Support gym without configure (#162)

* Major surgery to avoid gym.Env.configure

* track flashgames

* Forward configure through Timer-Render-Throttle

Logger now takes print_frequency in constructor

* More configure fixes. py.test passes locally

* Upgrade to gym 0.8

* Require 0.8.1

* configure bugfixes

* Record env metadata & spec. Use semantics.async instead of spec.vnc

* Bump tox version of gym for travis

* Fix race condition when recording from multiple threads on non-async envs

* sub-bump

* Bump

setup.py

@@ -1,7 +1,7 @@
 from setuptools import setup, find_packages
 
 setup(name='universe',
-      version='0.21.4-dev',
+      version='0.21.5',
       description="Universe: a software platform for measuring and training an AI's general intelligence across the world's supply of games, websites and other applications.",
       url='https://github.com/openai/universe',
       author='OpenAI',
@@ -14,7 +14,7 @@
           'docker-pycreds==0.2.1',
           'fastzbarlight>=0.0.13',
           'go-vncdriver>=0.4.8',
-          'gym>=0.7,<0.8', # gym v0.8 removes env.configure, which required for this version of universe
+          'gym>=0.8.1',
           'Pillow>=3.3.0',
           'PyYAML>=3.12',
           'six>=1.10.0',

tests/functional/test_core_envs_semantics.py

@@ -90,9 +90,10 @@ def test_nice_vnc_semantics_match(spec, matcher, wrapper):
     spaces.seed(0)
 
     vnc_env = spec.make()
+    if vnc_env.metadata.get('configure.required', False):
+        vnc_env.configure(remotes=1)
     vnc_env = wrapper(vnc_env)
     vnc_env = wrappers.Unvectorize(vnc_env)
-    vnc_env.configure(remotes=1)
 
     env = gym.make(spec._kwargs['gym_core_id'])
 

tests/functional/test_envs.py

@@ -26,12 +26,13 @@ def test_smoke(env_id):
     logging.getLogger().setLevel(logging.INFO)
 
     env = gym.make(env_id)
-    env = wrappers.Unvectorize(env)
+    if env.metadata.get('configure.required', False):
+        if os.environ.get('FORCE_LATEST_UNIVERSE_DOCKER_RUNTIMES'):  # Used to test universe-envs in CI
+            configure_with_latest_docker_runtime_tag(env)
+        else:
+            env.configure(remotes=1)
 
-    if os.environ.get('FORCE_LATEST_UNIVERSE_DOCKER_RUNTIMES'):  # Used to test universe-envs in CI
-        configure_with_latest_docker_runtime_tag(env)
-    else:
-        env.configure(remotes=1)
+    env = wrappers.Unvectorize(env)
 
     env.reset()
     _rollout(env, timestep_limit=60*30) # Check a rollout

tox.ini

@@ -11,7 +11,7 @@ skipsdist=True
 passenv=DISPLAY DOCKER_USERNAME DOCKER_PASSWORD FORCE_LATEST_UNIVERSE_DOCKER_RUNTIMES TRAVIS*
 deps =
     pytest
-    gym[atari]<0.8
+    gym[atari]<0.9
     docker-py==1.10.3
     Pillow
     autobahn

universe/envs/dummy_vnc_env.py

@@ -36,7 +36,7 @@ def __init__(self):
         self.observation_space = spaces.VNCObservationSpace()
         self.action_space = spaces.VNCActionSpace()
 
-    def _configure(self, remotes=None,
+    def configure(self, remotes=None,
                    client_id=None,
                    start_timeout=None, docker_image=None,
                    ignore_clock_skew=False, disable_action_probes=False,
@@ -45,7 +45,6 @@ def _configure(self, remotes=None,
                    observer=False,
                    _n=3,
     ):
-        super(DummyVNCEnv, self)._configure()
         self.n = _n
         self._reward_buffers = [rewarder.RewardBuffer('dummy:{}'.format(i)) for i in range(self.n)]
         self._started = True

universe/envs/vnc_env.py

@@ -135,18 +135,18 @@ def _seed(self, seed):
         self._seed_value = seed
         return [seed]
 
-    def _configure(self, remotes=None,
-                   client_id=None,
-                   start_timeout=None, docker_image=None,
-                   ignore_clock_skew=False, disable_action_probes=False,
-                   vnc_driver=None, vnc_kwargs=None,
-                   rewarder_driver=None,
-                   replace_on_crash=False, allocate_sync=True,
-                   observer=False, api_key=None,
-                   record=False,
-                   sample_env_ids=None,
+    def configure(self, remotes=None,
+                  client_id=None,
+                  start_timeout=None, docker_image=None,
+                  ignore_clock_skew=False, disable_action_probes=False,
+                  vnc_driver=None, vnc_kwargs=None,
+                  rewarder_driver=None,
+                  replace_on_crash=False, allocate_sync=True,
+                  observer=False, api_key=None,
+                  record=False,
+                  sample_env_ids=None,
     ):
-        """Standard Gym hook to configure the environment.
+        """Universe method to configure the environment.
 
