(scenes) > 0
), "Expected dataset contains separate episode file per scene."
dataset_config.defrost()
dataset_config.CONTENT_SCENES = scenes[:PARTIAL_LOAD_SCENES]
dataset_config.SCENES_DIR = os.path.join(
os.getcwd(), DEFAULT_SCENE_PATH_PREFIX
)
dataset_config.freeze()
partial_dataset = make_dataset(
id_dataset=dataset_config.TYPE, config=dataset_config
)
assert (
len(partial_dataset.scene_ids) == PARTIAL_LOAD_SCENES
), "Number of loaded scenes doesn't correspond."
print(partial_dataset.episodes[0].scene_id)
assert os.path.exists(
partial_dataset.episodes[0].scene_id
), "Scene file {} doesn't exist using absolute path".format(
partial_dataset.episodes[0].scene_id
)
def test_single_pointnav_dataset():
dataset_config = get_config().DATASET
if not PointNavDatasetV1.check_config_paths_exist(dataset_config):
pytest.skip("Test skipped as dataset files are missing.")
scenes = PointNavDatasetV1.get_scenes_to_load(config=dataset_config)
assert (len(scenes) >
0), "Expected dataset contains separate episode file per scene."
dataset = PointNavDatasetV1(config=dataset_config)
assert len(dataset.episodes) > 0, "The dataset shouldn't be empty."
assert (len(
dataset.scene_ids) == 2), "The test dataset scenes number is wrong."
check_json_serializaiton(dataset)
def test_multiple_files_pointnav_dataset(cfg_path: str):
dataset_config = get_config(cfg_path).DATASET
if not PointNavDatasetV1.check_config_paths_exist(dataset_config):
pytest.skip("Test skipped as dataset files are missing.")
scenes = PointNavDatasetV1.get_scenes_to_load(config=dataset_config)
assert (len(scenes) >
0), "Expected dataset contains separate episode file per scene."
dataset_config.defrost()
dataset_config.CONTENT_SCENES = scenes
dataset_config.freeze()
partial_dataset = make_dataset(id_dataset=dataset_config.TYPE,
config=dataset_config)
check_json_serializaiton(partial_dataset)
def test_multiple_files_pointnav_dataset():
dataset_config = get_config(CFG_MULTI_TEST).DATASET
if not PointNavDatasetV1.check_config_paths_exist(dataset_config):
pytest.skip("Test skipped as dataset files are missing.")
scenes = PointNavDatasetV1.get_scenes_to_load(config=dataset_config)
assert (len(scenes) >
0), "Expected dataset contains separate episode file per scene."
dataset_config.defrost()
dataset_config.CONTENT_SCENES = scenes[:PARTIAL_LOAD_SCENES]
dataset_config.freeze()
partial_dataset = make_dataset(id_dataset=dataset_config.TYPE,
config=dataset_config)
assert (len(partial_dataset.scene_ids) == PARTIAL_LOAD_SCENES
), "Number of loaded scenes doesn't correspond."
check_json_serializaiton(partial_dataset)
def construct_envs(args):
env_configs = []
baseline_configs = []
args_list = []
# TODO check params consistency here
basic_config = cfg_env(config_paths=[args.task_config])
print("loading scenes ...")
