* Implement behavioral cloning for cc/dc, fc/rnn, state/observations.
* Re-organize folder structure in anticipation of unitytrainers as a package.
* Create demo environment BananaImitation to validate behavioral cloning.
* Fixes#336
* Reorganized python tests into separate folder, and make individiual test files for different (sub) modules.
* Add tests for trainer_controller, PPO, and behavioral cloning. More to come soon.
* Minor bug fixes discovered while writing tests.
* Reworked GirdWorld to reset much faster.
* Cleaned ObservationToTex and reworked GetObservationMatrixList to be 3x faster.
* On Demand Decision : Use RequestDecision and RequestAction
* New Agent Inspector : Use it to set On Demand Decision
* New BrainParameters interface
* LSTM memory size is now set in python
* New C# API
* Semantic Changes
* Replaced RunMDP
* New Bouncer Environment to test On Demand Dscision
* [Previous Text Actions] Renamed previous_action to previous_vector_action
added previous_text_action to the BrainInfo
* [Semantics] Carried the modifications to the semantics of previous_vector_action to the trainers
Fixes the following issues:
* Missing component reference in BananaRL environment.
* Neural Network for multiple visual observations was not properly generated.
* Episode time-out value estimate bootstrapping used incorrect observation as input.
This PR makes the following changes:
* Moves clipping of continuous control model into model itself. Output is now always [-1, 1].
* Internal model values are now clipped between [-3, 3] before being rescaled to [-1, 1] for output. * This improves training performance by providing a wider range of values within which the pdf of the gaussian can fall. Output of [-1, 1] is used to be more environment-creator friendly.
* Fixes issue where epsilon was erroneously being used to reconstruct old probabilities during PPO update, leading to reduced learning performance.
* Introduce ScaleAction() function within python to easily rescale values from [-1, 1] to arbitrary range.
* Re-train all CC models using improved algorithm. All performance levels are equal or improved. In the case of Crawler, improvement is drastic.
* Update documentation appropriately.
* Made miscellaneous minor code style and optimization improvements within environments.
* [Cold Fix] Split the way cummulative rewards and episode length are counted
The reward is appended at each step to the cummulative reward
The episode count is ONLY incremented when d_t+1 is false
* Adds implementation of Curiosity-driven Exploration by Self-supervised Prediction (https://arxiv.org/abs/1705.05363) to PPO trainer.
* To enable, set use_curiosity flag to true in hyperparameter file.
* Includes refactor of unitytrainers model code to accommodate new feature.
* Adds new Pyramids environment (w/ documentation). Environment contains sparse reward, and can only be solved using PPO+Curiosity.
In the case the agent is done imediately after spawning, its stats are empty because the stats need at least 2 successive experieces to create the stats.
By specifying the default value of 0, the error does no longer appear
* [Initial Commit]
Modified the model.py file and the ppo/trainer.py file to use masked actions
* Preliminary modifications to the python side of the code to enable action masking
* Preliminary modifications to the C# side of the code to enable action masking
* Preliminary modifications to the communication side of the code to enable action masking
* Implemented action masking for BC
Note : The actions of the teacher are not masked
* More error messages for the action masking
* fix pytests
* Added Documentation
* Address comment
* Addressed Comments on docs
* Addressed second comment on docs
* Addressed comments for the python side of the code
* Created the action masker and associated unit tests
* Addressed comments on the C# side
* Addressed the comment regarding action_masking_name
* Addressed the comments
* Initial Commit
* attempt at refactor
* Put all static methods into the CoreInternalBrain
* improvements
* more testing
* modifications
* renamed epsilon
* misc
* Now supports discrete actions
* added discrete support and RNN and visual. Left to do is refactor and save variables into models
* code cleaning
* made a tensor generator and applier
* fix on the models.py file
* Moved the Checks to a different Class
* Added some unit tests
* BugFix
* Need to generate the output tensors as well as inputs before executing the graph
* Made NodeNames static and created a new namespace
* Added comments to the TensorAppliers
* Started adding comments on the TensorGenerators code
* Added comments for the Tensor Generator
* Moving the helper classes into a separate folder
* Added initial comments to the TensorChecks
* Renamed NodeNames -> TensorNames
* Removing warnings in tests
* Now using Aut...
