* Fix return type for load, learn in BaseAlgorithm
* Update changelog
* Add typing extensions to dependencies
* Import directly from typing for python >3.11
* Reorder changelog to reflect merge order
* Roll back to typevar solution
* Updated changelog
* Remove typing extensions requirement
* Update base_class.py
* Remove final point in changelog
* Additional type fixes across project
Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com>
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Fix replay_buffer_class type annotation
* Update changelog
* Further replacement of same type annotation issue
* Formatting
* Rolled back formatting changes for consistency
* Use higher resolution time and round up to eps
* Update changelog
* Add test case
* Fix formatting, time()->time_ns
* Bugfix: ns is integer not float
* Move test to better place
* Divide by 1e9 earlier
* Replacing the policy registry with policy "aliases"
* Fixing import order and SAC
* Changing arg. order to be sure policy_aliases is a kwarg
* Import orders
* Removing pytype error check
* Reformat
* Fix alias import
* Not using mutable {} as default for policy_aliases
* Empty aliases initialization
* Using static attributes for policy_aliases
* Fixing isort
* Fixing back bad merge
* Running isort
* Fixing aliases for A2C and PPO
* Using f-string
* Moving policy_aliases definition position
* Adding change in the changelog
* Update version
Co-authored-by: Antonin Raffin <antonin.raffin@ensta.org>
* Removing dead code for handling time limits (see #829)
* Mentionning remove_time_limit_termination in the changelog
* Update changelog.rst
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Add multi-env training support for SAC
* Fix for dict obs
* Pytype fixes
* Fix assert on number of envs
* Remove for loop
* Add support for Dict obs
* Start cleanup
* Update doc and bug fix
* Add support for vectorized action noise
and add multi env example for off-policy
* Update version
* Bug fix with VecNormalize
* Update README table
* Update variable names
* Update changelog and version
* Update doc and fix for `gradient_steps=-1`
* Add test for `gradient_steps=-1`
* Disable pytype pyi errors
* Fix for DQN
* Update comment on deepcopy
* Remove episode_reward field
* Fix RolloutReturn
* Avoid modification by reference
* Fix error message
Co-authored-by: Anssi <kaneran21@hotmail.com>
* Store number of timesteps at the beginning of each learn cycle
* Update changelog
* Set default _num_timesteps_at_start in the contructor
* Test case for FPS logger
* Adjust test to cover both on-policy and off-policy algorithms
* Fix formatting
* Update test and add comment
* Fix test
Co-authored-by: Oleksii Kachaiev <okachaiev@riotgames.com>
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Add `system_env_info`
* Add `print_system_info` to load
and store system info at save time
* Remove TODO
* Rename to `get_system_info`
* Import as sb3 for consistency
* Update changelog
* Add warning for old SB3 versions
* Use underscore litteral for more clarity
* Use a consistent key to log the total timesteps
This changes the timestep logging key of on-policy algorithms from
`time/total_timesteps` to `time/total timesteps` (note the
underscore/space). The off-policy algorithms and the eval callback
already use the latter, so this behavior is more consistent.
* Use underscores instead of spaces in logging keys
Most keys already followed this policy and consistent behavior is
friendlier to new users.
