def policy_fn(name, ob_space, ac_space):
return MlpPolicy(
name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=32,
num_hid_layers=2)
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(name=name,
ob_space=env.observation_space,
ac_space=env.action_space,
hid_sizes=config["hid_sizes"],
gaussian_fixed_var=config["gaussian_fixed_var"],
use_obfilter=config["use_obfilter"])
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(
name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=64, #params["hid_size"],
num_hid_layers=1, #params["num_hid_layers"]
)
def policy_fn(name, ob_space, ac_space, sess=None, placeholders=None):
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=32,
num_hid_layers=2,
sess=sess,
placeholders=placeholders)
def create_policy(name, problem):
ob_space = problem.observation_space
ac_space = problem.action_space
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=64,
num_hid_layers=2)
def train(env_id, num_timesteps, hidden_size, num_hidden_layers, seed, rank):
with U.make_session(3) as sess:
worker_seed = seed + 10000 * rank
set_global_seeds(worker_seed)
# env = bench.Monitor(env, logger.get_dir() and
# osp.join(logger.get_dir(), str(rank)))
try:
env = gym.make(env_id)
env.seed(worker_seed)
# Rendering and saving callback
episode = 0
def episode_callback(locals, globals):
nonlocal episode
episode += 1
print("----- Episode {} -----".format(episode))
env.render()
if episode % 20 == 0:
save(sess)
# Policy function
policy_fn = lambda name, ob_space, ac_space: MlpPolicy(
name=name,
ob_space=env.observation_space,
ac_space=env.action_space,
hid_size=hidden_size,
num_hid_layers=num_hidden_layers
)
# Learning
trpo_mpi.learn(
env,
policy_fn,
timesteps_per_batch=1024,
max_kl=0.01,
cg_iters=10,
cg_damping=0.1,
max_timesteps=num_timesteps,
gamma=0.99,
lam=0.98,
vf_iters=5,
vf_stepsize=1e-3,
callback=episode_callback)
finally:
env.close()
def policy_fn(name, ob_space, ac_space):
"""Create policy for baselines.
Args:
name (str): Policy name.
ob_space (gym.spaces.Box) : Observation space.
ac_space (gym.spaces.Box) : Action space.
Returns:
baselines.ppo1.mlp_policy: MLP policy for baselines.
"""
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=hyper_parameters['hidden_sizes'][0],
num_hid_layers=len(hyper_parameters['hidden_sizes']))
def load(checkpoint, env_id, hidden_size, num_hidden_layers):
with U.single_threaded_session() as sess:
try:
env = gym.make(env_id)
policy = MlpPolicy(
name="pi",
ob_space=env.observation_space,
ac_space=env.action_space,
hid_size=hidden_size,
num_hid_layers=num_hidden_layers)
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
seg_gen = traj_segment_generator(
policy, env, 1024, True, human_render=True)
# Generate trajectory segments until stopped
for _ in seg_gen:
pass
finally:
env.close()
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(name=name,
ob_space=env.observation_space,
ac_space=env.action_space,
hid_size=32,
num_hid_layers=2)
def __init__(self, name, ob_space, ac_space, hid_size, num_hid_layers):
MlpPolicy.__init__(self, name, ob_space, ac_space, hid_size,
num_hid_layers)
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=hidden_sizes[0],
num_hid_layers=len(hidden_sizes))
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(name=name,
ob_space=env.observation_space,
ac_space=env.action_space,
hid_size=rl_params['pi_hidden_size'],
num_hid_layers=rl_params['pi_hidden_layer'])
def __init__(self, name, *args, session=None, **kwargs):
MlpPolicy.__init__(self, name, *args, **kwargs)
NetworkSaverMLP.__init__(self, network_id=name)
#print(x)
#print(y)
line.set_data(*(x, y))
time = float(i) * 0.0005
time_text.set_text('time = %.2f' % time)
return line, time_text
U.make_session(num_cpu=1).__enter__()
env = TwoDofArmEnv(ActiveMuscles='agonist',
actionParameterization=True,
sim_length=0.2)
pol = MlpPolicy("pi",
env.observation_space,
env.action_space,
hid_size=64,
num_hid_layers=2)
U.initialize()
U.load_state('reacher')
o = env.reset()
time = 0.
data = np.empty((1, 8))
while time < 5.0:
print(time)
ac, vpred = pol.act(False, o)
o, r, d, look = env.step(ac)
data = np.append(data, look['data'], axis=0)
def policy_fn(name, ob_space, ac_space):
return MlpPolicy(name=name,
ob_space=env.observation_space,
ac_space=env.action_space,
hid_sizes=[32, 32])