def set_policy(self):
env = gym.make(self.cfg['env_name'])
set_seed(self.cfg['seed'])
torch.set_default_tensor_type('torch.DoubleTensor')
state_dim, action_dim, is_disc_action = get_gym_info(env)
running_state = ZFilter((state_dim,), clip=5)
if is_disc_action:
policy_net = DiscretePolicy(state_dim, action_dim)
else:
policy_net = DiagnormalPolicy(state_dim, action_dim, log_std=self.cfg['log_std'])
return env, running_state, policy_net
def env_factory(thread_id):
env = gym.make(args.env_name)
env.seed(args.seed + thread_id)
return env
state_dim, action_dim, is_disc_action = get_gym_info(env_factory)
running_state = ZFilter((state_dim, ), clip=5)
# Define actor, critic and their optimizers
if is_disc_action:
policy_net = DiscretePolicy(state_dim, action_dim)
else:
policy_net = DiagnormalPolicy(state_dim, action_dim, log_std=args.log_std)
value_net = ValueFunction(state_dim)
device = torch.device("cuda" if use_gpu and args.gpu else "cpu")
if use_gpu and args.gpu:
policy_net = policy_net.to(device)
value_net = value_net.to(device)
running_state = ZFilter((state_dim, ), clip=5)
optimizer_policy = torch.optim.Adam(policy_net.parameters(), lr=args.lr_policy)
optimizer_value = torch.optim.Adam(value_net.parameters(), lr=args.lr_value)
cfg = Cfg(parse=args)
agent = ActorCriticAgent("A2c" + args.dis,
env_factory,
torch.set_default_tensor_type('torch.DoubleTensor')
set_seed(args.seed)
def env_factory(thread_id):
env = gym.make(args.env_name)
env.seed(args.seed + thread_id)
return env
state_dim, action_dim, is_disc_action = get_gym_info(env_factory)
running_state = ZFilter((state_dim, ), clip=5)
# Define actor, critic and their optimizers
assert not is_disc_action
policy_net = DiagnormalPolicy(state_dim, action_dim, log_std=args.log_std)
value_net = ValueFunction(state_dim)
device = torch.device("cuda" if use_gpu and args.gpu else "cpu")
if use_gpu and args.gpu:
policy_net = policy_net.to(device)
value_net = value_net.to(device)
optimizer_policy = torch.optim.Adam(policy_net.parameters(),
lr=args.learning_rate)
optimizer_value = torch.optim.Adam(value_net.parameters(),
lr=args.learning_rate)
cfg = Cfg(parse=args)
cfg["estimate_adv_and_target"] = False
agent = ActorCriticAgent("TSac" + args.dis,
set_seed(args.seed)
def env_factory(thread_id):
env = gym.make(args.env_name)
env.seed(args.seed + thread_id)
return env
state_dim, action_dim, is_disc_action = get_gym_info(env_factory)
running_state = ZFilter((state_dim,), clip=5)
# Define actor, critic and their optimizers
assert not is_disc_action
policy_net = AdditiveDiagnormalPolicy(state_dim, action_dim, args.policy_num, args.alpha, args.gpu)
temporary_policy = DiagnormalPolicy(state_dim, action_dim)
value_net = ValueFunction(state_dim)
device = torch.device("cuda" if use_gpu and args.gpu else "cpu")
if use_gpu and args.gpu:
policy_net = policy_net.to(device)
temporary_policy = temporary_policy.to(device)
value_net = value_net.to(device)
optimizer_policy = torch.optim.Adam(temporary_policy.parameters(), lr=args.learning_rate)
optimizer_value = torch.optim.Adam(value_net.parameters(), lr=args.learning_rate)
cfg = Cfg(parse=args)
agent = ActorCriticAgent("GbPPO" + args.dis, env_factory, temporary_policy,
value_net, cfg, running_state=running_state)
gbpo_ppo = GbpoUpdater(policy_net, temporary_policy, value_net, optimizer_policy, optimizer_value, cfg)
def env_factory(thread_id):
env = gym.make(args.env_name)
env.seed(args.seed + thread_id)
return env
set_seed(args.seed)
torch.set_default_tensor_type('torch.DoubleTensor')
state_dim, action_dim, _ = get_gym_info(env_factory)
running_state = ZFilter((state_dim,), clip=5)
# Define actor, critic and discriminator
hidden = [10 * state_dim, math.ceil(math.sqrt(50 * state_dim)), 5]
policy_net = DiagnormalPolicy(state_dim, action_dim, hidden=hidden, log_std=args.log_std)
value_net = ValueFunction(state_dim)
del hidden
if args.variate == 'linear':
variate_net = LinearVariate(state_dim, action_dim)
elif args.variate == 'quadratic':
variate_net = QuadraticVariate(state_dim, action_dim)
else:
variate_net = MlpVariate(state_dim, action_dim)
nets = {'policy': policy_net,
'value': value_net,
'variate': variate_net}
if use_gpu and args.gpu: