Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

object nav refactor #15

Open
wants to merge 10 commits into
base: procthor_stretch_objnav
Choose a base branch
from
Open

object nav refactor #15

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
4df063f
debugging procthor objnav
ehsanik May 23, 2022
40060a1
objnav config and sampler refactor 1 - did not work
rosehendrix May 24, 2022
e73c25c
refactor 2
rosehendrix May 24, 2022
67612ff
refactor 3 - complete 2camera
rosehendrix May 24, 2022
ca86f82
refactor 4 - cleanup and condense
rosehendrix May 24, 2022
1430f21
refactor 5 - update and test other experiment configs (narrow, roboth…
rosehendrix May 24, 2022
aab3d72
incorporate kiana changs to viz
rosehendrix May 24, 2022
26b5eb3
fix cherry-pick
rosehendrix May 25, 2022
7cc20d2
complete refactor 6 - n-camera model and preprocessors to mixin, prop…
rosehendrix May 25, 2022
13b40f8
update training pipeline for distributed nodes
rosehendrix May 26, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
133 changes: 50 additions & 83 deletions manipulathor_baselines/procthor_baselines/experiments/ithor/obj_nav_2camera_ithor_wide.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,27 +16,25 @@

from utils.stretch_utils.stretch_ithor_arm_environment import StretchManipulaTHOREnvironment

from manipulathor_baselines.procthor_baselines.experiments.ithor.obj_nav_for_procthor import ProcTHORObjectNavBaseConfig
from manipulathor_baselines.procthor_baselines.experiments.ithor.obj_nav_base_config import ObjectNavBaseConfig
from utils.procthor_utils.all_rooms_object_nav_task_sampler import AllRoomsObjectNavTaskSampler
from utils.procthor_utils.procthor_object_nav_tasks import StretchObjectNavTask
from utils.stretch_utils.stretch_constants import PROCTHOR_COMMIT_ID, STRETCH_ENV_ARGS
from utils.stretch_utils.stretch_constants import STRETCH_MANIPULATHOR_COMMIT_ID, STRETCH_ENV_ARGS
from manipulathor_utils.debugger_util import ForkedPdb

from manipulathor_baselines.procthor_baselines.models.clip_preprocessors import ClipResNetPreprocessor
from manipulathor_baselines.procthor_baselines.models.clip_objnav_ncamera_model import ResnetTensorNavNCameraActorCritic
# from manipulathor_baselines.procthor_baselines.models.clip_preprocessors import ClipResNetPreprocessor
from manipulathor_baselines.procthor_baselines.models.clip_resnet_ncamera_preprocess_mixin import \
ClipResNetPreprocessNCameraGRUActorCriticMixin
from allenact_plugins.clip_plugin.clip_preprocessors import ClipResNetPreprocessor
from allenact.base_abstractions.preprocessor import Preprocessor
from allenact.utils.experiment_utils import Builder
from utils.stretch_utils.stretch_visualizer import StretchObjNavImageVisualizer
from ithor_arm.ithor_arm_viz import TestMetricLogger

from scripts.dataset_generation.find_categories_to_use import FULL_LIST_OF_OBJECTS, ROBOTHOR_TRAIN, ROBOTHOR_VAL
from ithor_arm.ithor_arm_constants import TRAIN_OBJECTS, TEST_OBJECTS
from allenact_plugins.navigation_plugin.objectnav.models import ResnetTensorNavActorCritic



class ithorObjectNavClipResnet50RGBOnly2CameraWideFOV(
ProcTHORObjectNavBaseConfig
ObjectNavBaseConfig
):
"""An Object Navigation experiment configuration in iThor with RGB
input."""
Expand All @@ -54,35 +52,31 @@ class ithorObjectNavClipResnet50RGBOnly2CameraWideFOV(
for i in range(26, 31)
]


ALL_SCENES = TRAIN_SCENES + TEST_SCENES + VALID_SCENES

# OBJECT_TYPES = tuple(sorted(TRAIN_OBJECTS))
with open('datasets/objects/robothor_habitat2022.yaml', 'r') as f:
OBJECT_TYPES=yaml.safe_load(f)

# UNSEEN_OBJECT_TYPES = tuple(sorted(TEST_OBJECTS))


NOISE_LEVEL = 0
distance_thr = 1.0 # match procthor config
mean = np.array([0.48145466, 0.4578275, 0.40821073])
stdev = np.array([0.26862954, 0.26130258, 0.27577711])
SENSORS = [
RGBSensorThor(
height=224,
width=224,
height=ObjectNavBaseConfig.SCREEN_SIZE,
width=ObjectNavBaseConfig.SCREEN_SIZE,
use_resnet_normalization=True,
mean=mean,
stdev=stdev,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres",
),
RGBSensorStretchKinectBigFov(
height=224,
width=224,
height=ObjectNavBaseConfig.SCREEN_SIZE,
width=ObjectNavBaseConfig.SCREEN_SIZE,
use_resnet_normalization=True,
mean=mean,
stdev=stdev,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_arm",
),
GoalObjectTypeThorSensor(
Expand All @@ -93,86 +87,59 @@ class ithorObjectNavClipResnet50RGBOnly2CameraWideFOV(
MAX_STEPS = 500
if platform.system() == "Darwin":
MAX_STEPS = 100
SENSORS += [
RGBSensorStretchKinectBigFov(
height=ObjectNavBaseConfig.SCREEN_SIZE,
width=ObjectNavBaseConfig.SCREEN_SIZE,
use_resnet_normalization=True,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_arm_only_viz",
),
RGBSensorThor(
height=ObjectNavBaseConfig.SCREEN_SIZE,
width=ObjectNavBaseConfig.SCREEN_SIZE,
use_resnet_normalization=True,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_only_viz",
),
]

TASK_SAMPLER = AllRoomsObjectNavTaskSampler
TASK_TYPE = StretchObjectNavTask
ENVIRONMENT_TYPE = StretchManipulaTHOREnvironment
POTENTIAL_VISUALIZERS = [StretchObjNavImageVisualizer, TestMetricLogger]


NUM_PROCESSES = 40
CLIP_MODEL_TYPE = "RN50"


def __init__(self):
super().__init__()
self.REWARD_CONFIG['goal_success_reward'] = 10.0
self.REWARD_CONFIG['step_penalty'] = -0.01
self.REWARD_CONFIG['failed_stop_reward'] = 0.0
self.REWARD_CONFIG['reached_horizon_reward'] = 0.0
self.REWARD_CONFIG['shaping_weight'] = 1.0

self.ENV_ARGS = copy.deepcopy(STRETCH_ENV_ARGS)
self.ENV_ARGS['visibilityDistance'] = self.distance_thr
self.ENV_ARGS['environment_type'] = self.ENVIRONMENT_TYPE #TODO this is nto the best choice
# self.ENV_ARGS['scene'] = 'Procedural'
self.ENV_ARGS['renderInstanceSegmentation'] = 'False'
self.ENV_ARGS['commit_id'] = PROCTHOR_COMMIT_ID
self.ENV_ARGS['allow_flipping'] = True
self.ENV_ARGS['renderInstanceSegmentation'] = False
self.ENV_ARGS['commit_id'] = STRETCH_MANIPULATHOR_COMMIT_ID

self.preprocessing_and_model = ClipResNetPreprocessNCameraGRUActorCriticMixin(
sensors=self.SENSORS,
clip_model_type=self.CLIP_MODEL_TYPE,
screen_size=self.SCREEN_SIZE,
)

def preprocessors(self) -> Sequence[Union[Preprocessor, Builder[Preprocessor]]]:
return self.preprocessing_and_model.preprocessors()

@classmethod
@final
def preprocessors(cls) -> Sequence[Union[Preprocessor, Builder[Preprocessor]]]:
preprocessors = []
# rgb_sensor = next((s for s in cls.SENSORS if isinstance(s, RGBSensorThor)), None)

preprocessors.append(
ClipResNetPreprocessor(
rgb_input_uuid="rgb_lowres",
clip_model_type="RN50",
pool=False,
output_uuid="rgb_clip_resnet",
visualize=cls.VISUALIZE
)
)
preprocessors.append(
ClipResNetPreprocessor(
rgb_input_uuid="rgb_lowres_arm",
clip_model_type="RN50",
pool=False,
output_uuid="rgb_clip_resnet_arm",
visualize=cls.VISUALIZE
)

def create_model(self, **kwargs) -> nn.Module:
return self.preprocessing_and_model.create_model(
num_actions=len(self.TASK_TYPE.class_action_names()), **kwargs,
visualize=self.VISUALIZE
)
return preprocessors

@classmethod
@final
def create_model(cls, **kwargs) -> nn.Module:
goal_sensor_uuid = next(
(s.uuid for s in cls.SENSORS if isinstance(s, GoalObjectTypeThorSensor)),
None,
)
resnet_preprocessor_uuids = ["rgb_clip_resnet","rgb_clip_resnet_arm"]

return ResnetTensorNavNCameraActorCritic(
action_space=gym.spaces.Discrete(len(cls.TASK_TYPE.class_action_names())),
observation_space=kwargs["sensor_preprocessor_graph"].observation_spaces,
goal_sensor_uuid=goal_sensor_uuid,
resnet_preprocessor_uuids=resnet_preprocessor_uuids,
hidden_size=512,
goal_dims=32,
add_prev_actions=True,
)
# return ResnetTensorNavActorCritic(
# action_space=gym.spaces.Discrete(len(cls.TASK_TYPE.class_action_names())),
# observation_space=kwargs["sensor_preprocessor_graph"].observation_spaces,
# goal_sensor_uuid=goal_sensor_uuid,
# rgb_resnet_preprocessor_uuid="rgb_clip_resnet",
# depth_resnet_preprocessor_uuid="rgb_clip_resnet_arm", # a convenient lie - can't use with a depth sensor too
# hidden_size=512,
# goal_dims=32,
# add_prev_actions=True,
# )

@classmethod
def tag(cls):
Expand Down
166 changes: 29 additions & 137 deletions ...ulathor_baselines/procthor_baselines/experiments/ithor/obj_nav_2camera_procthor_narrow.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,168 +1,60 @@
import platform
import random
from typing import Sequence, Union
from typing_extensions import final

import gym
import numpy as np
from torch import nn
import yaml
import copy

from allenact_plugins.ithor_plugin.ithor_sensors import RGBSensorThor
from utils.stretch_utils.stretch_thor_sensors import RGBSensorStretchKinect, RGBSensorStretchIntel
from allenact_plugins.ithor_plugin.ithor_sensors import GoalObjectTypeThorSensor


from utils.stretch_utils.stretch_ithor_arm_environment import StretchManipulaTHOREnvironment

from manipulathor_baselines.procthor_baselines.experiments.ithor.obj_nav_for_procthor import ProcTHORObjectNavBaseConfig
from utils.procthor_utils.procthor_object_nav_task_samplers import ProcTHORObjectNavTaskSampler
from utils.procthor_utils.procthor_object_nav_tasks import StretchObjectNavTask
from utils.stretch_utils.stretch_constants import PROCTHOR_COMMIT_ID, STRETCH_ENV_ARGS
from manipulathor_utils.debugger_util import ForkedPdb

from manipulathor_baselines.procthor_baselines.models.clip_preprocessors import ClipResNetPreprocessor
from manipulathor_baselines.procthor_baselines.models.clip_objnav_ncamera_model import ResnetTensorNavNCameraActorCritic
from allenact.base_abstractions.preprocessor import Preprocessor
from allenact.utils.experiment_utils import Builder
from utils.stretch_utils.stretch_visualizer import StretchObjNavImageVisualizer
from ithor_arm.ithor_arm_viz import TestMetricLogger
from allenact_plugins.navigation_plugin.objectnav.models import ResnetTensorNavActorCritic

from manipulathor_baselines.procthor_baselines.experiments.ithor.obj_nav_for_procthor_clip_resnet50_rgb_only import ProcTHORObjectNavClipResnet50RGBOnly
from allenact_plugins.clip_plugin.clip_preprocessors import ClipResNetPreprocessor


class ProcTHORObjectNavClipResnet50RGBOnly2CameraNarrowFOV(
ProcTHORObjectNavBaseConfig
ProcTHORObjectNavClipResnet50RGBOnly
):
"""An Object Navigation experiment configuration in iThor with RGB
input."""

with open('datasets/objects/robothor_habitat2022.yaml', 'r') as f:
OBJECT_TYPES=yaml.safe_load(f)

NOISE_LEVEL = 0
distance_thr = 1.0 # match procthor config
mean = np.array([0.48145466, 0.4578275, 0.40821073])
stdev = np.array([0.26862954, 0.26130258, 0.27577711])
SENSORS = [
RGBSensorStretchIntel(
height=224,
width=224,
height=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
width=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
use_resnet_normalization=True,
mean=mean,
stdev=stdev,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres",
),
RGBSensorStretchKinect(
height=224,
width=224,
height=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
width=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
use_resnet_normalization=True,
mean=mean,
stdev=stdev,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_arm",
),
GoalObjectTypeThorSensor(
object_types=OBJECT_TYPES,
),
]

