* Fix naming conventions for consistency
* Add generalization link to ML-Agents Overview
* Add generalization to main Readme
* Include types of samplers available for use
* Removed obsolete 'TestDstWrongShape' test as it does not reflect how Barracuda tensors work
* Added proper test cleanup, to avoid warning messages from finalizer thread.
* Add Sampler and SamplerManager
* Enable resampling of reset parameters during training
* Documentation for Sampler and example YAML configuration file
Bringing bucket of temp memory allocation optimizations:
* switched to Barracuda backed tensor across the board, helps to leverage allocators and reuse of the internal buffers
* added Barracuda 0.2.4 release, which bring another set of temp memory allocation fixes
* Timer proof-of-concept
* micro optimizations
* add some timers
* cleanup, add asserts
* Cleanup (no start/end methods) and handle exceptions
* unit test and decorator
* move output code, add a decorator
* cleanup
* module docstring
* actually write the timings when done with training
* use __qualname__ instead
* add a few more timers
* fix mock import
* fix unit test
* get timers from worker process (WIP)
* clean up timer merging
* typo
* WIP
* cleanup merging code
* bad merge
* undo accidental change
* remove reset command
* fix style
* fix unit tests
* fix unit tests (they got overwrote in merge)
* get timer root though a function
* timer around communicate
* Removes unused SubprocessEnvManager import in trainer_controller
* Removes unused `steps` argument to `TrainerController._save_model`
* Consolidates unnecessary branching for curricula in
`TrainerController.advance`
* Moves `reward_buffer` into `TFPolicy` from `PPOPolicy` and adds
`BCTrainer` support so that we don't have a broken interface /
undefined behavior when BCTrainer is used with curricula.
* Don't 0 value bootstrap for GAIL and Curiosity
* Add gradient penalties to GAN to help with stability
* Add gail_config.yaml with GAIL examples
* Cleaned up trainer_config.yaml and unnecessary gammas
* Documentation updates
* Code cleanup
SubprocessEnvManager takes steps synchronously to reproduce old
behavior, meaning all parallel environments will need to wait for
the slowest environment to take a step. If some steps take much
longer than others, this can lead to a substantial overall slowdown
in practice. We've seen extreme cases where we see almost a 2x
speedup from using asynchronous stepping, with no downside for our
faster environments. (Bouncer 16% improvement, Walker 14% improvement
in tests).
This PR changes the SubprocessEnvManager to use async stepping.
This means on the "step" call the environment manager will enqueue
step requests to workers, and then only wait until at least one
step has been completed before returning.
Fixes shuffling issue with newer versions of numpy (#1798).
* make get_value_estimates output a dict of floats
* Use np.append instead of convert to list, unconvert
* Add type hints and test for get_value_estimates
Based on the new reward signals architecture, add BC pretrainer and GAIL for PPO. Main changes:
- A new GAILRewardSignal and GAILModel for GAIL/VAIL
- A BCModule component (not a reward signal) to do pretraining during RL
- Documentation for both of these
- Change to Demo Loader that lets you load multiple demo files in a folder
- Example Demo files for all of our tested sample environments (for future regression testing)