* Minor edit and bump version
Co-authored-by: Antonin Raffin <antonin.raffin@ensta.org>
* make sure DQN policy is always in correct mode - train or eval
* make set_training_mode an abstract method of the base policy - safer
* update docstring of _build method to note that the target network is put into eval mode
* use set_training_mode to put the dqn target network into eval mode
* use set_training_mode to set the training model of the q-network
* move set_training_mode abstract method from BasePolicy to BaseModel
* set train and eval mode for TD3
* make sure critic is always in correct mode during train
* set train and eval mode for SAC
* add comment re batch norm and dropout
* set train and eval mode for A2C and PPO
* add tests for collect rollouts with batch norm
* fix formatting
* update change log
* update version
* remove Optional typing for batch size - causing type check to fail
* Fix scipy dependency for toy text envs
* implement set_training_mode method in BaseModel
* move all tests of train/eval mode to test_train_eval_mode
* call learn with learning_starts = total_timesteps to test that collect_rollouts does not update batch norm
* remove extra calls to set_training_mode in train method of TD3 and SAC
* Allow gradient_steps=0
* Refactor tests
* Add comment + use aliases
* Typos
Co-authored-by: Antonin Raffin <antonin.raffin@ensta.org>
* training and evaluation: call model.train() and model.eval() to enable and disable dropout and batchnorm
* Add comment documentation
* Fix train and eval for the Actor class
* Run black
* Add github handle to changelog
* Add unit tests for PPO and DQN
* Refactor unit test
* Run black
* unit test: add a dropout layer and check that calling predict with deterministic=True is deterministic
* documentation: add bugfix description to changelog
* unit test: use learning_starts=0, decrease the size of the network and use more training steps
* on policy algorithms: call policy.train() and policy.eval() instead of disable_training and enable_training as it is a th.nn.module
* Rename unit test
* unit test: use drop out probability of 0.5
* Call policy.train and policy.eval
* Fixes + update tests
* Remove unneeded eval
Co-authored-by: David Blom <davidsblom@gmail.com>
Co-authored-by: Antonin Raffin <antonin.raffin@ensta.org>
* Add support for custom objects
* Add python 3.8 to the CI
* Bump version
* PyType fixes
* [ci skip] Fix typo
* Add note about slow-down + fix typos
* Minor edits to the doc
* Bug fix for DQN
* Update test
* Add test for custom objects
* Removed unneeded overrides of feature_extractor and normalize_images in the TD3 Actor.
* Add learning rate schedule example (#248)
* Add learning rate schedule example
* Update docs/guide/examples.rst
Co-authored-by: Adam Gleave <adam@gleave.me>
* Address comments
Co-authored-by: Adam Gleave <adam@gleave.me>
* Add supported action spaces checks (#254)
* Add supported action spaces checks
* Address comment
* Use `pass` in an abstractmethod instead of deleting the arguments.
* Remove the "deterministic" keyword from the forward method of the TD3 Actor since it always is deterministic anyways.
* Rename _get_data to _get_data_to_reconstruct_model.
_get_data was too generic and could have meant anything.
* Remove the n_episodes_rollout parameter and allow passing tuples as train_freq instead.
* Fix docstring of `train_freq` parameter.
* Black fixes.
* Fix TD3 delayed update + rename `_get_data()`
* Fix TD3 test
* Normalize `train_freq` to a tuple in the constructor and turn the warning into an assert.
* Make one step the default train frequency.
* Black fixes.
* Change np.bool to bool.
* Use the tuple format to specify an amount of steps in terms of steps or episodes in the collect_collouts of the off policy algorithm.
* Use the tuple format to specify an amount of steps in terms of steps or episodes in the collect_collouts of HER.
* Use named tuple for train freq
* Rename train_freq to train_every and TrainFreq to ExperienceDuration. Also add some type annotations and documentation.
* Black fixes.
* Revert to train_freq
* Fix terminal observation issues
* Typo
* Fix action noise bug in HER
* Add assert when loading HER models
* Update version
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
Co-authored-by: Adam Gleave <adam@gleave.me>
* Add callback signature to the learning rate type annotations.
* Add callback signature to the learning rate schedule type annotations.
* Add missing type annotations for learning rate callbacks.
* Add signature to old-style learning and evaluation callbacks.
* Add signature to env wrapper callback.
* Add type annotation to closure function.
* Use MaybeCallback more consistently.
* Update changelog.
* Remove now unused List import.
* Fix import order.
* Add type alias for learning rate schedules.
* Optimize imports.
* Fix messed up import.
* Remove resolved TODO.
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Small docstring improvements related to the notion of Rollout
* documented changes in changelog.rst, added myself to contributers
* Minor edits
Co-authored-by: Stefan Heid <stefan.heid@upb.de>
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Added working her version, Online sampling is missing.
