wandb.save(os.path.abspath(__file__))
# TRY NOT TO MODIFY: seeding
device = torch.device(
'cuda' if torch.cuda.is_available() and args.cuda else 'cpu')
env = gym.make(args.gym_id)
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"
help="the time it takes from start-e to go end-e")
args = parser.parse_args()
if not args.seed:
args.seed = int(time.time())
# TRY NOT TO MODIFY: setup the environment
env = gym.make(args.gym_id)
env.seed(args.seed)
env.action_space.np_random.seed(args.seed)
env.observation_space.np_random.seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = args.torch_deterministic
input_shape, preprocess_obs_fn = preprocess_obs_space(env.observation_space)
output_shape, preprocess_ac_fn = preprocess_ac_space(env.action_space,
stochastic=False)
# TODO: initialize agent here:
er = ReplayBuffer(args.buffer_size)
class QNetwork(nn.Module):
def __init__(self):
super(QNetwork, self).__init__()
self.fc1 = nn.Linear(input_shape, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, output_shape)
def forward(self, x):
x = preprocess_obs_fn(x)
x = F.relu(self.fc1(x))
