94bfb32cc1
1. add policy.eval() in all test scripts' "watch performance" 2. remove dict return support for collector preprocess_fn 3. add `__contains__` and `pop` in batch: `key in batch`, `batch.pop(key, deft)` 4. exact n_episode for a list of n_episode limitation and save fake data in cache_buffer when self.buffer is None (#184) 5. fix tensorboard logging: h-axis stands for env step instead of gradient step; add test results into tensorboard 6. add test_returns (both GAE and nstep) 7. change the type-checking order in batch.py and converter.py in order to meet the most often case first 8. fix shape inconsistency for torch.Tensor in replay buffer 9. remove `**kwargs` in ReplayBuffer 10. remove default value in batch.split() and add merge_last argument (#185) 11. improve nstep efficiency 12. add max_batchsize in onpolicy algorithms 13. potential bugfix for subproc.wait 14. fix RecurrentActorProb 15. improve the code-coverage (from 90% to 95%) and remove the dead code 16. fix some incorrect type annotation The above improvement also increases the training FPS: on my computer, the previous version is only ~1800 FPS and after that, it can reach ~2050 (faster than v0.2.4.post1).
84 lines
2.4 KiB
Python
84 lines
2.4 KiB
Python
import numpy as np
|
|
from typing import Union, Optional
|
|
from abc import ABC, abstractmethod
|
|
|
|
|
|
class BaseNoise(ABC, object):
|
|
"""The action noise base class."""
|
|
|
|
def __init__(self, **kwargs) -> None:
|
|
super().__init__()
|
|
|
|
@abstractmethod
|
|
def __call__(self, **kwargs) -> np.ndarray:
|
|
"""Generate new noise."""
|
|
raise NotImplementedError
|
|
|
|
def reset(self) -> None:
|
|
"""Reset to the initial state."""
|
|
pass
|
|
|
|
|
|
class GaussianNoise(BaseNoise):
|
|
"""Class for vanilla gaussian process,
|
|
used for exploration in DDPG by default.
|
|
"""
|
|
|
|
def __init__(self,
|
|
mu: float = 0.0,
|
|
sigma: float = 1.0):
|
|
super().__init__()
|
|
self._mu = mu
|
|
assert 0 <= sigma, 'noise std should not be negative'
|
|
self._sigma = sigma
|
|
|
|
def __call__(self, size: tuple) -> np.ndarray:
|
|
return np.random.normal(self._mu, self._sigma, size)
|
|
|
|
|
|
class OUNoise(BaseNoise):
|
|
"""Class for Ornstein-Uhlenbeck process, as used for exploration in DDPG.
|
|
Usage:
|
|
::
|
|
|
|
# init
|
|
self.noise = OUNoise()
|
|
# generate noise
|
|
noise = self.noise(logits.shape, eps)
|
|
|
|
For required parameters, you can refer to the stackoverflow page. However,
|
|
our experiment result shows that (similar to OpenAI SpinningUp) using
|
|
vanilla gaussian process has little difference from using the
|
|
Ornstein-Uhlenbeck process.
|
|
"""
|
|
|
|
def __init__(self,
|
|
mu: float = 0.0,
|
|
sigma: float = 0.3,
|
|
theta: float = 0.15,
|
|
dt: float = 1e-2,
|
|
x0: Optional[Union[float, np.ndarray]] = None
|
|
) -> None:
|
|
super(BaseNoise, self).__init__()
|
|
self._mu = mu
|
|
self._alpha = theta * dt
|
|
self._beta = sigma * np.sqrt(dt)
|
|
self._x0 = x0
|
|
self.reset()
|
|
|
|
def __call__(self, size: tuple, mu: Optional[float] = None) -> np.ndarray:
|
|
"""Generate new noise. Return a ``numpy.ndarray`` which size is equal
|
|
to ``size``.
|
|
"""
|
|
if self._x is None or self._x.shape != size:
|
|
self._x = 0
|
|
if mu is None:
|
|
mu = self._mu
|
|
r = self._beta * np.random.normal(size=size)
|
|
self._x = self._x + self._alpha * (mu - self._x) + r
|
|
return self._x
|
|
|
|
def reset(self) -> None:
|
|
"""Reset to the initial state."""
|
|
self._x = self._x0
|