b900fdf6f2
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>
81 lines
3.1 KiB
Python
81 lines
3.1 KiB
Python
from typing import Any, SupportsFloat
|
|
|
|
import gymnasium as gym
|
|
import numpy as np
|
|
from packaging import version
|
|
|
|
|
|
class ContinuousToDiscrete(gym.ActionWrapper):
|
|
"""Gym environment wrapper to take discrete action in a continuous environment.
|
|
|
|
:param gym.Env env: gym environment with continuous action space.
|
|
:param action_per_dim: number of discrete actions in each dimension
|
|
of the action space.
|
|
"""
|
|
|
|
def __init__(self, env: gym.Env, action_per_dim: int | list[int]) -> None:
|
|
super().__init__(env)
|
|
assert isinstance(env.action_space, gym.spaces.Box)
|
|
low, high = env.action_space.low, env.action_space.high
|
|
if isinstance(action_per_dim, int):
|
|
action_per_dim = [action_per_dim] * env.action_space.shape[0]
|
|
assert len(action_per_dim) == env.action_space.shape[0]
|
|
self.action_space = gym.spaces.MultiDiscrete(action_per_dim)
|
|
self.mesh = np.array(
|
|
[np.linspace(lo, hi, a) for lo, hi, a in zip(low, high, action_per_dim, strict=True)],
|
|
dtype=object,
|
|
)
|
|
|
|
def action(self, act: np.ndarray) -> np.ndarray: # type: ignore
|
|
# modify act
|
|
assert len(act.shape) <= 2, f"Unknown action format with shape {act.shape}."
|
|
if len(act.shape) == 1:
|
|
return np.array([self.mesh[i][a] for i, a in enumerate(act)])
|
|
return np.array([[self.mesh[i][a] for i, a in enumerate(a_)] for a_ in act])
|
|
|
|
|
|
class MultiDiscreteToDiscrete(gym.ActionWrapper):
|
|
"""Gym environment wrapper to take discrete action in multidiscrete environment.
|
|
|
|
:param gym.Env env: gym environment with multidiscrete action space.
|
|
"""
|
|
|
|
def __init__(self, env: gym.Env) -> None:
|
|
super().__init__(env)
|
|
assert isinstance(env.action_space, gym.spaces.MultiDiscrete)
|
|
nvec = env.action_space.nvec
|
|
assert nvec.ndim == 1
|
|
self.bases = np.ones_like(nvec)
|
|
for i in range(1, len(self.bases)):
|
|
self.bases[i] = self.bases[i - 1] * nvec[-i]
|
|
self.action_space = gym.spaces.Discrete(np.prod(nvec))
|
|
|
|
def action(self, act: np.ndarray) -> np.ndarray: # type: ignore
|
|
converted_act = []
|
|
for b in np.flip(self.bases):
|
|
converted_act.append(act // b)
|
|
act = act % b
|
|
return np.array(converted_act).transpose()
|
|
|
|
|
|
class TruncatedAsTerminated(gym.Wrapper):
|
|
"""A wrapper that set ``terminated = terminated or truncated`` for ``step()``.
|
|
|
|
It's intended to use with ``gym.wrappers.TimeLimit``.
|
|
|
|
:param gym.Env env: gym environment.
|
|
"""
|
|
|
|
def __init__(self, env: gym.Env):
|
|
super().__init__(env)
|
|
if not version.parse(gym.__version__) >= version.parse("0.26.0"):
|
|
raise OSError(
|
|
f"TruncatedAsTerminated is not applicable with gym version \
|
|
{gym.__version__}",
|
|
)
|
|
|
|
def step(self, act: np.ndarray) -> tuple[Any, SupportsFloat, bool, bool, dict[str, Any]]:
|
|
observation, reward, terminated, truncated, info = super().step(act)
|
|
terminated = terminated or truncated
|
|
return observation, reward, terminated, truncated, info
|