def randomString(stringLength=10):
letters = string.ascii_lowercase
return ''.join(random.choice(letters) for i in range(stringLength))
# Create logging and checkpointing folders.
random_experiment_name = randomString(10)
writer = SummaryWriter(log_dir="runs/" + random_experiment_name)
directory = os.path.join(args['save_dir'], random_experiment_name)
if not os.path.exists(directory):
os.makedirs(directory)
args["device"] = "cpu"
with open(os.path.join(directory, 'params.json'), 'w') as json_file:
json.dump(args, json_file)
agent = MRFAgent(args=args, writer=writer, added_u_dim = 0, temp=args["init_temp"])
save_dirs = os.path.join(directory, 'params_0')
agent.save_parameters(save_dirs)
# Evaluate initial parameter at training environment
avgs = []
num_hits, num_shoots = 0, 0
num_dones, per_worker_rew = [0] * args['num_envs'], [0] * args['num_envs']
agent.reset()
env_eval.seed([args['eval_init_seed'] + 1000 * rank for rank in range(args['num_envs'])])
obs = env_eval.reset()
while (all([k < args['eval_eps'] for k in num_dones])):
acts = agent.step(obs, eval=True)
num_shoots += sum([act == 7 for act in acts])
n_obs, rewards, dones, info = env_eval.step(acts)
return ''.join(random.choice(letters) for i in range(stringLength))
random_experiment_name = randomString(10)
writer = SummaryWriter(log_dir="runs/"+random_experiment_name)
directory = os.path.join(args['save_dir'], random_experiment_name)
if not os.path.exists(directory):
os.makedirs(directory)
with open(os.path.join(directory,'params.json'), 'w') as json_file:
json.dump(args, json_file)
with open(os.path.join('runs',random_experiment_name, 'params.json'), 'w') as json_file:
json.dump(args, json_file)
# Initialize the GPL-Q Agent
agent = MRFAgent(args=args, writer=writer, added_u_dim = 9)
# Define the training environment
num_players_train = args['num_players_train']
num_players_test = args['num_players_test']
def make_env(env_id, rank, seed=1285, effective_max_num_players=3, with_shuffle=False, gnn_input=True):
def _init():
env = gym.make(
env_id, seed=seed + rank,
effective_max_num_players=effective_max_num_players,
init_num_players=effective_max_num_players,
with_shuffle=with_shuffle,
gnn_input=gnn_input
)
return env
num_players_train = args['num_players_train']
num_players_test = args['num_players_test']
env = AsyncVectorEnv([
make_env(args,
i,
active_agents=num_players_train,
seed=args['seed'],
reward_scheme="sparse") for i in range(8)
])
args["device"] = "cpu"
writer = None
for idx in range(101):
agent = MRFAgent(args=args, writer=writer, added_u_dim=0)
load_dir = args['loading_dir'] + str(idx)
agent.load_parameters(load_dir)
obs_list = []
agent.reset()
obs = env.reset()
for i in range(3000):
print(idx, i)
obs_list.append(obs)
acts = agent.step(obs, eval=True)
n_obs, reward, done, info = env.step(acts)
obs = n_obs
agent.save_util_storage("utils_" + str(idx))