90 lines
3.1 KiB
Python
Raw Normal View History

from typing import Any
import numpy as np
from tianshou.data import (
Hindsight Experience Replay as a replay buffer (#753) ## implementation I implemented HER solely as a replay buffer. It is done by temporarily directly re-writing transitions storage (`self._meta`) during the `sample_indices()` call. The original transitions are cached and will be restored at the beginning of the next sampling or when other methods is called. This will make sure that. for example, n-step return calculation can be done without altering the policy. There is also a problem with the original indices sampling. The sampled indices are not guaranteed to be from different episodes. So I decided to perform re-writing based on the episode. This guarantees that the sampled transitions from the same episode will have the same re-written goal. This also make the re-writing ratio calculation slightly differ from the paper, but it won't be too different if there are many episodes in the buffer. In the current commit, HER replay buffer only support 'future' strategy and online sampling. This is the best of HER in term of performance and memory efficiency. I also add a few more convenient replay buffers (`HERVectorReplayBuffer`, `HERReplayBufferManager`), test env (`MyGoalEnv`), gym wrapper (`TruncatedAsTerminated`), unit tests, and a simple example (examples/offline/fetch_her_ddpg.py). ## verification I have added unit tests for almost everything I have implemented. HER replay buffer was also tested using DDPG on [`FetchReach-v3` env](https://github.com/Farama-Foundation/Gymnasium-Robotics). I used default DDPG parameters from mujoco example and didn't tune anything further to get this good result! (train script: examples/offline/fetch_her_ddpg.py). ![Screen Shot 2022-10-02 at 19 22 53](https://user-images.githubusercontent.com/42699114/193454066-0dd0c65c-fd5f-4587-8912-b441d39de88a.png)
2022-10-31 08:54:54 +09:00
HERReplayBuffer,
HERReplayBufferManager,
PrioritizedReplayBuffer,
PrioritizedReplayBufferManager,
ReplayBuffer,
ReplayBufferManager,
)
class VectorReplayBuffer(ReplayBufferManager):
"""VectorReplayBuffer contains n ReplayBuffer with the same size.
It is used for storing transition from different environments yet keeping the order
of time.
Remove kwargs in policy init (#950) Closes #947 This removes all kwargs from all policy constructors. While doing that, I also improved several names and added a whole lot of TODOs. ## Functional changes: 1. Added possibility to pass None as `critic2` and `critic2_optim`. In fact, the default behavior then should cover the absolute majority of cases 2. Added a function called `clone_optimizer` as a temporary measure to support passing `critic2_optim=None` ## Breaking changes: 1. `action_space` is no longer optional. In fact, it already was non-optional, as there was a ValueError in BasePolicy.init. So now several examples were fixed to reflect that 2. `reward_normalization` removed from DDPG and children. It was never allowed to pass it as `True` there, an error would have been raised in `compute_n_step_reward`. Now I removed it from the interface 3. renamed `critic1` and similar to `critic`, in order to have uniform interfaces. Note that the `critic` in DDPG was optional for the sole reason that child classes used `critic1`. I removed this optionality (DDPG can't do anything with `critic=None`) 4. Several renamings of fields (mostly private to public, so backwards compatible) ## Additional changes: 1. Removed type and default declaration from docstring. This kind of duplication is really not necessary 2. Policy constructors are now only called using named arguments, not a fragile mixture of positional and named as before 5. Minor beautifications in typing and code 6. Generally shortened docstrings and made them uniform across all policies (hopefully) ## Comment: With these changes, several problems in tianshou's inheritance hierarchy become more apparent. I tried highlighting them for future work. --------- Co-authored-by: Dominik Jain <d.jain@appliedai.de>
2023-10-08 17:57:03 +02:00
:param total_size: the total size of VectorReplayBuffer.
:param buffer_num: the number of ReplayBuffer it uses, which are under the same
configuration.
Other input arguments (stack_num/ignore_obs_next/save_only_last_obs/sample_avail)
are the same as :class:`~tianshou.data.ReplayBuffer`.
.. seealso::
Please refer to :class:`~tianshou.data.ReplayBuffer` for other APIs' usage.
