def _generate_real_env_with_kwargs(self, kwargs, env_name):
# generate environment class
# todo: make generic (e.g. check if class is existent in bandit.py or gridworld.py)
if env_name == "Bandit":
env = TimeLimit(BanditFixedPermutedGaussian())
elif env_name == "EmptyRoom22":
env = TimeLimit(EmptyRoom22())
elif env_name == "EmptyRoom23":
env = TimeLimit(EmptyRoom23())
elif env_name == "EmptyRoom33":
env = TimeLimit(EmptyRoom33())
elif env_name == "WallRoom":
env = TimeLimit(WallRoom())
elif env_name == "HoleRoom":
env = TimeLimit(HoleRoom())
elif env_name == "HoleRoomLarge":
env = TimeLimit(HoleRoomLarge())
elif env_name == "HoleRoomLargeShifted":
env = TimeLimit(HoleRoomLargeShifted())
elif env_name == "Cliff":
env = TimeLimit(Cliff())
else:
env = gym.make(env_name)
for key, value in kwargs.items():
setattr(env, key, value)
# for episode termination
env._max_episode_steps = int(kwargs["max_steps"])
# for model save/load
env.kwargs = kwargs
return env
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = args.torch_deterministic
env.seed(args.seed)
env.action_space.seed(args.seed)
env.observation_space.seed(args.seed)
input_shape, preprocess_obs_fn = preprocess_obs_space(env.observation_space,
device)
output_shape = preprocess_ac_space(env.action_space)
# respect the default timelimit
if int(args.episode_length):
if not isinstance(env, TimeLimit):
env = TimeLimit(env, int(args.episode_length))
else:
env._max_episode_steps = int(args.episode_length)
else:
args.episode_length = env._max_episode_steps if isinstance(
env, TimeLimit) else 200
if args.capture_video:
env = Monitor(env, f'videos/{experiment_name}')
assert isinstance(env.action_space,
Box), "only continuous action space is supported"
# ALGO LOGIC: initialize agent here:
class Policy(nn.Module):
def __init__(self):
super(Policy, self).__init__()
self.fc1 = nn.Linear(input_shape, 120)
self.fc2 = nn.Linear(120, 84)