MAX_STEPS = 500
if platform.system() == "Darwin":
MAX_STEPS = 100

TASK_SAMPLER = ProcTHORObjectNavTaskSampler
TASK_TYPE = StretchObjectNavTask
ENVIRONMENT_TYPE = StretchManipulaTHOREnvironment
POTENTIAL_VISUALIZERS = [StretchObjNavImageVisualizer, TestMetricLogger]


NUM_PROCESSES = 56

TRAIN_SCENES = [f'ProcTHOR{i}' for i in range(9999)] # 9999 is all of train
if platform.system() == "Darwin":
TRAIN_SCENES = [f'ProcTHOR{i}' for i in range(100)]

TEST_SCENES = [f'ProcTHOR{i}' for i in range(100)]
random.shuffle(TRAIN_SCENES)


def __init__(self):
super().__init__()
self.REWARD_CONFIG['goal_success_reward'] = 10.0
self.REWARD_CONFIG['step_penalty'] = -0.01
self.REWARD_CONFIG['failed_stop_reward'] = 0.0
self.REWARD_CONFIG['reached_horizon_reward'] = 0.0
self.REWARD_CONFIG['shaping_weight'] = 1.0

self.ENV_ARGS = copy.deepcopy(STRETCH_ENV_ARGS)
self.ENV_ARGS['visibilityDistance'] = self.distance_thr
self.ENV_ARGS['environment_type'] = self.ENVIRONMENT_TYPE #TODO this is nto the best choice
self.ENV_ARGS['scene'] = 'Procedural'
self.ENV_ARGS['renderInstanceSegmentation'] = 'False'
self.ENV_ARGS['commit_id'] = PROCTHOR_COMMIT_ID
self.ENV_ARGS['allow_flipping'] = True


@classmethod
@final
def preprocessors(cls) -> Sequence[Union[Preprocessor, Builder[Preprocessor]]]:
preprocessors = []
# rgb_sensor = next((s for s in cls.SENSORS if isinstance(s, RGBSensorThor)), None)

preprocessors.append(
ClipResNetPreprocessor(
rgb_input_uuid="rgb_lowres",
clip_model_type="RN50",
pool=False,
output_uuid="rgb_clip_resnet",
visualize=cls.VISUALIZE
)
)
preprocessors.append(
ClipResNetPreprocessor(
rgb_input_uuid="rgb_lowres_arm",
clip_model_type="RN50",
pool=False,
output_uuid="rgb_clip_resnet_arm",
visualize=cls.VISUALIZE
)
)
return preprocessors


@classmethod
@final
def create_model(cls, **kwargs) -> nn.Module:
goal_sensor_uuid = next(
(s.uuid for s in cls.SENSORS if isinstance(s, GoalObjectTypeThorSensor)),
None,
)
resnet_preprocessor_uuids = ["rgb_clip_resnet","rgb_clip_resnet_arm"]

return ResnetTensorNavNCameraActorCritic(
action_space=gym.spaces.Discrete(len(cls.TASK_TYPE.class_action_names())),
observation_space=kwargs["sensor_preprocessor_graph"].observation_spaces,
goal_sensor_uuid=goal_sensor_uuid,
resnet_preprocessor_uuids=resnet_preprocessor_uuids,
hidden_size=512,
goal_dims=32,
add_prev_actions=True,
)
# return ResnetTensorNavActorCritic(
# action_space=gym.spaces.Discrete(len(cls.TASK_TYPE.class_action_names())),
# observation_space=kwargs["sensor_preprocessor_graph"].observation_spaces,
# goal_sensor_uuid=goal_sensor_uuid,
# rgb_resnet_preprocessor_uuid="rgb_clip_resnet",
# depth_resnet_preprocessor_uuid="rgb_clip_resnet_arm", # a convenient lie - can't use with a depth sensor too
# hidden_size=512,
# goal_dims=32,
# add_prev_actions=True,
# )


@classmethod
def tag(cls):
return cls.__name__


SENSORS += [
RGBSensorStretchKinect(
height=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
width=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
use_resnet_normalization=True,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_arm_only_viz",
),
RGBSensorStretchIntel(
height=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
width=ProcTHORObjectNavClipResnet50RGBOnly.SCREEN_SIZE,
use_resnet_normalization=True,
mean=ClipResNetPreprocessor.CLIP_RGB_MEANS,
stdev=ClipResNetPreprocessor.CLIP_RGB_STDS,
uuid="rgb_lowres_only_viz",
),
]
Loading