* Enable buffer padding to be set other than 0
Allows buffer padding in AgentBufferField to be set to a custom value. In particular, 0-padding for `action_masks` causes a divide-by-zero error, and should be padded with 1’s instead.
This is done as a parameter passed to the `append` method, so that the pad value can be set right after the instantiation of an AgentBufferField.
* Move 'take_action' into Policy class
This refactor is part of Actor-Trainer separation. Since policies
will be distributed across actors in separate processes which share
a single trainer, taking an action should be the responsibility of
the policy.
This change makes a few smaller changes:
* Combines `take_action` logic between trainers, making it more
generic
* Adds an `ActionInfo` data class to be more explicit about the
data returned by the policy, only used by TrainerController and
policy for now.
* Moves trainer stats logic out of `take_action` and into
`add_experiences`
* Renames 'take_action' to 'get_action'
This commit adds support for running Unity environments in parallel.
An abstract base class was created for UnityEnvironment which a new
SubprocessUnityEnvironment inherits from.
SubprocessUnityEnvironment communicates through a pipe in order to
send commands which will be run in parallel to its workers.
A few significant changes needed to be made as a side-effect:
* UnityEnvironments are created via a factory method (a closure)
rather than being directly created by the main process.
* In mlagents-learn "worker-id" has been replaced by "base-port"
and "num-envs", and worker_ids are automatically assigned across runs.
* BrainInfo objects now convert all fields to numpy arrays or lists to
avoid serialization issues.
* WIP precommit on top level
* update CI
* circleci fixes
* intentionally fail black
* use --show-diff-on-failure in CI
* fix command order
* rebreak a file
* apply black
* WIP enable mypy
* run mypy on each package
* fix trainer_metrics mypy errors
* more mypy errors
* more mypy
* Fix some partially typed functions
* types for take_action_outputs
* fix formatting
* cleanup
* generate stubs for proto objects
* fix ml-agents-env mypy errors
* disallow-incomplete-defs for gym-unity
* Add CI notes to CONTRIBUTING.md
* Create new class (RewardSignal) that represents a reward signal.
* Add value heads for each reward signal in the PPO model.
* Make summaries agnostic to the type of reward signals, and log weighted rewards per reward signal.
* Move extrinsic and curiosity rewards into this new structure.
* Allow defining multiple reward signals in YAML file. Add documentation for this new structure.
At each step, an unused `last_reward` variable in the TF graph is
updated in our PPO trainer. There are also related unused methods
in various places in the codebase. This change removes them.
Previously in v0.8 we added parallel environments via the
SubprocessUnityEnvironment, which exposed the same abstraction as
UnityEnvironment while actually wrapping many parallel environments
via subprocesses.
Wrapping many environments with the same interface as a single
environment had some downsides, however:
* Ordering needed to be preserved for agents across different envs,
complicating the SubprocessEnvironment logic
* Asynchronous environments with steps taken out of sync with the
trainer aren't viable with the Environment abstraction
This PR introduces a new EnvManager abstraction which exposes a
reduced subset of the UnityEnvironment abstraction and a
SubprocessEnvManager implementation which replaces the
SubprocessUnityEnvironment.
* 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
* don't need fwd reference
* cleanup root
* always write timers, add comments
* undo accidental change
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)
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
* 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
* 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.
- Move common functions to trainer.py, model.pyfromppo/trainer.py, ppo/policy.pyandppo/model.py'
- Introduce RLTrainer class and move most of add_experiences and some common reward
signal code there. PPO and SAC will inherit from this, not so much BC Trainer.
- Add methods to Buffer to enable sampling, truncating, and save/loading.
- Add scoping to create encoders in model.py
* Adds evaluate_batch to reward signals. Evaluates on minibatch rather than on BrainInfo.
* Changes the way reward signal results are reported in rl_trainer so that we get the pure, unprocessed environment reward separate from the reward signals.