         Args:
 
@@ -552,7 +552,11 @@ def _render(self, mode='human', close=False):
                 self.vnc_session.render(self.connection_names[0])
 
     def __str__(self):
-        return 'VNCEnv<{}>'.format(self.spec.id)
+        if self.spec:
+            return '<VNCEnv{}>'.format(self.spec.id)
+        else:
+            return 'VNCEnv'
+
 
 class Mask(object):
     """Blocks the agent from interacting with the environment while the

universe/vectorized/core.py

@@ -15,11 +15,12 @@ class Env(gym.Env):
     # Number of remotes. User should set this.
     n = None
 
+
 class Wrapper(Env, gym.Wrapper):
     """Use this instead of gym.Wrapper iff you're wrapping a vectorized env,
     (or a vanilla env you wish to be vectorized).
     """
-    # If True and this is instantiated with a non-vectorized environment, 
+    # If True and this is instantiated with a non-vectorized environment,
     # automatically wrap it with the Vectorize wrapper.
     autovectorize = True
 
@@ -35,10 +36,12 @@ def __init__(self, env):
 
         self.env = env
 
-    def _configure(self, **kwargs):
-        super(Wrapper, self)._configure(**kwargs)
-        assert self.env.n is not None, "Did not set self.env.n: self.n={} self.env={} self={}".format(self.env.n, self.env, self)
-        self.n = self.env.n
+    @property
+    def n(self):
+        return self.env.n
+
+    def configure(self, **kwargs):
+        self.env.configure(**kwargs)
 
 class ObservationWrapper(Wrapper, gym.ObservationWrapper):
     pass
@@ -48,4 +51,3 @@ class RewardWrapper(Wrapper, gym.RewardWrapper):
 
 class ActionWrapper(Wrapper, gym.ActionWrapper):
     pass
-

universe/vectorized/multiprocessing_env.py

@@ -261,6 +261,7 @@ def close_n(worker_n):
 class MultiprocessingEnv(core.Env):
     metadata = {
         'runtime.vectorized': True,
+        'configure.required': True,
     }
 
     def __init__(self, env_id):
@@ -278,8 +279,7 @@ def __init__(self, env_id):
         self.observation_space = env.observation_space
         self.reward_range = env.reward_range
 
-    def _configure(self, n=1, pool_size=None, episode_limit=None):
-        super(MultiprocessingEnv, self)._configure()
+    def configure(self, n=1, pool_size=None, episode_limit=None):
         self.n = n
         self.envs = [self.spec.make() for _ in range(self.n)]
 

universe/vectorized/tests/test_monitoring.py

@@ -8,8 +8,8 @@
 def test_multiprocessing_env_monitoring():
     with helpers.tempdir() as temp:
         env = wrappers.WrappedMultiprocessingEnv('Pong-v3')
-        env = wrappers.Monitor(env, temp)
         env.configure(n=2)
+        env = wrappers.Monitor(env, temp)
         env.reset()
         for i in range(2):
             env.step([0, 0])
@@ -23,10 +23,10 @@ def test_multiprocessing_env_monitoring():
 def test_vnc_monitoring():
     with helpers.tempdir() as temp:
         env = gym.make('gym-core.Pong-v3')
+        env.configure(remotes=2)
         env = wrappers.GymCoreAction(env)
         env = wrappers.Monitor(env, temp)
 
-        env.configure(remotes=2)
         env.reset()
         for i in range(2):
             env.step([0, 0])

universe/vectorized/vectorize_filter.py

@@ -18,14 +18,13 @@ class VectorizeFilter(core.Wrapper):
     def __init__(self, env, filter_factory, *args, **kwargs):
         super(VectorizeFilter, self).__init__(env)
         self.filter_factory = filter_factory
+        self.filter_n = None
         self._args = args
         self._kwargs = kwargs
 
