def __init__(self, args=None):
self.env = None
self.sampler = None
self.experiment = None
self.c_level = None
self.pol = None
if args is None:
args = ArgsFromFile(self.default_args)
self.args = args
assert args.mirror in [False, "old", "new", "tuple"]
if args.mirror == "old":
args.mirror = True
args.mirror_tuples = args.mirror == "tuple"
self.env_kwargs = copy.deepcopy(getattr(self.args, "env_kwargs", {}))
device_name = "cpu" if args.cuda < 0 else "cuda:{}".format(args.cuda)
set_device(th.device(device_name))
self.num_updates = (int(args.num_total_frames) // args.num_steps //
args.num_processes)
self.seed_torch(args.seed)
self.setup_env()
if args.render or args.evaluate:
self.load_save(args.env, "last")
else:
if hasattr(args, "load_path"):
self.load_save(args.env, "last")
else:
self.setup_nets()
self.setup_optims()
self.setup_experiment()
with open(os.path.join(args.log, 'args.json'), 'w') as f:
json.dump(vars(args), f)
pprint(vars(args))
# modelの保存場所確保
if not os.path.exists(os.path.join(args.log, 'models')):
os.mkdir(os.path.join(args.log, 'models'))
# 乱数の種固定
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# GPU or CPU
device_name = 'cpu' if args.cuda < 0 else "cuda:{}".format(args.cuda)
device = torch.device(device_name)
set_device(device)
# logのcsvファイル確保
score_file = os.path.join(args.log, 'progress.csv')
logger.add_tabular_output(score_file)
# Gymのenviromentを生成
from pybullet_envs.bullet.kukaCamGymEnv import KukaCamGymEnv
env = KukaCamGymEnv(renders=False,
isDiscrete=False) # renders=Trueだとmachinaのtrain進まない。相性が悪い？
env = FlattenedObservationWrapper(env)
flattend_observation_space = env.flattend_observation_space
env = GymEnv(env,
log_dir=os.path.join(args.log, 'movie'),
record_video=args.record)