def __init__(self, args):
vis = visdom.Visdom()
assert vis.check_connection()
FloatTensor = torch.cuda.FloatTensor
LongTensor = torch.cuda.LongTensor
ByteTensor = torch.cuda.ByteTensor
Tensor = FloatTensor
args = parser.parse_args()
# setup the environment
self.env = MultiObjectiveEnv(args.env_name)
# get state / action / reward sizes
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
from crl.envelope.meta import MetaAgent
from crl.envelope.models import get_new_model
model = get_new_model(args.model, state_size, action_size, reward_size)
agent = MetaAgent(model, args, is_train=True)
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
model = get_new_model(args.model, state_size, action_size, reward_size)
dicts = torch.load("{}{}.pth.tar".format(args.save,
"m.{}_e.{}_n.{}".format(args.model, args.env_name, args.name)))
model.load_state_dict(dicts['state_dict'])
self.agent = MetaAgent(model, args, is_train=False)
class Paint_Booth():
def __init__(self, args):
vis = visdom.Visdom()
assert vis.check_connection()
FloatTensor = torch.cuda.FloatTensor
LongTensor = torch.cuda.LongTensor
ByteTensor = torch.cuda.ByteTensor
Tensor = FloatTensor
args = parser.parse_args()
# setup the environment
self.env = MultiObjectiveEnv(args.env_name)
# get state / action / reward sizes
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
from crl.envelope.meta import MetaAgent
from crl.envelope.models import get_new_model
model = get_new_model(args.model, state_size, action_size, reward_size)
agent = MetaAgent(model, args, is_train=True)
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
model = get_new_model(args.model, state_size, action_size, reward_size)
dicts = torch.load("{}{}.pth.tar".format(args.save,
"m.{}_e.{}_n.{}".format(args.model, args.env_name, args.name)))
model.load_state_dict(dicts['state_dict'])
self.agent = MetaAgent(model, args, is_train=False)
def run_one_episode(self, w):
w = np.abs(w) / np.linalg.norm(w, ord=1)
w_e = w / np.linalg.norm(w, ord=2)
ttrw = np.array([0.0, 0.0])
terminal = False
self.env.reset()
cnt = 0
while not terminal:
state = self.env.observe()
mask = self.env.env.get_action_out_mask()
action = self.agent.act(state, preference=torch.from_numpy(w).type(FloatTensor), mask=mask)
next_state, reward, terminal = self.env.step(action)
reward[0] = 1 - reward[0]
reward[1] = self.env.env.get_distortion(absolute=True, tollerance=15) / 5
if cnt > 300:
terminal = True
ttrw = ttrw + reward # * np.power(args.gamma, cnt)
cnt += 1
# ttrw_w = w.dot(ttrw) * w_e
return ttrw_w
act_1,
act_2,
# q__max,
loss / cnt)
if num_eps + 1 % 500 == 0:
agent.save(
args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name,
args.name))
# agent.save(args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name, args.name))
if __name__ == '__main__':
args = parser.parse_args()
# args.episode_num = 600
# setup the environment
env = MultiObjectiveEnv(args.env_name)
# get state / action / reward sizes
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
# generate an agent for initial training
agent = None
if args.method == 'crl-naive':
from crl.naive.meta import MetaAgent
from crl.naive.models import get_new_model
elif args.method == 'crl-envelope':
from crl.envelope.meta import MetaAgent
from crl.envelope.models import get_new_model
elif args.method == 'crl-energy':
if np.linalg.norm(a - b, ord=1) < 0.20 * np.linalg.norm(b):
cnt += 1.0
break
elif base == 2:
if np.linalg.norm(a - b, ord=1) < 0.3:
cnt += 1.0
break
return cnt / len(A)
################# Control Frontier #################
if args.pltcontrol:
# setup the environment
env = MultiObjectiveEnv(args.env_name)
# generate an agent for plotting
agent = None
if args.method == 'crl-naive':
from crl.naive.meta import MetaAgent
elif args.method == 'crl-envelope':
from crl.envelope.meta import MetaAgent
elif args.method == 'crl-energy':
from crl.energy.meta import MetaAgent
model = torch.load("{}{}.pkl".format(
args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name,
args.name)))
agent = MetaAgent(model, args, is_train=False)
# compute opt