def osimArmResume(ctxt=None,
snapshot_dir='data/local/experiment/osimArm_153',
seed=1):
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
runner.restore(snapshot_dir)
ddpg = runner._algo
env = GarageEnv(Arm2DVecEnv(visualize=True))
env.reset()
policy = ddpg.policy
env.render()
obs = env.step(env.action_space.sample())
steps = 0
n_steps = 100
while True:
if steps == n_steps:
env.close()
break
temp = policy.get_action(obs[0])
obs = env.step(temp[0])
env.render()
steps += 1
class TestNormalizedGym:
def setup_method(self):
self.env = GarageEnv(
normalize(gym.make('Pendulum-v0'),
normalize_reward=True,
normalize_obs=True,
flatten_obs=True))
def teardown_method(self):
self.env.close()
def test_does_not_modify_action(self):
a = self.env.action_space.sample()
a_copy = a
self.env.reset()
self.env.step(a)
assert a == a_copy
def test_flatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
self.env.render()
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
assert next_obs.shape == self.env.observation_space.low.shape
if done:
break
def test_unflatten(self):
for _ in range(10):
self.env.reset()
for _ in range(5):
action = self.env.action_space.sample()
next_obs, _, done, _ = self.env.step(action)
# yapf: disable
assert (self.env.observation_space.flatten(next_obs).shape
== self.env.observation_space.flat_dim)
# yapf: enable
if done:
break
def test_closes_box2d(self):
garage_env = GarageEnv(env_name='CarRacing-v0')
garage_env.render()
assert garage_env.env.viewer is not None
garage_env.close()
assert garage_env.env.viewer is None
def test_closes_mujoco(self):
garage_env = GarageEnv(env_name='Ant-v2')
garage_env.render()
assert garage_env.env.viewer is not None
garage_env.close()
assert garage_env.env.viewer is None
def osimArm(ctxt=None, seed=1):
"""Train TRPO with CartPole-v1 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
set_seed(seed)
with LocalTFRunner(ctxt) as runner:
env = GarageEnv(Arm2DVecEnv())
env.reset()
policy = ContinuousMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh)
exploration_policy = AddOrnsteinUhlenbeckNoise(env.spec,
policy,
sigma=0.2)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=tf.nn.relu)
replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))
ddpg = DDPG(env_spec=env.spec,
policy=policy,
policy_lr=1e-4,
qf_lr=1e-3,
max_path_length=200,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
target_update_tau=1e-2,
n_train_steps=50,
discount=0.9,
min_buffer_size=int(1e4),
exploration_policy=exploration_policy,
policy_optimizer=tf.compat.v1.train.AdamOptimizer,
qf_optimizer=tf.compat.v1.train.AdamOptimizer)
env.render()
obs = env.step(env.action_space.sample())
steps = 0
n_steps = 1000
while True:
if steps == n_steps:
env.close()
break
temp = policy.get_action(obs[0])
obs = env.step(temp[0])
env.render()
steps += 1
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