* Moves end_episode to rl_trainer
* Fixed bug with BCModule with RNN
* Normalize observations when adding experiences
This change moves normalization of vector observations into the trainer's
"add_experiences" interface.
Prior to this change, normalization occurred at inference time. This
was somewhat confusing since usually executing a forward pass shouldn't
have side-effects which would change the training step. Also, in a
asynchronous or distributed setting where we copy the neural network
weights from a trainer to a remote actor / inference worker we'd end up
with training issues because of the weights being different on the trainer
than the workers.
We have been ignoring unused imports and star imports via flake8. These are
both bad practice and grow over time without automated checking. This
commit attempts to fix all existing import errors and add back the corresponding
flake8 checks.
Our multi-GPU training had a regression such that freezing the
graph was broken. This change fixes that issue by making a few
changes:
* Removes the top level "tower" variable scope added by multi-GPU
so that the output nodes have correct names
* Removes the use of "freeze_graph" and replaces it with our own similar
functionality.
* Adds the "auto reuse" to network layers which require them
* Initial commit removing memories from C# and deprecating memory fields in proto
* initial changes to Python
* Adding functionalities
* Fixes
* adding the memories to the dictionary
* Fixing bugs
* tweeks
* Resolving bugs
* Recreating the proto
* Addressing comments
* Passing by reference does not work. Do not merge
* Fixing huge bug in Inference
* Applying patches
* fixing tests
* Addressing comments
* Renaming variable to reflect type
* test
* Modifying the .proto files
* attempt 1 at refactoring Python
* works for ppo hallway
* changing the documentation
* now works with both sac and ppo both training and inference
* Ned to fix the tests
* TODOs :
- Fix the demonstration recorder
- Fix the demonstration loader
- verify the intrinsic reward signals work
- Fix the tests on Python
- Fix the C# tests
* Regenerating the protos
* fix proto typo
* protos and modifying the C# demo recorder
* modified the demo loader
* Demos are loading
* IMPORTANT : THESE ARE THE FILES USED FOR CONVERSION FROM OLD TO NEW FORMAT
* Modified all the demo files
* Fixing all the tests
* fixing ci
* addressing comments
* removing reference to memories in the ll-api
This is the first in a series of PRs that intend to move the agent processing logic (add_experiences and process_experiences) out of the trainer and into a separate class. The plan is to do so in steps:
- Split the processing buffers (keeping track of agent trajectories and assembling trajectories) and update buffer (complete trajectories to be used for training) within the Trainer (this PR)
- Move the processing buffer and add/process experiences into a separate, outside class
- Change the data type of the update buffer to be a Trajectory
- Place and read Trajectories from queues, add subscription mechanism for both AgentProcessor and Trainers
* added team id and identifier concat to behavior parameters
* splitting brain params into brain name and identifiers
* set team id in prefab
* recieves brain_name and identifier on python side
* added team id and identifier concat to behavior parameters
* splitting brain params into brain name and identifiers
* set team id in prefab
* recieves brain_name and identifier on python side
* rebased with develop
* Correctly calls concatBehaviorIdentifiers
* added team id and identifier concat to behavior parameters
* splitting brain params into brain name and identifiers
* set team id in prefab
* recieves brain_name and identifier on python side
* rebased with develop
* Correctly calls concatBehaviorIdentifiers
* trainer_controller expects name_behavior_ids
* add_policy and create_policy separated
* adjusting tests to expect trainer.add_policy to be called
* fixing tests
* fixed naming ...
* [bug-fix] Increase height of wall in CrawlerStatic (#3650)
* [bug-fix] Improve performance for PPO with continuous actions (#3662)
* Corrected a typo in a name of a function (#3670)
OnEpsiodeBegin was corrected to OnEpisodeBegin in Migrating.md document
* Add Academy.AutomaticSteppingEnabled to migration (#3666)
* Fix editor port in Dockerfile (#3674)
* Hotfix memory leak on Python (#3664)
* Hotfix memory leak on Python
* Fixing
* Fixing a bug in the heuristic policy. A decision should not be requested when the agent is done
* [bug-fix] Make Python able to deal with 0-step episodes (#3671)
* adding some comments
Co-authored-by: Ervin T <ervin@unity3d.com>
* Remove vis_encode_type from list of required (#3677)
* Update changelog (#3678)
* Shorten timeout duration for environment close (#3679)
The timeout duration for closing an environment was set to the
same duration as the timeout when waiting ...
