Merge pull request #84 from allenai/advisor_into_master

Running black, optimizing imports, and inelegant merging of engines (currently kept as separate files).

.gitignore

@@ -135,4 +135,16 @@ dmypy.json
 *.pt
 
 # Default output dir
-experiment_output
+experiment_output
+
+# PDFs
+*.pdf
+
+# Tensorboard logs
+events.out.tfevents.*
+
+# TSV files
+*.tsv
+
+# tmp directory
+tmp/

Pipfile

@@ -37,6 +37,10 @@ numpy-quaternion = "*"
 pandas = "*"
 seaborn = "*"
 statsmodels = "*"
+gym-minigrid = "*"
+gin-config = "*"
+babyai = {editable = true,git = "https://github.com/unnat/babyai.git",ref = "ff645fe00ea8412a29bd5e2d6f79ae1595d229a7"}
+localreg = "*"
 
 [requires]
 python_version = "3.6"

ddmain.py

@@ -1,4 +1,5 @@
-"""Entry point to training/validating/testing for a user given experiment name"""
+"""Entry point to training/validating/testing for a user given experiment
+name."""
 
 import argparse
 import importlib
@@ -11,7 +12,7 @@
 
 from onpolicy_sync.runner import OnPolicyRunner
 from rl_base.experiment_config import ExperimentConfig
-from utils.system import init_logging, LOGGER
+from utils.system import get_logger
 
 
 def _get_args():
@@ -31,8 +32,8 @@ def _get_args():
         default="",
         required=False,
         help="Add an extra tag to the experiment when trying out new ideas (will be used"
-             "as a suffix of the experiment name). It also has to be provided when testing on"
-             "the trained model.",
+        "as a suffix of the experiment name). It also has to be provided when testing on"
+        "the trained model.",
     )
 
     parser.add_argument(
@@ -70,7 +71,7 @@ def _get_args():
         action="store_true",
         required=False,
         help="for training, if checkpoint is specified, use it as model initialization and "
-             "restart training with current config",
+        "restart training with current config",
     )
     parser.set_defaults(restart=False)
 
@@ -152,11 +153,9 @@ def _load_config(args) -> Tuple[ExperimentConfig, Dict[str, Tuple[str, str]]]:
 
 
 def main():
-    init_logging()
-
     args = _get_args()
 
-    LOGGER.info("Running with args {}".format(args))
+    get_logger().info("Running with args {}".format(args))
 
     ptitle("Master: {}".format("Training" if args.test_date is None else "Testing"))
 

