help="If specified, sets the recording dir, default=Disabled")
parser.add_argument("-s",
"--save",
type=int,
help="If specified, save every N-th step as an image")
parser.add_argument("--acktr",
default=False,
action='store_true',
help="Enable Acktr-specific tweaks")
args = parser.parse_args()
get_link_state = rospy.ServiceProxy("/gazebo/get_link_state", GetLinkState)
pitch = 0
rospy.Subscriber('/Bobby/imu', Imu, get_angular_vel)
counter = 0
env = make_env(args)
if args.record:
env = wrappers.Monitor(env, args.record)
net = model.ModelActor(env.observation_space.shape[0],
env.action_space.shape[0], args.hid)
if args.acktr:
opt = kfac.KFACOptimizer(net)
net.load_state_dict(torch.load(args.model))
obs = env.reset()
total_reward = 0.0
total_steps = 0
while True:
obs_v = torch.FloatTensor(obs)
GAMMA = 0.99
REWARD_STEPS = 5
BATCH_SIZE = 32
LEARNING_RATE_ACTOR = 1e-5
LEARNING_RATE_CRITIC = 1e-3
ENTROPY_BETA = 1e-3
ENVS_COUNT = 16
if __name__ == "__main__":
parser = make_parser()
args, device, save_path, test_env, maxeps, maxsec = parse_args(
parser, "a2c")
envs = [make_env(args) for _ in range(ENVS_COUNT)]
net_act, net_crt = make_nets(args, envs[0], device)
writer = SummaryWriter(comment="-a2c_" + args.name)
agent = model.AgentA2C(net_act, device=device)
exp_source = ptan.experience.ExperienceSourceFirstLast(
envs, agent, GAMMA, steps_count=REWARD_STEPS)
opt_act = optim.Adam(net_act.parameters(), lr=LEARNING_RATE_ACTOR)
opt_crt = optim.Adam(net_crt.parameters(), lr=LEARNING_RATE_CRITIC)
batch = []
best_reward = None
tstart = time.time()