def evaluate_single_episode(agent_ids, agent_paths):
num_agents = agent_ids.size
agents = []
for id_ in agent_ids:
my_agent_file = environment_utils.read_file(agent_full_paths[id_])
agents.append(environment_utils.get_last_callable(my_agent_file))
halite_scores = evaluate_game(
"halite", agents, num_episodes=1,
configuration={"agentExec": "LOCAL",
# "agentTimeout": 10000, # Uncomment while debugging
# "actTimeout": 10000, # Uncomment while debugging
})[0]
halite_scores = [-1 if h is None else h for h in halite_scores]
episode_rewards = np.zeros((num_agents))
for i in range(0, num_agents-1):
for j in range(i+1, num_agents):
if halite_scores[i] == halite_scores[j]:
first_score = 0.5
else:
first_score = int(halite_scores[i] > halite_scores[j])
episode_rewards[i] += first_score
episode_rewards[j] += 1-first_score
episode_rewards /= (num_agents-1)
return episode_rewards, halite_scores
def sample_from_config_or_path(config_or_path, return_callable):
if isinstance(config_or_path, str):
agent_file = environment_utils.read_file(config_or_path)
if return_callable:
try:
return kaggle_agent.get_last_callable(agent_file)
except:
return environment_utils.get_last_callable(agent_file)
else:
return agent_file
else:
return sample_from_config(config_or_path)
def action_handler(args):
args = utils.structify({
"action":
utils.get(args, str, "list", ["action"]),
"agents":
utils.get(args, list, [], ["agents"]),
"configuration":
utils.get(args, dict, {}, ["configuration"]),
"environment":
args.get("environment", None),
"episodes":
utils.get(args, int, 1, ["episodes"]),
"steps":
utils.get(args, list, [], ["steps"]),
"render":
utils.get(args, dict, {"mode": "json"}, ["render"]),
"debug":
utils.get(args, bool, False, ["debug"])
})
for index, agent in enumerate(args.agents):
agent = utils.read_file(agent, agent)
args.agents[index] = utils.get_last_callable(agent, agent)
if args.action == "list":
return action_list(args)
if args.environment == None:
return {"error": "Environment required."}
try:
if args.action == "http-server":
return {"error": "Already running a http server."}
elif args.action == "evaluate":
return action_evaluate(args)
elif args.action == "step":
return action_step(args)
elif args.action == "run":
return action_run(args)
elif args.action == "load":
return action_load(args)
else:
return {"error": "Unknown Action"}
except Exception as e:
return {"error": str(e), "trace": traceback.format_exc()}
def main():
"""
Makes sure that the `submission_standalone.py` file is valid.
"""
out = sys.stdout
submission = utils.read_file("submission_standalone.py")
agent = utils.get_last_callable(submission)
sys.stdout = out
env = make("halite", debug=True)
env.run([agent, agent])
if not env.state[0].status == env.state[1].status == "DONE":
raise ValidationError(
"`submission_standalone.py` file is not vaild. ",
f"agent #1 state: '{env.state[0].status}', ",
f"agent #2 state: '{env.state[1].status}'",
)
print("`submission_standalone.py` file is valid.")
agents_paths_video_names.append((agent_paths, video_name))
if generate_self_play_videos:
for i in range(num_agents):
agent_path = agent_full_paths[i]
agent_paths = [agent_path, agent_path, agent_path, agent_path]
video_name = agent_extensions[i][:-3] + " ***self play***"
agents_paths_video_names.append((agent_paths, video_name))
if agents_paths_video_names:
# Load all agent callables once (not really that much more performant)
agent_callables = {}
for i in range(num_agents):
agent_path = agent_full_paths[i]
agent_file = environment_utils.read_file(agent_path)
agent_callables[agent_path] = environment_utils.get_last_callable(
agent_file)
env = make_environment("halite", configuration={"agentExec": "LOCAL"})
for agent_paths, video_name in agents_paths_video_names:
agents = []
for p in agent_paths:
agents.append(agent_callables[p])
env.reset(num_agents=len(agents))
env.run(agents)
# Save the HTML recording in the videos folder
game_recording = env.render(mode="html", width=800, height=600)
videos_folder = os.path.join(agents_folder, '../Videos')
Path(videos_folder).mkdir(parents=True, exist_ok=True)
video_path = os.path.join(videos_folder, video_name+'.html')
f.write(" {}".format(line))
# f.write(f' state_dict = {state_dict}\n') # TODO: without loading the state dict the params are initialized randomly
f.write(
f' action_space = {action_space}\n' +
f' observation_space = {observation_space}\n' +
' model = QNet(action_space, observation_space, batch_norm=True)\n'
+
# ' model.load_state_dict(state_dict)\n' + # TODO: see above
' model.eval()\n' +
' return int(model.get_action(observation))\n')
print("model written to", save_file)
if __name__ == '__main__':
args = _parse()
if args.write:
env = {'connectx': ConnectX}[args.env]()
file = os.path.join('submission', args.env, 'submission.py')
load_path = os.path.join('out', args.name, 'models', args.name + '.pt')
write_agent_to_file(env, file, QNet, load_path)
out = sys.stdout
submission = utils.read_file(file)
agent = utils.get_last_callable(submission)
sys.stdout = out
env = make(args.env, debug=True)
env.run([agent, agent])
print("Success!" if env.state[0].status == env.state[1].
status == "DONE" else "Failed...")