kaggle-environments 0.2.1__py3-none-any.whl → 1.20.1__py3-none-any.whl

kaggle_environments/envs/football/football.py

@@ -0,0 +1,277 @@
+import importlib
+import json
+import shutil
+import uuid
+from os import path
+from pathlib import Path
+
+import numpy as np
+from gfootball.env import football_action_set
+
+
+def renderer(state, env):
+    html_renderer(env)
+
+
+def run_right_agent(obs):
+    # keep running right.
+    return [5] * obs.controlled_players
+
+
+def run_left_agent(obs):
+    # keep running left.
+    return [1] * obs.controlled_players
+
+
+def do_nothing_agent(obs):
+    # do nothing.
+    return [0] * obs.controlled_players
+
+
+def builtin_ai_agent(obs):
+    # execute builtin AI behavior.
+    return [19] * obs.controlled_players
+
+
+agents = {
+    "run_right": run_right_agent,
+    "run_left": run_left_agent,
+    "do_nothing": do_nothing_agent,
+    "builtin_ai": builtin_ai_agent,
+}
+
+
+def parse_single_player(obs_raw_entry):
+    # Remove pixel information.
+    if "frame" in obs_raw_entry:
+        del obs_raw_entry["frame"]
+    for k, v in obs_raw_entry.items():
+        if type(v) == np.ndarray:
+            obs_raw_entry[k] = v.tolist()
+    return obs_raw_entry
+
+
+def try_get_video(env, keep_running=False):
+    if not env.football_video_path:
+        internal_env = m_envs[env.configuration.id]
+        while not hasattr(internal_env, "_env"):
+            internal_env = internal_env.env
+        if not keep_running and internal_env._env._step == -1:
+            # Generate no-op step, so that video is available.
+            internal_env.step([0] * (env.configuration.team_1 + env.configuration.team_2))
+        trace = internal_env._env._trace
+        if trace:
+            trace._dump_config["episode_done"]._min_frequency = 0
+            dumps = trace.process_pending_dumps(True)
+            env.football_video_path = retrieve_video_link(dumps)
+        if not env.football_video_path:
+            return
+        if keep_running:
+            trace.write_dump("episode_done")
+    if "LiveVideoPath" in env.info and env.info["LiveVideoPath"] is not None:
+        target_path = Path(env.info["LiveVideoPath"])
+        target_path.parent.mkdir(parents=True, exist_ok=True)
+        shutil.move(env.football_video_path, target_path)
+        env.football_video_path = env.info["LiveVideoPath"]
+
+
+def update_observations_and_rewards(configuration, state, obs, rew=None):
+    """Updates agent-visible observations given 'raw' observations from environment.
+    Observations in 'obs' are coming directly from the environment and are in 'raw' format.
+    """
+    state[0].observation.controlled_players = configuration.team_1
+    state[1].observation.controlled_players = configuration.team_2
+
+    assert len(obs) == configuration.team_1 + configuration.team_2
+    if rew is not None:
+        state[0].reward = rew
+        state[1].reward = -rew
+    state[0].observation.players_raw = [parse_single_player(obs[x]) for x in range(configuration.team_1)]
+    state[1].observation.players_raw = [
+        parse_single_player(obs[x + configuration.team_1]) for x in range(configuration.team_2)
+    ]
+
+
+def mark_invalid(agent, message):
+    agent.status = "INVALID"
+    agent.reward = -100
+    agent.info.debug_info = message
+
+
+def maybe_terminate(env, state):
+    if state[0].status != "ACTIVE" or state[1].status != "ACTIVE":
+        if state[0].status == "ACTIVE":
+            state[0].status = "DONE"
+            state[0].reward = 100
+            state[0].info.debug_info = "Opponent forfeited. You win."
+        elif not state[0].reward:
+            state[0].reward = -100
+        if state[1].status == "ACTIVE":
+            state[1].status = "DONE"
+            state[1].reward = 100
+            state[1].info.debug_info = "Opponent forfeited. You win."