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
if len(scenes) > 0:
assert len(scenes) >= args.num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_split_size = int(np.floor(len(scenes) / args.num_processes))
print("using ", args.num_processes, " processes and ", scene_split_size,
" scenes per process")
for i in range(args.num_processes):
config_env = cfg_env(config_paths=[args.task_config])
config_env.defrost()
if len(scenes) > 0:
config_env.DATASET.CONTENT_SCENES = scenes[i *
scene_split_size:(i +
1) *
scene_split_size]
if i < args.num_processes_on_first_gpu:
gpu_id = 0
else:
gpu_id = int((i - args.num_processes_on_first_gpu) //
args.num_processes_per_gpu) + args.sim_gpu_id
# gpu_id = min(torch.cuda.device_count() - 1, gpu_id)
gpu_id = 0
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = gpu_id
config_env.freeze()
env_configs.append(config_env)
config_baseline = cfg_baseline()
baseline_configs.append(config_baseline)
args_list.append(args)
envs = VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(
zip(args_list, env_configs, baseline_configs,
range(args.num_processes)))),
)
# envs = make_env_fn(args_list[0], env_configs[0], config_baseline=baseline_configs[0], rank=42)
print("returning with environment")
return envs
def make_env_fn(config_env, config_baseline, rank):
dataset = PointNavDatasetV1(config_env.DATASET)
config_env.defrost()
config_env.SIMULATOR.SCENE = dataset.episodes[0].scene_id
config_env.freeze()
env = NavRLEnv(config_env=config_env,
config_baseline=config_baseline,
dataset=dataset)
env.seed(rank)
return env
def test_demo_notebook():
config = habitat.get_config("configs/tasks/pointnav_mp3d.yaml")
config.defrost()
config.DATASET.SPLIT = "val"
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip(
"Please download the Matterport3D PointNav val dataset and Matterport3D val scenes"
)
else:
pytest.main(["--nbval-lax", "notebooks/habitat-api-demo.ipynb"])
def test_demo_notebook():
config = habitat.get_config("configs/tasks/pointnav_rgbd.yaml")
config.defrost()
config.DATASET.SPLIT = "val"
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip("Please download the habitat test scenes")
else:
pytest.main([
"--nbval-lax",
"notebooks/relative_camera_views_transform_and_warping_demo.ipynb",
])
def make_env_fn(args, config_env, rank):
dataset = PointNavDatasetV1(config_env.DATASET)
config_env.defrost()
config_env.SIMULATOR.SCENE = dataset.episodes[0].scene_id
print("Loading {}".format(config_env.SIMULATOR.SCENE))
config_env.freeze()
env = Neural_SLAM_Env(args=args, rank=rank,
config_env=config_env, dataset=dataset
)
env.seed(rank)
return env
def make_env_fn(args, config_env, config_baseline, rank):
print("-------------- condig_env ---------------")
print(config_env)
print("-----------------------------------------")
dataset = PointNavDatasetV1(config_env.DATASET)
print("Loading {}".format(config_env.SIMULATOR.SCENE))
env = PointNavEnv(args=args,
rank=rank,
config_env=config_env,
config_baseline=config_baseline,
dataset=dataset)
env.seed(rank)
return env
def make_env_fn(config_env, config_baseline, rank, target_dim, log_dir,
visdom_name, visdom_log_file, vis_interval, visdom_server,
visdom_port, swap_building_k_episodes, map_kwargs,
reward_kwargs, should_record, seed):
if config_env.DATASET.SPLIT == 'val':
datasetfile_path = config_env.DATASET.POINTNAVV1.DATA_PATH.format(
split=config_env.DATASET.SPLIT)
dataset = PointNavDatasetV1()
with gzip.open(datasetfile_path, "rt") as f:
dataset.from_json(f.read())
else:
dataset = PointNavDatasetV1(config_env.DATASET)
config_env.defrost()
config_env.SIMULATOR.SCENE = dataset.episodes[0].scene_id
config_env.freeze()
env = MidlevelNavRLEnv(config_env=config_env,
config_baseline=config_baseline,
dataset=dataset,
target_dim=target_dim,
map_kwargs=map_kwargs,
reward_kwargs=reward_kwargs)
env.episodes = shuffle_episodes(env, swap_every_k=swap_building_k_episodes)
env.seed(seed)
if should_record and visdom_log_file is not None:
print("SETTING VISDOM MONITOR WITH VIS INTERVAL", vis_interval)
env = VisdomMonitor(env,
directory=os.path.join(log_dir, visdom_name),
video_callable=lambda x: x % vis_interval == 0,
uid=str(rank),
server=visdom_server,
port=visdom_port,
visdom_log_file=visdom_log_file,
visdom_env=visdom_name)
return env
def test_demo_notebook():
config = habitat.get_config("configs/tasks/pointnav_mp3d.yaml")
config.defrost()
config.DATASET.SPLIT = "val"
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip(
"Please download the Matterport3D PointNav val dataset and Matterport3D val scenes"
)
else:
pytest.main(["--nbval-lax", "notebooks/habitat-api-demo.ipynb"])
