This PR seeks to bring habitat-API inline with this. StaticPointGoalSensor is now just a PointGoalSensor and the old PointGoalSensor is now a PointGoalWithGPSCompassSensor. There are also now explicit GPS and Compass sensors, such that you can fairly easily compute PointGoalWithGPSCompassSensor given [PointGoalSensor, GPSSensor, CompassSensor].
The PointGoal* sensors now also include {2d, 3d} x {Cartesian, Polar} support.

Explicitly call close for vector env to avoid issues with file writer prematurely close buffer when tensorboard enabled. See #184 for more info.

Making eval() in ppo_trainer retrieve the name (sensor_uuid) of performance metric used instead of always using spl. This change is necessary for ppo_trainer to support new tasks that might have performance metric other than spl.

With the recent refactor/redesign of habitat_baselines, all experiment parameters can be specified through config files. Thus it makes sense to turn the python script run in circle-ci into a separate test file in our test/. This also makes it more extensible when we add more trainers in the future.

* fix baseline test import and docstring

* lr scheduler and tb writer

* refactor config

* More Env typing fixes

* Move log_file config entry

* fix tb-related VectorEnv del bug

* change config entries to list

* fix env class retrieval

* change default opts to empty list

* changes according to comments

* Refactor baselines to allow for extension

* Make clip an after-backward

* Loop over rollout step instead of function for full rollout

Update baseline test import and fix docstring 

* refactor and  add generic trainer class

* fix to pass tests

* change BaseModel to BaseTrainer

* fix tensorboard import causing CI failure

* modify circle-ci test script accordingly

* doc, typing and other changes

* rename BASELINE to TRAINER

* merge from upstream master

* Update Habitat-API to allow for no rendering sensors (#139)

Update Habitat-API to allow for no rendering sensors

* Added installation requirements step for sim installation in CI setup. (#159)

Added installation requirements step for sim installation in CI setup

* move RolloutStorage to utils

* add environments.py

* make ckpt dir if not exist

* small fixes according to comments

* changes according to comments

* update readme

* fix old config compatibility

* fix missed isort lint

* episode stats changed to dict of dicts

* handle non-unique episode_id

* use tuple as key

* update while loop checking

Co-Authored-By: Erik Wijmans <ewijmans2@gmail.com>

* Add checkpoint progress tracking for evalute_ppo. Now when specified with a checkpoint directory,
* Evaluate_ppo will evaluate checkpoints in chronological order, and constantly check for new checkpoint.
* Add tensorboard visualization to both train_ppo and evaluate_ppo
* Add video generation for navigation episode evaluation. Generated videos can be either saved locally or visualized through tensorboard.
* Add shortest path visualization

Fix issue with POINTNAVV1 fix.

SimulatorActions is no longer an enum. It is an extendable singleton. It allows you to extend the action space but it is quite strict about it. Actions cannot be put on different ints, actions cannot be remove.
There is now a registry for action space configurations.

* Modify partitioning logic to handle left-over scenes due to integer division

Added description how to use config system, fixed missing copyright headers, improved example and isort CI (#125)

* Added description how to use config system related to #20.
* Added missing copyright headers
* Improve registry example to be cleaner
* Adde habitat_sim as first party lib to isort CI setup

* Generalize PPO with both PointNav goal sensors
* generalize simple agents and fix lint
* Removed dataset specific nodes

Updated baselines readme, fixed links, added link to checkpoints.

Expose other measures through get_info in NavRLEnv

Fixes and upgrades the collision logic.

Adds documentation for habitat-api, to view the docs follow instructions in docs/README.md and view the html files from the build folder.

- add config option NUM_SAMPLE_EPISODE, with default value -1 that
indicates use all episodes. When set value to non-negative integer,
pointna_dataset will sample the specified number of episodes.
- fix the functionality of merging multiple config files. Now can specify
more than one config file like this
	--task-config "a.yaml" "b.yaml"

Our registry pattern is far from idea and thus we haven't used it consistent. This PR aims to resolve this by copying the registry pattern used in Pythia (as it is much nicer) and then registering all current measures and sensors and using the register to query for those instead of getattr as this is a more clearly extendable pattern (as compared to monkey patching).

* Builds and installs habitat-sim with python setup.py install
* Added pytorch installation and running habitat baseline test in prediction mode. That maybe too much of testing, but decided to keep it for now.
* Have ideas how to speed up and parallelize jobs. Will follow up in next PRs.
* Run isort and black on whole code to make it happy.
* Includes parallelization of installation steps and use files as flags that step is finished. CircleCI has no good way to parallelize steps and have dependencies one the same machine.
* Cache is used to save conda dependencies installed and build simulator based on commit SHA.

* Fix the broken tests by attr pr

* Update main README to properly install gym

* Update baseline README for gym install

Changing action space names to say what they actually do, adding the look up and look down actions.

* add extras_require in setup.py

* add exclude in setup.py for baselines

* add pytest as requirement

* Implement accessing setup() params within custom commands

Renamed baselines folder to habitat_baselines