def test_get_action(self, obs_dim, action_dim):
env = GarageEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
policy = CategoricalMLPPolicy(env_spec=env.spec)
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions(
[obs.flatten(), obs.flatten(),
obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_get_action(self, obs_dim, action_dim, obs_type):
assert obs_type in ['discrete', 'dict']
if obs_type == 'discrete':
env = GymEnv(
DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
else:
env = GymEnv(
DummyDictEnv(obs_space_type='box', act_space_type='discrete'))
policy = CategoricalMLPPolicy(env_spec=env.spec)
obs = env.reset()[0]
if obs_type == 'discrete':
obs = obs.flatten()
action, _ = policy.get_action(obs)
assert env.action_space.contains(action)
actions, _ = policy.get_actions([obs, obs, obs])
for action in actions:
assert env.action_space.contains(action)
def test_get_action(self, obs_dim, action_dim):
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
obs_var = tf.compat.v1.placeholder(
tf.float32,
shape=[None, None, env.observation_space.flat_dim],
name='obs')
policy = CategoricalMLPPolicy(env_spec=env.spec)
policy.build(obs_var)
obs = env.reset()
action, _ = policy.get_action(obs.flatten())
assert env.action_space.contains(action)
actions, _ = policy.get_actions(
[obs.flatten(), obs.flatten(),
obs.flatten()])
for action in actions:
assert env.action_space.contains(action)
def test_get_action(self, mock_rand, obs_dim, action_dim):
mock_rand.return_value = 0
env = TfEnv(DummyDiscreteEnv(obs_dim=obs_dim, action_dim=action_dim))
with mock.patch(('garage.tf.policies.'
'categorical_mlp_policy.MLPModel'),
new=SimpleMLPModel):
policy = CategoricalMLPPolicy(env_spec=env.spec)
env.reset()
obs, _, _, _ = env.step(1)
action, prob = policy.get_action(obs)
expected_prob = np.full(action_dim, 0.5)
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(prob['prob'], expected_prob)
actions, probs = policy.get_actions([obs, obs, obs])
for action, prob in zip(actions, probs['prob']):
assert env.action_space.contains(action)
assert action == 0
assert np.array_equal(prob, expected_prob)