"""
def __init__(self, total_size: int, buffer_num: int, **kwargs: Any) -> None:
assert buffer_num > 0
size = int(np.ceil(total_size / buffer_num))
buffer_list = [ReplayBuffer(size, **kwargs) for _ in range(buffer_num)]
super().__init__(buffer_list)
class PrioritizedVectorReplayBuffer(PrioritizedReplayBufferManager):
"""PrioritizedVectorReplayBuffer contains n PrioritizedReplayBuffer with same size.
It is used for storing transition from different environments yet keeping the order
of time.
Remove kwargs in policy init (#950) Closes #947 This removes all kwargs from all policy constructors. While doing that, I also improved several names and added a whole lot of TODOs. ## Functional changes: 1. Added possibility to pass None as `critic2` and `critic2_optim`. In fact, the default behavior then should cover the absolute majority of cases 2. Added a function called `clone_optimizer` as a temporary measure to support passing `critic2_optim=None` ## Breaking changes: 1. `action_space` is no longer optional. In fact, it already was non-optional, as there was a ValueError in BasePolicy.init. So now several examples were fixed to reflect that 2. `reward_normalization` removed from DDPG and children. It was never allowed to pass it as `True` there, an error would have been raised in `compute_n_step_reward`. Now I removed it from the interface 3. renamed `critic1` and similar to `critic`, in order to have uniform interfaces. Note that the `critic` in DDPG was optional for the sole reason that child classes used `critic1`. I removed this optionality (DDPG can't do anything with `critic=None`) 4. Several renamings of fields (mostly private to public, so backwards compatible) ## Additional changes: 1. Removed type and default declaration from docstring. This kind of duplication is really not necessary 2. Policy constructors are now only called using named arguments, not a fragile mixture of positional and named as before 5. Minor beautifications in typing and code 6. Generally shortened docstrings and made them uniform across all policies (hopefully) ## Comment: With these changes, several problems in tianshou's inheritance hierarchy become more apparent. I tried highlighting them for future work. --------- Co-authored-by: Dominik Jain <d.jain@appliedai.de>
2023-10-08 17:57:03 +02:00
:param total_size: the total size of PrioritizedVectorReplayBuffer.
:param buffer_num: the number of PrioritizedReplayBuffer it uses, which are
under the same configuration.
Other input arguments (alpha/beta/stack_num/ignore_obs_next/save_only_last_obs/
sample_avail) are the same as :class:`~tianshou.data.PrioritizedReplayBuffer`.
.. seealso::
Please refer to :class:`~tianshou.data.ReplayBuffer` for other APIs' usage.
"""
def __init__(self, total_size: int, buffer_num: int, **kwargs: Any) -> None:
assert buffer_num > 0
size = int(np.ceil(total_size / buffer_num))
buffer_list = [PrioritizedReplayBuffer(size, **kwargs) for _ in range(buffer_num)]
super().__init__(buffer_list)
def set_beta(self, beta: float) -> None:
for buffer in self.buffers:
buffer.set_beta(beta)
Hindsight Experience Replay as a replay buffer (#753) ## implementation I implemented HER solely as a replay buffer. It is done by temporarily directly re-writing transitions storage (`self._meta`) during the `sample_indices()` call. The original transitions are cached and will be restored at the beginning of the next sampling or when other methods is called. This will make sure that. for example, n-step return calculation can be done without altering the policy. There is also a problem with the original indices sampling. The sampled indices are not guaranteed to be from different episodes. So I decided to perform re-writing based on the episode. This guarantees that the sampled transitions from the same episode will have the same re-written goal. This also make the re-writing ratio calculation slightly differ from the paper, but it won't be too different if there are many episodes in the buffer. In the current commit, HER replay buffer only support 'future' strategy and online sampling. This is the best of HER in term of performance and memory efficiency. I also add a few more convenient replay buffers (`HERVectorReplayBuffer`, `HERReplayBufferManager`), test env (`MyGoalEnv`), gym wrapper (`TruncatedAsTerminated`), unit tests, and a simple example (examples/offline/fetch_her_ddpg.py). ## verification I have added unit tests for almost everything I have implemented. HER replay buffer was also tested using DDPG on [`FetchReach-v3` env](https://github.com/Farama-Foundation/Gymnasium-Robotics). I used default DDPG parameters from mujoco example and didn't tune anything further to get this good result! (train script: examples/offline/fetch_her_ddpg.py). ![Screen Shot 2022-10-02 at 19 22 53](https://user-images.githubusercontent.com/42699114/193454066-0dd0c65c-fd5f-4587-8912-b441d39de88a.png)
2022-10-31 08:54:54 +09:00
class HERVectorReplayBuffer(HERReplayBufferManager):
"""HERVectorReplayBuffer contains n HERReplayBuffer with same size.
