Fix 2 bugs and refactor RunningMeanStd to support dict obs norm (#695)

* fix #689

* fix #672

* refactor RMS class

* fix #688

docs/tutorials/cheatsheet.rst

@@ -1,9 +1,13 @@
 Cheat Sheet
 ===========
 
-This page shows some code snippets of how to use Tianshou to develop new algorithms / apply algorithms to new scenarios.
+This page shows some code snippets of how to use Tianshou to develop new
+algorithms / apply algorithms to new scenarios.
 
-By the way, some of these issues can be resolved by using a ``gym.Wrapper``. It could be a universal solution in the policy-environment interaction. But you can also use the batch processor :ref:`preprocess_fn`.
+By the way, some of these issues can be resolved by using a ``gym.Wrapper``.
+It could be a universal solution in the policy-environment interaction. But
+you can also use the batch processor :ref:`preprocess_fn` or vectorized
+environment wrapper :class:`~tianshou.env.VectorEnvWrapper`.
 
 
 .. _network_api:
@@ -22,6 +26,18 @@ Build New Policy
 See :class:`~tianshou.policy.BasePolicy`.
 
 
+.. _eval_policy:
+
+Manually Evaluate Policy
+------------------------
+
+If you'd like to manually see the action generated by a well-trained agent:
+::
+
+    # assume obs is a single environment observation
+    action = policy(Batch(obs=np.array([obs]))).act[0]
+
+
 .. _customize_training:
 
 Customize Training Process

docs/tutorials/dqn.rst

@@ -256,6 +256,12 @@ Watch the Agent's Performance
     collector = ts.data.Collector(policy, env, exploration_noise=True)
     collector.collect(n_episode=1, render=1 / 35)
 
+If you'd like to manually see the action generated by a well-trained agent:
+::
+
+    # assume obs is a single environment observation
+    action = policy(Batch(obs=np.array([obs]))).act[0]
+
 
 .. _customized_trainer:
 

test/base/test_env.py

@@ -211,6 +211,14 @@ def test_vecenv(size=10, num=8, sleep=0.001):
         v.close()
 
 
+def test_attr_unwrapped():
+    train_envs = DummyVectorEnv([lambda: gym.make("CartPole-v1")])
+    train_envs.set_env_attr("test_attribute", 1337)
+    assert train_envs.get_env_attr("test_attribute") == [1337]
+    assert hasattr(train_envs.workers[0].env, "test_attribute")
+    assert hasattr(train_envs.workers[0].env.unwrapped, "test_attribute")
+
+
 def test_env_obs_dtype():
     for obs_type in ["array", "object"]:
         envs = SubprocVectorEnv(
@@ -349,6 +357,7 @@ if __name__ == '__main__':
     test_venv_wrapper_envpool()
     test_env_obs_dtype()
     test_vecenv()
+    test_attr_unwrapped()
     test_async_env()
     test_async_check_id()
     test_env_reset_optional_kwargs()

tianshou/env/venv_wrappers.py

@@ -68,24 +68,17 @@ class VectorEnvNormObs(VectorEnvWrapper):
     """An observation normalization wrapper for vectorized environments.
 
     :param bool update_obs_rms: whether to update obs_rms. Default to True.
-    :param float clip_obs: the maximum absolute value for observation. Default to
-        10.0.
-    :param float epsilon: To avoid division by zero.
     """
 
     def __init__(
         self,
         venv: BaseVectorEnv,
         update_obs_rms: bool = True,
-        clip_obs: float = 10.0,
-        epsilon: float = np.finfo(np.float32).eps.item(),
     ) -> None:
         super().__init__(venv)
         # initialize observation running mean/std
         self.update_obs_rms = update_obs_rms
         self.obs_rms = RunningMeanStd()
-        self.clip_max = clip_obs
-        self.eps = epsilon
 
     def reset(
         self,
@@ -127,8 +120,7 @@ class VectorEnvNormObs(VectorEnvWrapper):
 