# NB: Force a gc collect run as it can take a little bit for
# the cleanup to happen after the notebook and we get
# a double context crash!
gc.collect()
def test_demo_notebook():
config = habitat.get_config("configs/tasks/pointnav_rgbd.yaml")
config.defrost()
config.DATASET.SPLIT = "val"
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip("Please download the habitat test scenes")
else:
pytest.main([
"--nbval-lax",
"notebooks/relative_camera_views_transform_and_warping_demo.ipynb",
])
# NB: Force a gc collect run as it can take a little bit for
# the cleanup to happen after the notebook and we get
# a double context crash!
gc.collect()
def construct_envs(args):
env_configs = []
baseline_configs = []
basic_config = cfg_env(config_paths=args.task_config, opts=args.opts)
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
if len(scenes) > 0:
random.shuffle(scenes)
assert len(scenes) >= args.num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_split_size = int(np.floor(len(scenes) / args.num_processes))
for i in range(args.num_processes):
config_env = cfg_env(config_paths=args.task_config, opts=args.opts)
config_env.defrost()
if len(scenes) > 0:
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = scenes[
i * scene_split_size:(i + 1) * scene_split_size]
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = args.sim_gpu_id
agent_sensors = args.sensors.strip().split(",")
for sensor in agent_sensors:
assert sensor in ["RGB_SENSOR", "DEPTH_SENSOR"]
config_env.SIMULATOR.AGENT_0.SENSORS = agent_sensors
config_env.freeze()
env_configs.append(config_env)
config_baseline = cfg_baseline()
baseline_configs.append(config_baseline)
logger.info("config_env: {}".format(config_env))
envs = habitat.VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(zip(env_configs, baseline_configs,
range(args.num_processes)))),
)
return envs
def main():
basic_config = cfg_env(config_paths=
["env/habitat/habitat_lab/configs/" + args.task_config])
basic_config.defrost()
basic_config.DATASET.SPLIT = args.split
basic_config.freeze()
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
config_env = cfg_env(config_paths=
["env/habitat/habitat_lab/configs/" + args.task_config])
config_env.defrost()
config_env.DATASET.CONTENT_SCENES = scenes
gpu_id = 0
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = gpu_id
agent_sensors = []
agent_sensors.append("RGB_SENSOR")
agent_sensors.append("DEPTH_SENSOR")
config_env.SIMULATOR.AGENT_0.SENSORS = agent_sensors
config_env.ENVIRONMENT.MAX_EPISODE_STEPS = args.max_episode_length
config_env.ENVIRONMENT.ITERATOR_OPTIONS.SHUFFLE = False
config_env.SIMULATOR.RGB_SENSOR.WIDTH = args.env_frame_width
config_env.SIMULATOR.RGB_SENSOR.HEIGHT = args.env_frame_height
config_env.SIMULATOR.RGB_SENSOR.HFOV = args.hfov
config_env.SIMULATOR.RGB_SENSOR.POSITION = [0, args.camera_height, 0]
config_env.SIMULATOR.DEPTH_SENSOR.WIDTH = args.env_frame_width
config_env.SIMULATOR.DEPTH_SENSOR.HEIGHT = args.env_frame_height
config_env.SIMULATOR.DEPTH_SENSOR.HFOV = args.hfov
config_env.SIMULATOR.DEPTH_SENSOR.POSITION = [0, args.camera_height, 0]
config_env.SIMULATOR.TURN_ANGLE = 10
config_env.DATASET.SPLIT = args.split
config_env.freeze()
env = make_env_fn(args, config_env, 0)
obs, inf = env.reset()
print("done")
def _load_test_data():
configs = []
datasets = []
for i in range(NUM_ENVS):
config = get_config(CFG_TEST)
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip("Please download Habitat test data to data folder.")
datasets.append(
habitat.make_dataset(id_dataset=config.DATASET.TYPE,
config=config.DATASET))
config.defrost()
config.SIMULATOR.SCENE = datasets[-1].episodes[0].scene_id
if not os.path.exists(config.SIMULATOR.SCENE):
pytest.skip("Please download Habitat test data to data folder.")
config.freeze()
configs.append(config)
return configs, datasets
def test_pointnav_episode_generator():
config = get_config(CFG_TEST)
config.defrost()
config.DATASET.SPLIT = "val"
config.ENVIRONMENT.MAX_EPISODE_STEPS = 500
config.freeze()
if not PointNavDatasetV1.check_config_paths_exist(config.DATASET):
pytest.skip("Test skipped as dataset files are missing.")
with habitat.Env(config) as env:
env.seed(config.SEED)
random.seed(config.SEED)
generator = pointnav_generator.generate_pointnav_episode(
sim=env.sim,
shortest_path_success_distance=config.TASK.SUCCESS_DISTANCE,
shortest_path_max_steps=config.ENVIRONMENT.MAX_EPISODE_STEPS,
)
episodes = []
for i in range(NUM_EPISODES):
episode = next(generator)
episodes.append(episode)
for episode in pointnav_generator.generate_pointnav_episode(
sim=env.sim,
num_episodes=NUM_EPISODES,
shortest_path_success_distance=config.TASK.SUCCESS_DISTANCE,
shortest_path_max_steps=config.ENVIRONMENT.MAX_EPISODE_STEPS,
geodesic_to_euclid_min_ratio=0,
):
episodes.append(episode)
assert len(episodes) == 2 * NUM_EPISODES
env.episode_iterator = iter(episodes)
for episode in episodes:
check_shortest_path(env, episode)
dataset = habitat.Dataset()
dataset.episodes = episodes
assert (
dataset.to_json()
), "Generated episodes aren't json serializable."
def make_habitat_vector_env(num_processes=2,
target_dim=7,
preprocessing_fn=None,
log_dir=None,
visdom_name='main',
visdom_log_file=None,
visdom_server='localhost',
visdom_port='8097',
vis_interval=200,
scenes=None,
val_scenes=['Greigsville', 'Pablo', 'Mosquito'],
num_val_processes=0,
swap_building_k_episodes=10,
gpu_devices=[0],
collate_obs_before_transform=False,
map_kwargs={},
reward_kwargs={},
seed=42):
assert map_kwargs[
'map_building_size'] > 0, 'Map building size must be positive!'
default_reward_kwargs = {
'slack_reward': -0.01,
'success_reward': 10,
'use_visit_penalty': False,
'visit_penalty_coef': 0,
'penalty_eps': 999,
}
for k, v in default_reward_kwargs.items():
if k not in reward_kwargs:
reward_kwargs[k] = v
habitat_path = os.path.dirname(os.path.dirname(habitat.__file__))
task_config = os.path.join(habitat_path,
'configs/tasks/pointnav_gibson_train.yaml')
basic_config = cfg_env(config_file=task_config)
basic_config.defrost()
basic_config.DATASET.POINTNAVV1.DATA_PATH = os.path.join(
habitat_path, basic_config.DATASET.POINTNAVV1.DATA_PATH)
basic_config.freeze()
if scenes is None:
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
random.shuffle(scenes)
val_task_config = os.path.join(
habitat_path, 'configs/tasks/pointnav_gibson_val_mini.yaml')
val_cfg = cfg_env(config_file=val_task_config)
val_cfg.defrost()
val_cfg.DATASET.SPLIT = "val"
val_cfg.freeze()
scenes = [s for s in scenes if s not in val_scenes]
if num_val_processes > 0 and len(val_scenes) % num_val_processes != 0:
warnings.warn(
"Please make num_val_processes ({}) evenly divide len(val_scenes) ({}) or some buildings may be overrepresented"
.format(num_val_processes, len(val_scenes)))
env_configs = []
baseline_configs = []
encoders = []
target_dims = []
is_val = []
# Assign specific buildings to each process
train_process_scenes = [[]
for _ in range(num_processes - num_val_processes)]
for i, scene in enumerate(scenes):
train_process_scenes[i % len(train_process_scenes)].append(scene)
if num_val_processes > 0:
val_process_scenes = [[] for _ in range(num_val_processes)]
for i, scene in enumerate(val_scenes):
val_process_scenes[i % len(val_process_scenes)].append(scene)
for i in range(num_processes):
config_env = cfg_env(task_config)
config_env.defrost()
config_env.DATASET.POINTNAVV1.DATA_PATH = os.path.join(
habitat_path, basic_config.DATASET.POINTNAVV1.DATA_PATH)
if i < num_processes - num_val_processes:
config_env.DATASET.SPLIT = 'train'
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = train_process_scenes[
i]
else:
config_env.DATASET.SPLIT = 'val'
val_i = i - (num_processes - num_val_processes)
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = val_process_scenes[
val_i]
print("Env {}:".format(i),
config_env.DATASET.POINTNAVV1.CONTENT_SCENES)
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = gpu_devices[
i % len(gpu_devices)]
agent_sensors = ["RGB_SENSOR"]
config_env.SIMULATOR.AGENT_0.SENSORS = agent_sensors
config_env.SIMULATOR.SCENE = os.path.join(habitat_path,
config_env.SIMULATOR.SCENE)
config_env.freeze()
env_configs.append(config_env)
config_baseline = cfg_baseline()
baseline_configs.append(config_baseline)
encoders.append(preprocessing_fn)
target_dims.append(target_dim)
should_record = [(i == 0 or i == (num_processes - num_val_processes))
for i in range(num_processes)]
envs = HabitatPreprocessVectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(
zip(
env_configs,
baseline_configs,
range(num_processes),
target_dims,
[log_dir for _ in range(num_processes)],
[visdom_name for _ in range(num_processes)],
[visdom_log_file for _ in range(num_processes)],
[vis_interval for _ in range(num_processes)],
[visdom_server for _ in range(num_processes)],
[visdom_port for _ in range(num_processes)],
[swap_building_k_episodes for _ in range(num_processes)],
[map_kwargs for _ in range(num_processes)],
[reward_kwargs for _ in range(num_processes)],
should_record,
[seed + i for i in range(num_processes)],
))),
preprocessing_fn=preprocessing_fn,
collate_obs_before_transform=collate_obs_before_transform)
envs.observation_space = envs.observation_spaces[0]
envs.action_space = spaces.Discrete(3)
envs.reward_range = None
envs.metadata = None
envs.is_embodied = True
return envs
def test_new_actions():
if not PointNavDatasetV1.check_config_paths_exist(
config=habitat.get_config().DATASET):
pytest.skip("Please download Habitat test data to data folder.")
new_actions.main()
def test_register_new_sensors_and_measures():
if not PointNavDatasetV1.check_config_paths_exist(
config=habitat.get_config().DATASET):
pytest.skip("Please download Habitat test data to data folder.")
register_new_sensors_and_measures.main()
def test_visualizations_example():
if not PointNavDatasetV1.check_config_paths_exist(
config=habitat.get_config().DATASET):
pytest.skip("Please download Habitat test data to data folder.")
visualization_examples.main()
def test_readme_example():
if not PointNavDatasetV1.check_config_paths_exist(
config=habitat.get_config().DATASET):
pytest.skip("Please download Habitat test data to data folder.")
example()
def construct_envs(args):
env_configs = []
# baseline_configs = []
args_list = []
basic_config = cfg_env(
config_paths=["env/habitat/habitat_lab/configs/" + args.task_config])
basic_config.defrost()
basic_config.DATASET.SPLIT = args.split
basic_config.freeze()
scenes = PointNavDatasetV1.get_scenes_to_load(basic_config.DATASET)
if len(scenes) > 0:
assert len(scenes) >= args.num_processes, (
"reduce the number of processes as there "
"aren't enough number of scenes")
scene_split_size = int(np.floor(len(scenes) / args.num_processes))
for i in range(args.num_processes):
config_env = cfg_env(config_paths=[
"env/habitat/habitat_lab/configs/" + args.task_config
])
config_env.defrost()
if len(scenes) > 0:
config_env.DATASET.CONTENT_SCENES = scenes[i *
scene_split_size:(i +
1) *
scene_split_size]
if i < args.num_processes_on_first_gpu:
gpu_id = 0
else:
gpu_id = int((i - args.num_processes_on_first_gpu) //
args.num_processes_per_gpu) + args.sim_gpu_id
gpu_id = min(torch.cuda.device_count() - 1, gpu_id)
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = gpu_id
agent_sensors = []
agent_sensors.append("RGB_SENSOR")
agent_sensors.append("DEPTH_SENSOR")
config_env.SIMULATOR.AGENT_0.SENSORS = agent_sensors
config_env.ENVIRONMENT.MAX_EPISODE_STEPS = args.max_episode_length
config_env.ENVIRONMENT.ITERATOR_OPTIONS.SHUFFLE = False
config_env.SIMULATOR.RGB_SENSOR.WIDTH = args.env_frame_width
config_env.SIMULATOR.RGB_SENSOR.HEIGHT = args.env_frame_height
config_env.SIMULATOR.RGB_SENSOR.HFOV = args.hfov
config_env.SIMULATOR.RGB_SENSOR.POSITION = [0, args.camera_height, 0]
config_env.SIMULATOR.DEPTH_SENSOR.WIDTH = args.env_frame_width
config_env.SIMULATOR.DEPTH_SENSOR.HEIGHT = args.env_frame_height
config_env.SIMULATOR.DEPTH_SENSOR.HFOV = args.hfov
config_env.SIMULATOR.DEPTH_SENSOR.POSITION = [0, args.camera_height, 0]
config_env.SIMULATOR.TURN_ANGLE = 10
config_env.DATASET.SPLIT = args.split
config_env.freeze()
env_configs.append(config_env)
# Baseline configs are not used
# config_baseline = cfg_baseline()
# baseline_configs.append(config_baseline)
args_list.append(args)
envs = VectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(zip(args_list, env_configs, range(args.num_processes)))),
)
return envs
def make_habitat_vector_env(
scenario='PointNav',
num_processes=2,
target_dim=7,
preprocessing_fn=None,
log_dir=None,
visdom_name='main',
visdom_log_file=None,
visdom_server='localhost',
visdom_port='8097',
vis_interval=200,
train_scenes=None,
val_scenes=None,
num_val_processes=0,
swap_building_k_episodes=10,
gpu_devices=[0],
map_kwargs={},
reward_kwargs={},
seed=42,
test_mode=False,
debug_mode=False,
scenario_kwargs={},
):
assert map_kwargs[
'map_building_size'] > 0, 'Map building size must be positive!'
default_reward_kwargs = {
'slack_reward': -0.01,
'success_reward': 10,
'use_visit_penalty': False,
'visit_penalty_coef': 0,
'penalty_eps': 999,
'sparse': False,
'dist_coef': 1.0,
}
for k, v in default_reward_kwargs.items():
if k not in reward_kwargs:
reward_kwargs[k] = v
habitat_path = os.path.dirname(os.path.dirname(habitat.__file__))
if scenario == 'PointNav' or scenario == 'Exploration':
task_config = os.path.join(habitat_path,
'configs/tasks/pointnav_gibson_train.yaml')
# only difference is that Exploration needs DEPTH_SENSOR but that is added in the Env
# task_config = os.path.join(habitat_path, 'configs/tasks/exploration_gibson.yaml')
env_configs = []
baseline_configs = []
encoders = []
target_dims = []
is_val = []
# Assign specific episodes to each process
config_env = cfg_env(task_config)
# Load dataset
print('Loading val dataset (partition by episode)...')
datasetfile_path = config_env.DATASET.POINTNAVV1.DATA_PATH.format(
split='val')
dataset = PointNavDatasetV1()
with gzip.open(datasetfile_path, "rt") as f:
dataset.from_json(f.read())
val_datasets = get_splits(dataset, max(num_val_processes, 1))
# for d in val_datasets:
# d.episodes = [d.episodes[0]]
print('Loaded.')
print('Loading train dataset (partition by building)...')
train_datasets = []
if num_processes - num_val_processes > 0:
# dataset = PointNavDatasetV1(config_env.DATASET)
train_datasets = [
None for _ in range(num_processes - num_val_processes)
]
print('Loaded.')
# Assign specific buildings to each process
if num_processes > num_val_processes:
train_process_scenes = [[] for _ in range(num_processes -
num_val_processes)]
if train_scenes is None:
train_scenes = PointNavDatasetV1.get_scenes_to_load(
config_env.DATASET)
random.shuffle(train_scenes)
for i, scene in enumerate(train_scenes):
train_process_scenes[i % len(train_process_scenes)].append(scene)
# If n processes > n envs, some processes can use all envs
for j, process in enumerate(train_process_scenes):
if len(process) == 0:
train_process_scenes[j] = list(train_scenes)
get_scenes = lambda d: list(
Counter([e.scene_id.split('/')[-1].split(".")[0]
for e in d.episodes]).items())
for i in range(num_processes):
config_env = cfg_env(task_config)
config_env.defrost()
if i < num_processes - num_val_processes:
config_env.DATASET.SPLIT = 'train'
# config_env.DATASET.POINTNAVV1.CONTENT_SCENES = get_scenes(train_datasets[i])
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = train_process_scenes[
i]
else:
val_i = i - (num_processes - num_val_processes)
config_env.DATASET.SPLIT = 'val'
if val_scenes is not None:
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = val_scenes
else:
config_env.DATASET.POINTNAVV1.CONTENT_SCENES = get_scenes(
val_datasets[val_i])
print("Env {}:".format(i),
config_env.DATASET.POINTNAVV1.CONTENT_SCENES)
config_env.SIMULATOR.HABITAT_SIM_V0.GPU_DEVICE_ID = gpu_devices[
i % len(gpu_devices)]
config_env.SIMULATOR.SCENE = os.path.join(habitat_path,
config_env.SIMULATOR.SCENE)
config_env.SIMULATOR.AGENT_0.SENSORS = ["RGB_SENSOR"]
# Now define the config for the sensor
# config.TASK.AGENT_POSITION_SENSOR = habitat.Config()
# config.TASK.AGENT_POSITION_SENSOR.TYPE = "agent_position_sensor"
# config.TASK.SENSORS.append("AGENT_POSITION_SENSOR")
config_env.TASK.MEASUREMENTS.append('COLLISIONS')
config_env.freeze()
env_configs.append(config_env)
config_baseline = cfg_baseline()
baseline_configs.append(config_baseline)
encoders.append(preprocessing_fn)
target_dims.append(target_dim)
should_record = [(i == 0 or i == (num_processes - num_val_processes))
for i in range(num_processes)]
if debug_mode:
env = make_env_fn(
scenario,
env_configs[0],
baseline_configs[0],
0,
0,
1,
target_dim,
log_dir,
visdom_name,
visdom_log_file,
vis_interval,
visdom_server,
visdom_port,
swap_building_k_episodes,
map_kwargs,
reward_kwargs,
False,
seed, # TODO set should_record to True
test_mode,
(train_datasets + val_datasets)[0],
scenario_kwargs)
envs = PreprocessEnv(env, preprocessing_fn=preprocessing_fn)
else:
envs = HabitatPreprocessVectorEnv(