* [bug-fix] Fix regression in --initialize-from feature (#4086)
* Fixed text in GettingStarted page specifying the logdir for tensorboard. Before it was in a directory summaries which no longer existed. Results are now saved to the results dir. (#4085)
* [refactor] Remove nonfunctional `output_path` option from TrainerSettings (#4087)
* Reverting bug introduced in #4071 (#4101)
Co-authored-by: Scott <Scott.m.jordan91@gmail.com>
Co-authored-by: Vincent-Pierre BERGES <vincentpierre@unity3d.com>
* Experiment branch for comparing torch
* Updates and merging ervin changes
* improvements on experiment_torch.py
* Better printing of results
* preliminary gpu experiment
* Testing gpu
* Prepare to see a lot of commits, because I like my IDE and I am testing on a server and I am using git to sync the two
* Prepare to see a lot of commits, because I like my IDE and I am testing on a server and I am using git to sync the two
* _
* _
* _
* _
* _
* _
* _
* _
* Attempt at gpu on tf. Does not work
* _
* _
* _
* _
* _
* _
* _
* _
* _
* _
* _
* Fixing learn.py
* Update Dockerfile
* Separate send environment data from reset (#4128)
* Fixed a typo on ML-Agents-Overview.md (#4130)
Fixed redundant "to" word from the sentence since it is probably a typo in document.
* Updated the badge’s link to point to the newest doc version
* Replaced all of the doc to release_3_doc
* Fix 3DBall and 3DBallHard SAC regressions (#4132)
* Move memory validation to settings
* Update docs
* Add settings test
* Update to release_3 in installation.md (#4144)
* rename to SideChannelManager +backcompat (#4137)
* Remove comment about logo with --help (#4148)
* [bugfix] Make FoodCollector heuristic playable (#4147)
* Make FoodCollector heuristic playable
* Update changelog
* script to check for old release links and references (#4153)
* Remove package validation suite from Project (#4146)
* RayPerceptionSensor: handle empty and invalid tags (#4155...
* Added Reward Providers for Torch
* Use NetworkBody to encode state in the reward providers
* Integrating the reward prodiders with ppo and torch
* work in progress, integration with PPO. Not training properly Pyramids at the moment
* Integration in PPO
* Removing duplicate file
* Gail and Curiosity working
* addressing comments
* Enfore float32 for tests
* enfore np.float32 in buffer
This change adds an export to .nn for each checkpoint generated by
RLTrainer and adds a NNCheckpointManager to track the generated
checkpoints and final model in training_status.json.
Co-authored-by: Jonathan Harper <jharper+moar@unity3d.com>
* Begin porting work
* Add ResNet and distributions
* Dynamically construct actor and critic
* Initial optimizer port
* Refactoring policy and optimizer
* Resolving a few bugs
* Share more code between tf and torch policies
* Slightly closer to running model
* Training runs, but doesn’t actually work
* Fix a couple additional bugs
* Add conditional sigma for distribution
* Fix normalization
* Support discrete actions as well
* Continuous and discrete now train
* Mulkti-discrete now working
* Visual observations now train as well
* GRU in-progress and dynamic cnns
* Fix for memories
* Remove unused arg
* Combine actor and critic classes. Initial export.
* Support tf and pytorch alongside one another
* Prepare model for onnx export
* Use LSTM and fix a few merge errors
* Fix bug in probs calculation
* Optimize np -> tensor operations
* Time action sample funct...