+        elif not state[1].reward:
+            state[1].reward = -100
+        try_get_video(env)
+        return True
+    return False
+
+
+def football_env():
+    # Use lazy-import to avoid this heavy dependency unless it is really needed.
+    return importlib.import_module("gfootball.env")
+
+
+# Global dictionary with active environments.
+m_envs = {}
+
+
+def cleanup(env):
+    global m_envs
+    del m_envs[env.configuration.id]
+
+
+def cleanup_all():
+    global m_envs
+    del m_envs
+
+
+def retrieve_video_link(dumps):
+    for entry in dumps:
+        if entry["name"] == "episode_done":
+            print("Received video link.")
+            return entry["video"]
+    return None
+
+
+def interpreter(state, env):
+    global m_envs
+    if "id" not in env.configuration or env.configuration.id is None:
+        env.configuration.id = str(uuid.uuid4())
+
+    if (env.configuration.id not in m_envs) or env.done:
+        if env.configuration.id not in m_envs:
+            print(
+                "Staring a new environment %s: with scenario: %s"
+                % (env.configuration.id, env.configuration.scenario_name)
+            )
+
+            other_config_options = {}
+            # Use webm to encode videos (so that you can see them in the browser).
+            other_config_options["video_format"] = "webm"
+            if env.configuration.running_in_notebook:
+                assert not env.configuration.render, "Render is not supported inside notebook environment."
+
+            env.football_video_path = None
+            if "TeamNames" in env.info:
+                names = env.info["TeamNames"]
+                assert len(names) == 2
+                other_config_options["custom_display_stats"] = [
+                    "LEFT PLAYER: %s" % names[0],
+                    "RIGHT PLAYER: %s" % names[1],
+                ]
+            m_envs[env.configuration.id] = football_env().create_environment(
+                env_name=env.configuration.scenario_name,
+                stacked=False,
+                # We use 'raw' representation to transfer data between server and agents.
+                representation="raw",
+                logdir=path.join(env.configuration.logdir, env.configuration.id),
+                write_goal_dumps=False,
+                write_full_episode_dumps=env.configuration.save_video,
+                write_video=env.configuration.save_video,
+                render=env.configuration.render,
+                number_of_left_players_agent_controls=env.configuration.team_1,
+                number_of_right_players_agent_controls=env.configuration.team_2,
+                other_config_options=other_config_options,
+            )
+        else:
+            print(
+                "Resetting environment %s: with scenario: %s" % (env.configuration.id, env.configuration.scenario_name)
+            )
+        obs = m_envs[env.configuration.id].reset()
+        update_observations_and_rewards(configuration=env.configuration, state=state, obs=obs)
+    if env.done:
+        return state
+
+    if maybe_terminate(env, state):
+        return state
+
+    # verify actions.
+    controlled_players = env.configuration.team_1
+    action_set = football_action_set.action_set_dict["default"]
+
+    try:
+        for action in state[0].action:
+            football_action_set.named_action_from_action_set(action_set, action)
+    except Exception:
+        mark_invalid(state[0], "Invalid action provided: %s." % state[0].action)
+    if len(state[0].action) != env.configuration.team_1:
+        mark_invalid(
+            state[0],
+            "Invalid number of actions provided: Expected %d, got %d."
+            % (env.configuration.team_1, len(state[0].action)),
+        )
+    actions_to_env = state[0].action
+
+    try:
+        for action in state[1].action:
+            football_action_set.named_action_from_action_set(action_set, action)
+    except Exception:
+        mark_invalid(state[1], "Invalid action provided: %s." % state[1].action)
+    if len(state[1].action) != env.configuration.team_2:
+        mark_invalid(
+            state[1],
+            "Invalid number of actions provided: Expected %d, got %d."
+            % (env.configuration.team_2, len(state[1].action)),
+        )
+    if env.configuration.team_2:
+        actions_to_env = actions_to_env + state[1].action
+
+    if maybe_terminate(env, state):
+        return state
+    obs, rew, done, info = m_envs[env.configuration.id].step(actions_to_env)
+
+    if "dumps" in info:
+        env.football_video_path = retrieve_video_link(info["dumps"])
+    update_observations_and_rewards(
+        configuration=env.configuration, state=state, obs=obs, rew=obs[0]["score"][0] - obs[0]["score"][1]
+    )
+
+    ## TODO: pass other information from 'info' to the state/agent.