-    def _configure(self, **kwargs):
-        super(VectorizeFilter, self)._configure(**kwargs)
-        self.filter_n = [self.filter_factory(*self._args, **self._kwargs) for _ in range(self.n)]
-
     def _reset(self):
+        if self.filter_n is None:
+            self.filter_n = [self.filter_factory(*self._args, **self._kwargs) for _ in range(self.n)]
         observation_n = self.env.reset()
         observation_n = [filter._after_reset(observation) for filter, observation in zip(self.filter_n, observation_n)]
         return observation_n

universe/wrappers/diagnostics.py

@@ -6,15 +6,12 @@
 
 # Not used in core; but used in play_flashgames
 class Diagnostics(vectorized.Wrapper):
-    def _configure(self, **kwargs):
-        super(Diagnostics, self)._configure(**kwargs)
-        self.diagnostics = self.unwrapped.diagnostics
 
     def _step(self, action_n):
         observation_n, reward_n, done_n, info = self.env.step(action_n)
         # We want this to be above Mask, so we know whether or not a
         # particular index is resetting.
-        if self.diagnostics:
+        if self.unwrapped.diagnostics:
             with pyprofile.push('vnc_env.diagnostics.add_metadata'):
-                self.diagnostics.add_metadata(observation_n, info['n'])
+                self.unwrapped.diagnostics.add_metadata(observation_n, info['n'])
         return observation_n, reward_n, done_n, info

universe/wrappers/experimental/random_env.py

@@ -14,8 +14,8 @@ def __init__(self, env, env_ids):
         super(RandomEnv, self).__init__(env)
         self.env_ids = env_ids
 
-    def _configure(self, **kwargs):
-        super(RandomEnv, self)._configure(sample_env_ids=self.env_ids, **kwargs)
+    def configure(self, **kwargs):
+        super(RandomEnv, self).configure(sample_env_ids=self.env_ids, **kwargs)
 
     def _reset(self):
         observation_n = self.env.reset()

universe/wrappers/joint.py

@@ -4,9 +4,6 @@
 class Joint(vectorized.Wrapper):
     def __init__(self, env_m):
         self.env_m = env_m
-        for env in self.env_m:
-            if not env._configured:
-                raise error.Error('Joint env should have been initialized: {}'.format(env))
 
         # TODO: generalize this. Doing so requires adding a vectorized
         # space mode.
@@ -15,26 +12,21 @@ def __init__(self, env_m):
 
         self.pool = pool.ThreadPool(min(len(env_m), 5))
 
-    def _close(self):
-        if hasattr(self, 'pool'):
-            self.pool.close()
+        self._n = sum(env.n for env in self.env_m)
+        self.metadata = self.metadata.copy()
+        self.metadata['render.modes'] = self.env_m[0].metadata['render.modes']
 
     @property
-    def spec(self):
-        return None
+    def n(self):
+        return self._n
 
-    @spec.setter
-    def spec(self, value):
-        pass
+    def _close(self):
+        if hasattr(self, 'pool'):
+            self.pool.close()
 
     def _render(self, mode='human', close=False):
         return self.env_m[0]._render(mode=mode, close=close)
 
-    def _configure(self, **kwargs):
-        self.n = sum(env.n for env in self.env_m)
-        self.metadata = self.metadata.copy()
-        self.metadata['render.modes'] = self.env_m[0].metadata['render.modes']
-
     def _reset(self):
         # Keep all env[0] action on the main thread, in case we ever
         # need to render. Otherwise we get segfaults from the

universe/wrappers/logger.py

@@ -23,18 +23,21 @@ def stats(count):
     return s
 
 class Logger(vectorized.Wrapper):
-    def __init__(self, env):
+    metadata = {
+        'configure.required': True
+    }
+    def __init__(self, env, print_frequency=5):
         super(Logger, self).__init__(env)
-
-    def _configure(self, print_frequency=5, **kwargs):
         self.print_frequency = print_frequency
         extra_logger.info('Running VNC environments with Logger set to print_frequency=%s. To change this, pass "print_frequency=k" or "print_frequency=None" to "env.configure".', self.print_frequency)
-        super(Logger, self)._configure(**kwargs)
-        self._clear_step_state()
-        self.metadata['render.modes'] = self.env.metadata['render.modes']
-
+        if self.n is not None:
+            self._clear_step_state()
         self._last_step_time = None
 
+    def configure(self, **kwargs):
+        self.env.configure(**kwargs)
+        self._clear_step_state()
+
     def _clear_step_state(self):
         self.frames = 0
         self.last_print = time.time()

universe/wrappers/multiprocessing_env.py

@@ -30,13 +30,15 @@ def _step(self, action_n):
         return observation_n, reward_n, done_n, info
 
 class EpisodeID(vectorized.Wrapper):
+    metadata = {
+        'configure.required': True
+    }
     """
-For each episode, return its id, and also return the total number of contiguous 
-episodes that are now done. 
+For each episode, return its id, and also return the total number of contiguous
+episodes that are now done.
 """
-    def _configure(self, episode_limit=None, **kwargs):
-        super(EpisodeID, self)._configure(**kwargs)
-        assert self.metadata.get('runtime.vectorized')
+    def configure(self, episode_limit=None, **kwargs):
+        self.env.configure(**kwargs)
         self.episode_limit = episode_limit
         self._clear_state()