     def _norm_obs(self, obs: np.ndarray) -> np.ndarray:
         if self.obs_rms:
-            obs = (obs - self.obs_rms.mean) / np.sqrt(self.obs_rms.var + self.eps)
-            obs = np.clip(obs, -self.clip_max, self.clip_max)
+            return self.obs_rms.norm(obs)  # type: ignore
         return obs
 
     def set_obs_rms(self, obs_rms: RunningMeanStd) -> None:

tianshou/env/worker/dummy.py

@@ -17,7 +17,7 @@ class DummyEnvWorker(EnvWorker):
         return getattr(self.env, key)
 
     def set_env_attr(self, key: str, value: Any) -> None:
-        setattr(self.env, key, value)
+        setattr(self.env.unwrapped, key, value)
 
     def reset(self, **kwargs: Any) -> Union[np.ndarray, Tuple[np.ndarray, dict]]:
         if "seed" in kwargs:

tianshou/env/worker/ray.py

@@ -14,7 +14,7 @@ except ImportError:
 class _SetAttrWrapper(gym.Wrapper):
 
     def set_env_attr(self, key: str, value: Any) -> None:
-        setattr(self.env, key, value)
+        setattr(self.env.unwrapped, key, value)
 
     def get_env_attr(self, key: str) -> Any:
         return getattr(self.env, key)

tianshou/env/worker/subproc.py

@@ -49,9 +49,9 @@ def _setup_buf(space: gym.Space) -> Union[dict, tuple, ShArray]:
     if isinstance(space, gym.spaces.Dict):
         assert isinstance(space.spaces, OrderedDict)
         return {k: _setup_buf(v) for k, v in space.spaces.items()}
-    elif isinstance(space, gym.spaces.Tuple):
-        assert isinstance(space.spaces, tuple)
-        return tuple([_setup_buf(t) for t in space.spaces])
+    elif isinstance(space, gym.spaces.Tuple):  # type: ignore
+        assert isinstance(space.spaces, tuple)  # type: ignore
+        return tuple([_setup_buf(t) for t in space.spaces])  # type: ignore
     else:
         return ShArray(space.dtype, space.shape)  # type: ignore
 
@@ -122,7 +122,7 @@ def _worker(
             elif cmd == "getattr":
                 p.send(getattr(env, data) if hasattr(env, data) else None)
             elif cmd == "setattr":
-                setattr(env, data["key"], data["value"])
+                setattr(env.unwrapped, data["key"], data["value"])
             else:
                 p.close()
                 raise NotImplementedError

tianshou/utils/statistics.py

@@ -1,5 +1,5 @@
 from numbers import Number
-from typing import List, Union
+from typing import List, Optional, Union
 
 import numpy as np
 import torch
@@ -70,15 +70,31 @@ class RunningMeanStd(object):
     """Calculates the running mean and std of a data stream.
 
     https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
+
+    :param mean: the initial mean estimation for data array. Default to 0.
+    :param std: the initial standard error estimation for data array. Default to 1.
+    :param float clip_max: the maximum absolute value for data array. Default to
+        10.0.
+    :param float epsilon: To avoid division by zero.
     """
 
     def __init__(
         self,
         mean: Union[float, np.ndarray] = 0.0,
-        std: Union[float, np.ndarray] = 1.0
+        std: Union[float, np.ndarray] = 1.0,
+        clip_max: Optional[float] = 10.0,
+        epsilon: float = np.finfo(np.float32).eps.item(),
     ) -> None:
         self.mean, self.var = mean, std
+        self.clip_max = clip_max
         self.count = 0
+        self.eps = epsilon
+
+    def norm(self, data_array: Union[float, np.ndarray]) -> Union[float, np.ndarray]:
+        data_array = (data_array - self.mean) / np.sqrt(self.var + self.eps)
+        if self.clip_max:
+            data_array = np.clip(data_array, -self.clip_max, self.clip_max)
+        return data_array
 
     def update(self, data_array: np.ndarray) -> None:
         """Add a batch of item into RMS with the same shape, modify mean/var/count."""