make_env_fn=make_env_fn,
env_fn_args=tuple(
tuple(
zip(
[scenario for _ in range(num_processes)],
env_configs,
baseline_configs,
range(num_processes),
[num_val_processes for _ in range(num_processes)],
[num_processes for _ in range(num_processes)],
target_dims,
[log_dir for _ in range(num_processes)],
[visdom_name for _ in range(num_processes)],
[visdom_log_file for _ in range(num_processes)],
[vis_interval for _ in range(num_processes)],
[visdom_server for _ in range(num_processes)],
[visdom_port for _ in range(num_processes)],
[
swap_building_k_episodes
for _ in range(num_processes)
],
[map_kwargs for _ in range(num_processes)],
[reward_kwargs for _ in range(num_processes)],
should_record,
[seed + i for i in range(num_processes)],
[test_mode for _ in range(num_processes)],
train_datasets + val_datasets,
[scenario_kwargs for _ in range(num_processes)],
))),
preprocessing_fn=preprocessing_fn,
)
envs.observation_space = envs.observation_spaces[0]
envs.action_space = spaces.Discrete(3)
envs.reward_range = None
envs.metadata = None
envs.is_embodied = True
return envs
def make_env_fn(scenario, config_env, config_baseline, rank, num_val_processes,
num_processes, target_dim, log_dir, visdom_name,
visdom_log_file, vis_interval, visdom_server, visdom_port,
swap_building_k_episodes, map_kwargs, reward_kwargs,
should_record, seed, test_mode, dataset, scenario_kwargs):
if config_env.DATASET.SPLIT == 'train':
dataset = PointNavDatasetV1(config_env.DATASET)
habitat_path = os.path.dirname(os.path.dirname(habitat.__file__))
for ep in dataset.episodes:
ep.scene_id = os.path.join(habitat_path, ep.scene_id)
config_env.defrost()
config_env.SIMULATOR.SCENE = dataset.episodes[0].scene_id
config_env.freeze()
if scenario == 'PointNav':
dataset.episodes = [
epi for epi in dataset.episodes if epi.info['geodesic_distance'] <
scenario_kwargs['max_geodesic_dist']
]
env = MidlevelNavRLEnv(config_env=config_env,
config_baseline=config_baseline,
dataset=dataset,
target_dim=target_dim,
map_kwargs=map_kwargs,
reward_kwargs=reward_kwargs,
loop_episodes=not test_mode,
scenario_kwargs=scenario_kwargs)
elif scenario == 'Exploration':
env = ExplorationRLEnv(config_env=config_env,
config_baseline=config_baseline,
dataset=dataset,
map_kwargs=map_kwargs,
reward_kwargs=reward_kwargs,
loop_episodes=not test_mode,
scenario_kwargs=scenario_kwargs)
else:
assert False, f'do not recognize scenario {scenario}'
if test_mode:
env.episodes = env.episodes
else:
env.episodes = shuffle_episodes(env,
swap_every_k=swap_building_k_episodes)
env.seed(seed)
if should_record and visdom_log_file is not None:
print(
f"Recording videos from env {rank} every {vis_interval} episodes (via visdom)"
)
env = VisdomMonitor(env,
directory=os.path.join(log_dir, visdom_name),
video_callable=lambda x: x % vis_interval == 0,
uid=str(rank),
server=visdom_server,
port=visdom_port,
visdom_log_file=visdom_log_file,
visdom_env=visdom_name)
return env
def test_shortest_path_follower_example():
if not PointNavDatasetV1.check_config_paths_exist(
config=habitat.get_config().DATASET):
pytest.skip("Please download Habitat test data to data folder.")
shortest_path_follower_example.main()
