def main():
args = get_args()
print_args(args)
log_dir = create_log_dir(args)
if not args.evaluate:
writer = SummaryWriter(log_dir)
SEED = 721
env = make_env(args) # "LaserTag-small2-v0" "SlimeVolleyPixel-v0"
print(env.observation_space, env.action_space)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
test(env, args)
env.close()
return
train(env, args, writer)
writer.export_scalars_to_json(os.path.join(log_dir, "all_scalars.json"))
writer.close()
env.close()
def main():
args = get_args()
args.noisy = True
args.double = True
args.dueling = True
args.prioritized_replay = True
args.c51 = True
args.multi_step = 3
args.load_agents = True
args.num_agents = 12
args.read_model = None
args.evaluate = False
print_args(args)
log_dir = create_log_dir(args)
if not args.evaluate:
writer = SummaryWriter(log_dir)
env = PanicEnv(num_agents=args.num_agents,
scenario_=Scenario.Two_Exits,
load_agents=True,
read_agents=False)
set_global_seeds(args.seed)
if args.evaluate:
test(env, args)
return
train(env, args, writer)
writer.export_scalars_to_json(os.path.join(log_dir, "all_scalars.json"))
writer.close()
def main():
args = get_args()
print_args(args)
model_path = f'models/bilateral_dqn/{args.env}'
os.makedirs(model_path, exist_ok=True)
log_dir = create_log_dir(args)
if not args.evaluate:
writer = SummaryWriter(log_dir)
SEED = 721
if args.num_envs == 1 or args.evaluate:
env = make_env(
args) # "SlimeVolley-v0", "SlimeVolleyPixel-v0" 'Pong-ram-v0'
else:
VectorEnv = [
DummyVectorEnv, SubprocVectorEnv
][1] # https://github.com/thu-ml/tianshou/blob/master/tianshou/env/venvs.py
env = VectorEnv([lambda: make_env(args) for _ in range(args.num_envs)])
print(env.observation_space, env.action_space)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
test(env, args, model_path)
env.close()
return
train(env, args, writer, model_path)
# writer.export_scalars_to_json(os.path.join(log_dir, "all_scalars.json"))
writer.close()
env.close()
def main():
args = get_args()
print_args(args)
if args.evaluate:
if args.env == "1DStatic":
env = Env1DStatic(args)
elif args.env == "1DDynamic":
env = Env1DDynamic_Validation(args)
elif args.env == "2DStatic":
env = Env2DStatic(args)
elif args.env == "2DDynamic":
env = Env2DDynamic_Validation(args)
elif args.env == "3DStatic":
env = Env3DStatic(args)
elif args.env == "3DDynamic":
env = Env3DDynamic_Validation(args)
else:
if args.env == "1DStatic":
env = Env1DStatic(args)
elif args.env == "1DDynamic":
env = Env1DDynamic(args)
elif args.env == "2DStatic":
env = Env2DStatic(args)
elif args.env == "2DDynamic":
env = Env2DDynamic(args)
elif args.env == "3DStatic":
env = Env3DStatic(args)
elif args.env == "3DDynamic":
env = Env3DDynamic(args)
datetime = time.time()
save_hyperparameters(args, datetime)
log_dir = create_log_dir(args)
writer = SummaryWriter(log_dir)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
validate(env, args)
else:
train(env, args, writer, datetime)
writer.flush()
writer.close()
env.close()
def main():
args = get_args()
log_dir = create_log_dir(args)
if not args.evaluate:
writer = SummaryWriter(log_dir)
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
test(env, args)
env.close()
return
train(env, args, writer)
writer.export_scalars_to_json(os.path.join(log_dir, "all_scalars.json"))
writer.close()
env.close()
def main():
args = get_args()
print_args(args)
log_dir = create_log_dir(args)
print("Log dir is:", log_dir)
if not args.evaluate:
writer = SummaryWriter(log_dir)
env = gym.make(args.env)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
test(env, args)
env.close()
return
train(env, args, writer)
writer.export_scalars_to_json(os.path.join(log_dir, "all_scalars.json"))
writer.close()
env.close()
def main():
args = get_args()
print_args(args)
log_dir = create_log_dir(args)
wandb.init(project=args.wandb_project,
name=args.wandb_name,
notes=args.wandb_notes,
config=args)
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
set_global_seeds(args.seed)
env.seed(args.seed)
if args.evaluate:
test(env, args)
env.close()
return
train(env, args)
env.close()
def main():
Exploiter = 'DQN'
EvaluatedModel = 'NashDQN'
args = get_args()
# args.against_baseline=False
print_args(args)
env = make_env(
args) # "SlimeVolley-v0", "SlimeVolleyPixel-v0" 'Pong-ram-v0'
print(env.observation_space, env.action_space)
model_prefix = model_metadata[args.env]
exploiter = load_exploiter(env, Exploiter, args)
evaluated_model = load_evaluated_model(env, EvaluatedModel, args)
model_dir = "models/nash_dqn/{}/{}/".format(args.env, model_prefix)
exploiter_dir = "models/nash_dqn/{}/{}/exploiter/".format(
args.env, model_prefix)
os.makedirs(model_dir, exist_ok=True)
os.makedirs(exploiter_dir, exist_ok=True)
log_dir = create_log_dir(args)
if not args.evaluate:
writer = SummaryWriter(log_dir)
set_global_seeds(args.seed)
env.seed(args.seed)
# Parse all models saved during training in order
filelist, epi_list = [], []
for filename in os.listdir(model_dir):
if filename.endswith("dqn"):
filelist.append(filename.split('_')[0] +
'_') # remove '_dqn' at end
epi_list.append(int(filename.split('_')[0]))
sort_idx = np.argsort(epi_list).tolist()
filelist = [x for _, x in sorted(zip(epi_list, filelist))
] # sort filelist according to the sorting of epi_list
epi_list.sort() # filelist.sort() will not give correct answer
print(epi_list)
# Evaluate/exploit all models saved during training in order
eval_data = {}
for f, i in zip(filelist, epi_list):
print('load model: ', i, model_dir, f)
# if i>17000:
evaluated_model.load_model(model_dir + f,
eval=True,
map_location='cuda:0')
exploiter_path = exploiter_dir + f
r, l = exploit(env,
evaluated_model,
exploiter,
args,
exploiter_path=exploiter_path)
eval_data[str(i)] = [r, l]
save_dir = 'data/{}/'.format(args.env)
os.makedirs(save_dir, exist_ok=True)
if args.fictitious:
save_dir += '/fictitious_eval_data.npy'
else:
save_dir += '/eval_data.npy'
np.save(save_dir, eval_data)
writer.close()
env.close()