Better handle "--More--" messages in NLE (#23)

* auto more wrapper

* keep the behavior the same

* fix order

* add unwrapped to wrapper works

Author: BartekCupial

balrog/config/config.yaml

@@ -47,7 +47,7 @@ envs:
     no_progress_timeout: 150    # Timeout for no progress in NLE
     savedir: null               # Directory to save NLE data; null disables saving
     save_ttyrec_every: 0        # Frequency of saving TTY recordings
-    skip_more: False            # Whether to skip the 'more' prompt in NLE
+    skip_more: True            # Whether to skip the 'more' prompt in NLE
   minihack_kwargs: 
     character: "@"
     max_episode_steps: 100
@@ -57,7 +57,7 @@ envs:
     savedir: null
     save_ttyrec_every: 0
     autopickup: False           
-    skip_more: False
+    skip_more: True
   babyai_kwargs:
     num_dists: 0                
   crafter_kwargs:

balrog/environments/minihack/minihack_env.py

@@ -3,7 +3,7 @@
 import gym
 import minihack  # NOQA: F401
 
-from balrog.environments.nle import NLELanguageWrapper
+from balrog.environments.nle import AutoMore, NLELanguageWrapper
 from balrog.environments.wrappers import GymV21CompatibilityV0, NLETimeLimit
 
 MINIHACK_ENVS = []
@@ -30,7 +30,9 @@ def make_minihack_env(env_name, task, config, render_mode: Optional[str] = None)
         ],
         **minihack_kwargs,
     )
-    env = NLELanguageWrapper(env, vlm=vlm, skip_more=skip_more)
+    if skip_more:
+        env = AutoMore(env)
+    env = NLELanguageWrapper(env, vlm=vlm)
 
     # wrap NLE with timeout
     env = NLETimeLimit(env)

balrog/environments/nle/__init__.py

@@ -1,5 +1,6 @@
 import enum
 
+from .auto_more import AutoMore
 from .base import NLELanguageWrapper
 
 

balrog/environments/nle/auto_more.py

@@ -0,0 +1,33 @@
+import gym
+from nle import nle_language_obsv
+from nle.nethack import actions as A
+
+
+class AutoMore(gym.Wrapper):
+    def __init__(self, env):
+        super().__init__(env)
+        self.nle_language = nle_language_obsv.NLELanguageObsv()
+
+    def reset(self, **kwargs):
+        obs = super().reset(**kwargs)
+        obs["text_message"] = self.nle_language.text_message(obs["tty_chars"]).decode("latin-1")
+
+        return obs
+
+    def step(self, action):
+        obs, reward, done, info = super().step(action)
+
+        message = self.nle_language.text_message(obs["tty_chars"]).decode("latin-1")
+
+        while "--More--" in message and not done:
+            message = message.replace("--More--", "\n")
+
+            action_index = self.env.actions.index(A.MiscAction.MORE)
+            obs, rew, done, info = super().step(action_index)
+            add = self.nle_language.text_message(obs["tty_chars"]).decode("latin-1")
+            message += add
+            reward += rew
+
+        obs["text_message"] = message
+
+        return obs, reward, done, info

balrog/environments/nle/base.py

@@ -14,14 +14,13 @@
 
 
 class NLELanguageWrapper(language_wrapper.NLELanguageWrapper):
-    def __init__(self, env, vlm=False, skip_more=False):
+    def __init__(self, env, vlm=False):
         super().__init__(env, use_language_action=True)
         self.nle_language = nle_language_obsv.NLELanguageObsv()
         self.language_action_space = self.create_action_space()
         self.env = env
         self.vlm = vlm
         self.done = False
-        self.skip_more = skip_more
 
         if not vlm:
             self.prompt_mode = "hybrid"
@@ -77,28 +76,15 @@ def nle_obsv_type(self, nle_obsv):
         else:
             raise ValueError(f'"{self.prompt_mode}" is not a valid prompt mode.')
 
-    def clean_message(self, nle_obsv):
-        message = self.nle_language.text_message(nle_obsv["tty_chars"]).decode("latin-1")
-        if not self.skip_more:
-            while "--More--" in message and not self.done:
-                message = message.replace("--More--", " ")
-                message = message.replace("\n", " ")
-
-                nle_obsv, reward, self.done, info = self.step("more")
-                add = self.nle_language.text_message(nle_obsv["obs"]["tty_chars"]).decode("latin-1")
-                message += add
-                return message, nle_obsv["obs"]
-        return message, nle_obsv
-
     def render(self, mode="human"):
         if mode == "tiles":
-            obs = self.env.last_observation
-            glyphs = obs[self.env._observation_keys.index("glyphs")]
+            obs = self.env.unwrapped.last_observation
+            glyphs = obs[self.env.unwrapped._observation_keys.index("glyphs")]
             return rgb_render_image(glyphs)
         elif mode == "tty_image":
-            obs = self.env.last_observation
-            tty_chars = obs[self.env._observation_keys.index("tty_chars")]
-            tty_colors = obs[self.env._observation_keys.index("tty_colors")]
+            obs = self.env.unwrapped.last_observation
+            tty_chars = obs[self.env.unwrapped._observation_keys.index("tty_chars")]
+            tty_colors = obs[self.env.unwrapped._observation_keys.index("tty_colors")]
             return tty_render_image(tty_chars, tty_colors)
         else:
             return super().render(mode)
@@ -150,7 +136,11 @@ def nle_obsv_to_language(self, nle_obsv):
             (dict): language observation
         """
 
-        message, nle_obsv = self.clean_message(nle_obsv)
+        message = (
+            nle_obsv["text_message"]
+            if "text_message" in nle_obsv
+            else self.nle_language.text_message(nle_obsv["tty_chars"]).decode("latin-1")
+        )
 
         glyphs = nle_obsv["glyphs"]
         blstats = nle_obsv["blstats"]

balrog/environments/nle/nle_env.py

@@ -3,7 +3,7 @@
 import gym
 import nle  # NOQA: F401
 
-from balrog.environments.nle import NLELanguageWrapper
+from balrog.environments.nle import AutoMore, NLELanguageWrapper
 from balrog.environments.wrappers import GymV21CompatibilityV0, NLETimeLimit
 
 NETHACK_ENVS = []
@@ -18,7 +18,9 @@ def make_nle_env(env_name, task, config, render_mode: Optional[str] = None):
     skip_more = nle_kwargs.pop("skip_more", False)
     vlm = True if config.agent.max_image_history > 0 else False
     env = gym.make(task, **nle_kwargs)
-    env = NLELanguageWrapper(env, vlm=vlm, skip_more=skip_more)
+    if skip_more:
+        env = AutoMore(env)
+    env = NLELanguageWrapper(env, vlm=vlm)
 
     # wrap NLE with timeout
     env = NLETimeLimit(env)