docs/overview/defining-a-new-task.md

@@ -188,7 +188,7 @@ Finally, we need to define methods to determine the number of available tasks (p
 ```python
 
     @property
-    def __len__(self) -> Union[int, float]:
+    def length(self) -> Union[int, float]:
         return float("inf")
 
     @property

docs/overview/experiment.md

@@ -193,8 +193,8 @@ defined which task we wish to train our agent to complete. This is done by imple
 ```python
 class ObjectNavThorPPOExperimentConfig(rl_base.experiment_config.ExperimentConfig):
     ...
-    @staticmethod
-    def make_sampler_fn(**kwargs) -> rl_base.task.TaskSampler:
+    @classmethod
+    def make_sampler_fn(cls, **kwargs) -> rl_base.task.TaskSampler:
         return rl_ai2thor.object_nav.task_samplers.ObjectNavTaskSampler(**kwargs)
     ...
 ```

experiments/object_nav_thor_a2c.py

@@ -131,8 +131,8 @@ def create_model(cls, **kwargs) -> nn.Module:
             object_type_embedding_dim=8,
         )
 
-    @staticmethod
-    def make_sampler_fn(**kwargs) -> TaskSampler:
+    @classmethod
+    def make_sampler_fn(cls, **kwargs) -> TaskSampler:
         return ObjectNavTaskSampler(**kwargs)
 
     @staticmethod

experiments/object_nav_thor_dagger_ppo.py

@@ -2,13 +2,13 @@
 import torch.optim as optim
 from torch.optim.lr_scheduler import LambdaLR
 
-from .object_nav_thor_ppo import ObjectNavThorPPOExperimentConfig
 from onpolicy_sync.losses import PPO
 from onpolicy_sync.losses.imitation import Imitation
 from onpolicy_sync.losses.ppo import PPOConfig
 from rl_ai2thor.ai2thor_sensors import RGBSensorThor, GoalObjectTypeThorSensor
 from rl_base.sensor import ExpertActionSensor
 from utils.experiment_utils import LinearDecay, Builder, PipelineStage, TrainingPipeline
+from .object_nav_thor_ppo import ObjectNavThorPPOExperimentConfig
 
 
 class ObjectNavThorDAggerPPOExperimentConfig(ObjectNavThorPPOExperimentConfig):

experiments/object_nav_thor_ppo.py

@@ -9,7 +9,6 @@
 from torch.optim.lr_scheduler import LambdaLR
 
 from models.object_nav_models import ObjectNavBaselineActorCritic
-
 from onpolicy_sync.losses import PPO
 from onpolicy_sync.losses.ppo import PPOConfig
 from rl_ai2thor.ai2thor_sensors import RGBSensorThor, GoalObjectTypeThorSensor
@@ -139,8 +138,8 @@ def create_model(cls, **kwargs) -> nn.Module:
             object_type_embedding_dim=8,
         )
 
-    @staticmethod
-    def make_sampler_fn(**kwargs) -> TaskSampler:
+    @classmethod
+    def make_sampler_fn(cls, **kwargs) -> TaskSampler:
         return ObjectNavTaskSampler(**kwargs)
 
     @staticmethod

experiments/objectnav_habitat_base.py

@@ -1,22 +1,22 @@
-from typing import Dict, Any, List, Optional
 from abc import abstractmethod
+from typing import Dict, Any, List, Optional
 
 import gym
+import habitat
 import torch
 import torch.nn as nn
 import torch.optim as optim
 from torch.optim.lr_scheduler import LambdaLR
 
-import habitat
-from onpolicy_sync.losses.ppo import PPOConfig
 from models.point_nav_models import PointNavActorCriticResNet50GRU
 from onpolicy_sync.losses import PPO
+from onpolicy_sync.losses.ppo import PPOConfig
 from rl_base.experiment_config import ExperimentConfig
+from rl_base.preprocessor import ObservationSet
 from rl_base.sensor import SensorSuite
 from rl_base.task import TaskSampler
-from rl_base.preprocessor import ObservationSet
-from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_task_samplers import ObjectNavTaskSampler
+from rl_habitat.habitat_tasks import ObjectNavTask
 from utils.experiment_utils import Builder, PipelineStage, TrainingPipeline, LinearDecay
 
 

experiments/objectnav_habitat_depth_deterministic_resnet50GRU_ppo.py

@@ -1,16 +1,15 @@
 import gym
+import habitat
 import torch.nn as nn
 from torchvision import models
 
-import habitat
-
+from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 from models.object_nav_models import ObjectNavResNetActorCritic
 from rl_base.sensor import SensorSuite
-from rl_habitat.habitat_tasks import ObjectNavTask
+from rl_habitat.habitat_preprocessors import ResnetPreProcessorHabitat
 from rl_habitat.habitat_sensors import DepthSensorHabitat, TargetObjectSensorHabitat
+from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_utils import construct_env_configs
-from rl_habitat.habitat_preprocessors import ResnetPreProcessorHabitat
-from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 
 
 class ObjectNavHabitatDepthDeterministicResNet50GRUPPOExperimentConfig(

experiments/objectnav_habitat_depth_deterministic_simpleconvGRU_ppo.py

@@ -1,14 +1,13 @@
 import gym
-import torch.nn as nn
-
 import habitat
+import torch.nn as nn
 
+from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 from models.object_nav_models import ObjectNavBaselineActorCritic
 from rl_base.sensor import SensorSuite
-from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_sensors import DepthSensorHabitat, TargetObjectSensorHabitat
+from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_utils import construct_env_configs
-from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 
 
 class ObjectNavHabitatDebthDeterministicSimpleConvGRUPPOExperimentConfig(

experiments/objectnav_habitat_rgb_deterministic_resnet50GRU_ppo.py

@@ -1,16 +1,15 @@
 import gym
+import habitat
 import torch.nn as nn
 from torchvision import models
 
-import habitat
-
+from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 from models.object_nav_models import ObjectNavResNetActorCritic
 from rl_base.sensor import SensorSuite
-from rl_habitat.habitat_tasks import ObjectNavTask
+from rl_habitat.habitat_preprocessors import ResnetPreProcessorHabitat
 from rl_habitat.habitat_sensors import RGBSensorHabitat, TargetObjectSensorHabitat
+from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_utils import construct_env_configs
-from rl_habitat.habitat_preprocessors import ResnetPreProcessorHabitat
-from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 
 
 class ObjectNavHabitatRGBDeterministicResNet50GRUPPOExperimentConfig(

experiments/objectnav_habitat_rgb_deterministic_simpleconvGRU_ppo.py

@@ -1,14 +1,13 @@
 import gym
-import torch.nn as nn
-
 import habitat
+import torch.nn as nn
 
+from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 from models.object_nav_models import ObjectNavBaselineActorCritic
 from rl_base.sensor import SensorSuite
-from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_sensors import RGBSensorHabitat, TargetObjectSensorHabitat
+from rl_habitat.habitat_tasks import ObjectNavTask
 from rl_habitat.habitat_utils import construct_env_configs
-from experiments.objectnav_habitat_base import ObjectNavHabitatBaseExperimentConfig
 
 
 class ObjectNavHabitatRGBDeterministicSimpleConvGRUPPOExperimentConfig(