+    if done:
+        for agent in range(2):
+            state[agent].status = "DONE"
+        try_get_video(env)
+
+    return state
+
+
+dirpath = path.dirname(__file__)
+jsonpath = path.abspath(path.join(dirpath, "football.json"))
+with open(jsonpath) as f:
+    specification = json.load(f)
+
+
+def html_renderer(env):
+    try_get_video(env, keep_running=True)
+    if not env.football_video_path:
+        raise Exception("No video found. Was environment created with save_video enabled?")
+
+    from base64 import b64encode
+
+    from IPython.display import HTML, display
+
+    video = open(env.football_video_path, "rb").read()
+    env.football_video_path = None
+    data_url = "data:video/webm;base64," + b64encode(video).decode()
+
+    html = (
+        """
+<video width=800 controls>
+  <source src="%s" type="video/webm">
+</video>
+"""
+        % data_url
+    )
+    display(HTML(html))
+    return ""

kaggle_environments/envs/football/helpers.py

@@ -0,0 +1,95 @@
+from enum import Enum
+from functools import wraps
+from typing import *
+
+
+class Action(Enum):
+    Idle = 0
+    Left = 1
+    TopLeft = 2
+    Top = 3
+    TopRight = 4
+    Right = 5
+    BottomRight = 6
+    Bottom = 7
+    BottomLeft = 8
+    LongPass = 9
+    HighPass = 10
+    ShortPass = 11
+    Shot = 12
+    Sprint = 13
+    ReleaseDirection = 14
+    ReleaseSprint = 15
+    Slide = 16
+    Dribble = 17
+    ReleaseDribble = 18
+
+
+sticky_index_to_action = [
+    Action.Left,
+    Action.TopLeft,
+    Action.Top,
+    Action.TopRight,
+    Action.Right,
+    Action.BottomRight,
+    Action.Bottom,
+    Action.BottomLeft,
+    Action.Sprint,
+    Action.Dribble,
+]
+
+
+class PlayerRole(Enum):
+    GoalKeeper = 0
+    CenterBack = 1
+    LeftBack = 2
+    RightBack = 3
+    DefenceMidfield = 4
+    CentralMidfield = 5
+    LeftMidfield = 6
+    RIghtMidfield = 7
+    AttackMidfield = 8
+    CentralFront = 9
+
+
+class GameMode(Enum):
+    Normal = 0
+    KickOff = 1
+    GoalKick = 2
+    FreeKick = 3
+    Corner = 4
+    ThrowIn = 5
+    Penalty = 6
+
+
+def human_readable_agent(agent: Callable[[Dict], Action]):
+    """
+    Decorator allowing for more human-friendly implementation of the agent function.
+
+    @human_readable_agent
+    def my_agent(obs):
+        ...
+        return football_action_set.action_right
+    """
+
+    @wraps(agent)
+    def agent_wrapper(obs) -> List[int]:
+        # Extract observations for the first (and only) player we control.
+        obs = obs["players_raw"][0]
+        # Turn 'sticky_actions' into a set of active actions (strongly typed).
+        obs["sticky_actions"] = {
+            sticky_index_to_action[nr] for nr, action in enumerate(obs["sticky_actions"]) if action
+        }
+        # Turn 'game_mode' into an enum.
+        obs["game_mode"] = GameMode(obs["game_mode"])
+        # In case of single agent mode, 'designated' is always equal to 'active'.
+        if "designated" in obs:
+            del obs["designated"]
+        # Conver players' roles to enum.
+        obs["left_team_roles"] = [PlayerRole(role) for role in obs["left_team_roles"]]
+        obs["right_team_roles"] = [PlayerRole(role) for role in obs["right_team_roles"]]
+
+        action = agent(obs)
+        return [action.value]
+
+    return agent_wrapper