It is used for storing transition from different environments yet keeping the order
of time.
Remove kwargs in policy init (#950) Closes #947 This removes all kwargs from all policy constructors. While doing that, I also improved several names and added a whole lot of TODOs. ## Functional changes: 1. Added possibility to pass None as `critic2` and `critic2_optim`. In fact, the default behavior then should cover the absolute majority of cases 2. Added a function called `clone_optimizer` as a temporary measure to support passing `critic2_optim=None` ## Breaking changes: 1. `action_space` is no longer optional. In fact, it already was non-optional, as there was a ValueError in BasePolicy.init. So now several examples were fixed to reflect that 2. `reward_normalization` removed from DDPG and children. It was never allowed to pass it as `True` there, an error would have been raised in `compute_n_step_reward`. Now I removed it from the interface 3. renamed `critic1` and similar to `critic`, in order to have uniform interfaces. Note that the `critic` in DDPG was optional for the sole reason that child classes used `critic1`. I removed this optionality (DDPG can't do anything with `critic=None`) 4. Several renamings of fields (mostly private to public, so backwards compatible) ## Additional changes: 1. Removed type and default declaration from docstring. This kind of duplication is really not necessary 2. Policy constructors are now only called using named arguments, not a fragile mixture of positional and named as before 5. Minor beautifications in typing and code 6. Generally shortened docstrings and made them uniform across all policies (hopefully) ## Comment: With these changes, several problems in tianshou's inheritance hierarchy become more apparent. I tried highlighting them for future work. --------- Co-authored-by: Dominik Jain <d.jain@appliedai.de>
2023-10-08 17:57:03 +02:00
:param total_size: the total size of HERVectorReplayBuffer.
:param buffer_num: the number of HERReplayBuffer it uses, which are
Hindsight Experience Replay as a replay buffer (#753) ## implementation I implemented HER solely as a replay buffer. It is done by temporarily directly re-writing transitions storage (`self._meta`) during the `sample_indices()` call. The original transitions are cached and will be restored at the beginning of the next sampling or when other methods is called. This will make sure that. for example, n-step return calculation can be done without altering the policy. There is also a problem with the original indices sampling. The sampled indices are not guaranteed to be from different episodes. So I decided to perform re-writing based on the episode. This guarantees that the sampled transitions from the same episode will have the same re-written goal. This also make the re-writing ratio calculation slightly differ from the paper, but it won't be too different if there are many episodes in the buffer. In the current commit, HER replay buffer only support 'future' strategy and online sampling. This is the best of HER in term of performance and memory efficiency. I also add a few more convenient replay buffers (`HERVectorReplayBuffer`, `HERReplayBufferManager`), test env (`MyGoalEnv`), gym wrapper (`TruncatedAsTerminated`), unit tests, and a simple example (examples/offline/fetch_her_ddpg.py). ## verification I have added unit tests for almost everything I have implemented. HER replay buffer was also tested using DDPG on [`FetchReach-v3` env](https://github.com/Farama-Foundation/Gymnasium-Robotics). I used default DDPG parameters from mujoco example and didn't tune anything further to get this good result! (train script: examples/offline/fetch_her_ddpg.py). ![Screen Shot 2022-10-02 at 19 22 53](https://user-images.githubusercontent.com/42699114/193454066-0dd0c65c-fd5f-4587-8912-b441d39de88a.png)
2022-10-31 08:54:54 +09:00
under the same configuration.
Other input arguments are the same as :class:`~tianshou.data.HERReplayBuffer`.
.. seealso::
Please refer to :class:`~tianshou.data.ReplayBuffer` for other APIs' usage.
"""
def __init__(self, total_size: int, buffer_num: int, **kwargs: Any) -> None:
assert buffer_num > 0
size = int(np.ceil(total_size / buffer_num))
buffer_list = [HERReplayBuffer(size, **kwargs) for _ in range(buffer_num)]
super().__init__(buffer_list)