* Updated test_her.
* Added first version of online her sampling. Still problems with tensor dimensions.
* Reformat
* Fixed tests
* Added some comments.
* Updated changelog.
* Add missing init file
* Fixed some small bugs.
* Reduced arguments for HER, small changes.
* Added getattr. Fixed bug for online sampling.
* Updated save/load funtions. Small changes.
* Added her to init.
* Updated save method.
* Updated her ratio.
* Move obs_wrapper
* Added DQN test.
* Fix potential bug
* Offline and online her share same sample_goal function.
* Changed lists into arrays.
* Updated her test.
* Fix online sampling
* Fixed action bug. Updated time limit for episodes.
* Updated convert_dict method to take keys as arguments.
* Renamed obs dict wrapper.
* Seed bit flipping env
* Remove get_episode_dict
* Add fast online sampling version
* Added documentation.
* Vectorized reward computation
* Vectorized goal sampling
* Update time limit for episodes in online her sampling.
* Fix max episode length inference
* Bug fix for Fetch envs
* Fix for HER + gSDE
* Reformat (new black version)
* Added info dict to compute new reward. Check her_replay_buffer again.
* Fix info buffer
* Updated done flag.
* Fixes for gSDE
* Offline her version uses now HerReplayBuffer as episode storage.
* Fix num_timesteps computation
* Fix get torch params
* Vectorized version for offline sampling.
* Modified offline her sampling to use sample method of her_replay_buffer
* Updated HER tests.
* Updated documentation
* Cleanup docstrings
* Updated to review comments
* Fix pytype
* Update according to review comments.
* Removed random goal strategy. Updated sample transitions.
* Updated migration. Removed time signal removal.
* Update doc
* Fix potential load issue
* Add VecNormalize support for dict obs
* Updated saving/loading replay buffer for HER.
* Fix test memory usage
* Fixed save/load replay buffer.
* Fixed save/load replay buffer
* Fixed transition index after loading replay buffer in online sampling
* Better error handling
* Add tests for get_time_limit
* More tests for VecNormalize with dict obs
* Update doc
* Improve HER description
* Add test for sde support
* Add comments
* Add comments
* Remove check that was always valid
* Fix for terminal observation
* Updated buffer size in offline version and reset of HER buffer
* Reformat
* Update doc
* Remove np.empty + add doc
* Fix loading
* Updated loading replay buffer
* Separate online and offline sampling + bug fixes
* Update tensorboard log name
* Version bump
* Bug fix for special case
Co-authored-by: Antonin Raffin <antonin.raffin@dlr.de>
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Add auto formatting with black and isort
* Reformat code
* Ignore typing errors
* Add note about line length
* Add minimum version for isort
* Add commit-checks
* Update docker image
* Fixed lost import (during last merge)
* Fix opencv dependency
* Add DDPG + TD3 with any number of critics
* Allow any number of critics for SAC
* Update doc
* [ci skip] Update DDPG example
* Remove unused parameter
* Add DDPG to identity test
* Fix computation with n_critics=1,3
* Update doc
* Apply suggestions from code review
Co-authored-by: Adam Gleave <adam@gleave.me>
* Update docstrings for off-policy algos
* Add check for sde
Co-authored-by: Adam Gleave <adam@gleave.me>
* Created DQN template according to the paper.
Next steps:
- Create Policy
- Complete Training
- Debug
* Changed Base Class
* refactor save, to be consistence with overriding the excluded_save_params function. Do not try to exclude the parameters twice.