* Move linear encoding to NetworkBody
* moved encoders to processors (#4420)
* fix bad merge
* Get it running
* Replace mentions of visual_encoders
* Remove output_size property
* Fix tests
* Fix some references
* Revert test_simple_rl
* Fix networks test
* Make curiosity test more accomodating
* Rename total_input_size
* [Bug fix] Fix bug in GAIL gradient penalty (#4425) (#4426)
Co-authored-by: Vincent-Pierre BERGES <vincentpierre@unity3d.com>
* Up number of steps
* Rename to visual_processors and vector_processors
Co-authored-by: andrewcoh <54679309+andrewcoh@users.noreply.github.com>
Co-authored-by: Andrew Cohen <andrew.cohen@unity3d.com>
Co-authored-by: Vincent-Pierre BERGES <vincentpierre@unity3d.com>
* Add torch_utils
* Use torch from torch_utils
* Add torch to banned modules in CI
* Better import error handling
* Fix flake8 errors
* Address comments
* Move networks to GPU if enabled
* Switch to torch_utils
* More flake8 problems
* Move reward providers to GPU/CPU
* Remove anothere set default tensor
* Fix banned import in test
* Moved components to the tf folder and moved the TrainerFactory to the `trainer` folder
* Addressing comments
* Editing the migrating doc
* fixing test
* Torch setup.py
* Set torch to default
* Make torch default in setup.py
* Remove indents
* Remove other instances of TF being used
* Add tensorboard to setup.py
* Adding correst setup commands for verifying torch is installed (#4524)
* Adding correst setup commands for verifying torch is installed
* Editing the test_requirments to add tf and remove torch
* Develop torchdefault raise outside setup (#4530)
* Torch not imported error to raise at first usage
* Torch not imported error to raise at first usage
* [refactor] Use PyTorch TensorBoard utils (#4518)
* Convert stats writer to use PyTorch TB support
* Use common function to print params
* Update test
* Bump tensorboard to 1.15 to fix the tests
* putting tensorboard 1.15.0 as min version requirement
Co-authored-by: vincentpierre <vincentpierre@unity3d.com>
* [Docs] Initial documentation changes for making...
* Proper dimensions for entropy, sum before bonus in PPO
* Make entropy reporting same as TF
* Always use separate critic
* Revert to shared
* Remove unneeded extra line
* Change entropy shape in test
* Change another entropy shape
* Add entropy summing to evaluate_actions
* Add notes about torch.abs(policy_loss)
* Make buffer type-agnostic
* Edit types of Apped method
* Change comment
* Collaborative walljump
* Make collab env harder
* Add group ID
* Add collab obs to trajectory
* Fix bug; add critic_obs to buffer
* Set group ids for some envs
* Pretty broken
* Less broken PPO
* Update SAC, fix PPO batching
* Fix SAC interrupted condition and typing
* Fix SAC interrupted again
* Remove erroneous file
* Fix multiple obs
* Update curiosity reward provider
* Update GAIL and BC
* Multi-input network
* Some minor tweaks but still broken
* Get next critic observations into value estimate
* Temporarily disable exporting
* Use Vince's ONNX export code
* Cleanup
* Add walljump collab YAML
* Lower max height
* Update prefab
* Update prefab
* Collaborative Hallway
* Set num teammates to 2
* Add config and group ids to HallwayCollab
* Fix bug with hallway collab
* E...
* Fix end episode for POCA, add warning for group reward if not POCA (#5113)
* Fix end episode for POCA, add warning for group reward if not POCA
* Add missing imports
* Use np.any, which is faster
* Don't clear update buffer, but don't append to it either
* Update changelog
* Address comments
* Make experience replay buffer saving more verbose
(cherry picked from commit 63e7ad44d96b7663b91f005ca1d88f4f3b11dd2a)
GitHub
eshvk
Arthur Juliani
Vincent Gao
Marwan Mattar
Joe Ward
vincentpierre
unityjeffrey
Deric Pang
Ervin T
Vincent(Yuan) Gao
Jonathan Harper
Chris Elion
Ervin Teng
Anupam Bhatnagar
Andrew Cohen
Christopher Goy
Ruo-Ping Dong
HH
brccabral