* Added simple DQN policy
* Finished learn and train function
- missing correct loss computation
* changed collect_rollouts to work with discrete space
* moved discrete space collect_rollouts to dqn
* basic dqn working
* deleted SDE related code
* added gradient clipping and moved greedy policy to policy
* changed policy to implement target network
and added soft update(in fact standart tau is 1 so hard update)
* fixed policy setup
* rebase target_update_intervall on _n_updates
* adapted all tests
all tests passing
* Move to stable-baseline3
* Fixes for DQN
* Fix tests + add CNNPolicy
* Allow any optimizer for DQN
* added some util functions to create a arbitrary linear schedule, fixed pickle problem with old exploration schedule
* more documentation
* changed buffer dtype
* refactor and document
* Added Sphinx Documentation
Updated changelog.rst
* removed custom collect_rollouts as it is no longer necessary
* Implemented suggestions to clean code and documentation.
* extracted some functions on tests to reduce duplicated code
* added support for exploration_fraction
* Fixed exploration_fraction
* Added documentation
* Fixed get_linear_fn -> proper progress scaling
* Merged master
* Added nature reference
* Changed default parameters to https://www.nature.com/articles/nature14236/tables/1
* Fixed n_updates to be incremented correctly
* Correct train_freq
* Doc update
* added special parameter for DQN in tests
* different fix for test_discrete
* Update docs/modules/dqn.rst
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Update docs/modules/dqn.rst
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Update docs/modules/dqn.rst
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Added RMSProp in optimizer_kwargs, as described in nature paper
* Exploration fraction is inverse of 50.000.000 (total frames) / 1.000.000 (frames with linear schedule) according to nature paper
* Changelog update for buffer dtype
* standard exlude parameters should be always excluded to assure proper saving only if intentionally included by ``include`` parameter
* slightly more iterations on test_discrete to pass the test
* added param use_rms_prop instead of mutable default argument
* forgot alpha
* using huber loss, adam and learning rate 1e-4
* account for train_freq in update_target_network
* Added memory check for both buffers
* Doc updated for buffer allocation
* Added psutil Requirement
* Adapted test_identity.py
* Fixes with new SB3 version
* Fix for tensorboard name
* Convert assert to warning and fix tests
* Refactor off-policy algorithms
* Fixes
* test: remove next_obs in replay buffer
* Update changelog
* Fix tests and use tmp_path where possible
* Fix sampling bug in buffer
* Do not store next obs on episode termination
* Fix replay buffer sampling
* Update comment
* moved epsilon from policy to model
* Update predict method
* Update atari wrappers to match SB2
* Minor edit in the buffers
* Update changelog
* Merge branch 'master' into dqn
* Update DQN to new structure
* Fix tests and remove hardcoded path
* Fix for DQN
* Disable memory efficient replay buffer by default
* Fix docstring
* Add tests for memory efficient buffer
* Update changelog
* Split collect rollout
* Move target update outside `train()` for DQN
* Update changelog
* Update linear schedule doc
* Cleanup DQN code
* Minor edit
* Update version and docker images
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* save_replay_buffer now receives as argument the file path instead of the folder path
* Update changelog.rst
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
* Split torch module code into torch_layers file
* Updated reference to CNN
* Change 'CxWxH' to 'CxHxW', as per common notion
* Fix missing import in policies.py
* Move PPOPolicy to OnlineActorCriticPolicy
* Create OnPolicyRLModel from PPO, and make A2C and PPO inherit
* Update A2C optimizer comment
* Clean weight init scales for clarity
* Fix A2C log_interval default parameter
* Rename 'progress' to 'progress_remaining
* Rename 'Models' to 'Algorithms'
* Rename 'OnlineActorCriticPolicy' to 'ActorCriticPolicy'
* Move static functions out from BaseAlgorithm
* Move on/off_policy base algorithms to their own files
* Add files for A2C/PPO
* Fix docs
* Fix pytype
* Update documentation on OnPolicyAlgorithm
* Add proper doctstring for on_policy rollout gathering
* Add bit clarification on the mlppolicy/cnnpolicy naming
* Move static function is_vectorized_policies to utils.py
* Checking docstrings, pep8 fixes
* Update changelog
* Clean changelog
* Remove policy warnings for sac/td3
* Add monitor_wrapper for OnPolicyAlgorithm. Clean tb logging variables. Add parameter keywords to OffPolicyAlgorithm super init
Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org>
